diff --git a/.github/workflows/create-tag.yml b/.github/workflows/create-tag.yml
index d8b3a9f6..9462bc4d 100644
--- a/.github/workflows/create-tag.yml
+++ b/.github/workflows/create-tag.yml
@@ -12,6 +12,7 @@ on:
           - acp
           - coder
           - console
+          - context-manager
           - database
           - email
           - harness
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index bfd129b8..0bf89907 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -7,6 +7,7 @@ on:
       - 'acp/v*'
       - 'coder/v*'
       - 'console/v*'
+      - 'context-manager/v*'
       - 'database/v*'
       - 'email/v*'
       - 'harness/v*'
diff --git a/README.md b/README.md
index afab3e69..ea3eb6cd 100644
--- a/README.md
+++ b/README.md
@@ -41,6 +41,7 @@ npx skills add iii-hq/iii --all
 | [`acp`](acp/) | Rust | Agent Client Protocol surface — stdio JSON-RPC, exposes iii agents as ACP sessions. |
 | [`harness`](harness/) | Node | TS port of the iii harness stack — bundles `harness` (provider registry + credentials/settings/permissions via the `configuration` worker), `turn-orchestrator`, `approval-gate`, `hook-fanout`, `models-catalog`, the `provider-*` workers, `llm-budget`, and `context-compaction` as one pnpm monorepo. Conversations persist in `session-manager`. See [`harness/README.md`](harness/README.md). |
 | [`session-manager`](session-manager/) | Rust | Durable, reactive, branching conversation store — fourteen `session::*` functions plus six trigger types; the transcript backend for `harness` and `console`. See [`session-manager/architecture/`](session-manager/architecture/). |
+| [`context-manager`](context-manager/) | Rust | Model-ready context assembly — four `context::*` functions for token counting, function-result pruning, and history compaction over caller-supplied messages. Storage-agnostic; summarisation via `llm-router` when installed. |
 | [`database`](database/) | Rust | PostgreSQL, MySQL, and SQLite client — query, execute, transactions, prepared statements, and change feeds. |
 | [`iii-directory`](iii-directory/) | Rust | Engine introspection (functions / triggers / workers), workers-registry proxy, and filesystem-backed skill + prompt reader. |
 | [`iii-lsp`](iii-lsp/) | Rust | Language Server for iii function ids, trigger configs, and worker discovery. Autocomplete / hover across JS/TS, Python, Rust. |
diff --git a/console/README.md b/console/README.md
index 67da94bb..90d6a3ac 100644
--- a/console/README.md
+++ b/console/README.md
@@ -33,7 +33,7 @@ A purpose-built agentic chat UI on top of [Lexical](https://lexical.dev). Lives
 - **Three modes** — `plan`, `ask`, and `agent` toggle right in the composer
 - **Live model picker** — provider-grouped from `models::list`; static fallback (OpenAI, Anthropic, Google) when the catalog is unreachable
 - **`@`-mentions** — fuzzy-search every function registered against the engine
-- **`/compact` slash command** — collapses conversation history via `context-compaction::compact_session`
+- **`/compact` slash command** — summarises conversation history via `context-compaction::compact_session` (a thin wrapper over the `context-manager` worker's `context::compact`); the durable transcript is untouched — a compaction bookkeeping entry is added and the summary anchors future turns
 - **Attachments** — multi-file picker with text/image previews
 - **Function calls** — running / pending / error cards, consecutive calls grouped, with **approve/deny** gating for pending approvals (`approval::resolve`)
 - **Streaming** — abortable mid-flight; thinking shimmer; collapsible thought messages
diff --git a/console/web/src/lib/functions.ts b/console/web/src/lib/functions.ts
index a19c863e..e9d116f4 100644
--- a/console/web/src/lib/functions.ts
+++ b/console/web/src/lib/functions.ts
@@ -17,10 +17,6 @@ export const STATIC_FUNCTIONS: FunctionEntry[] = [
     id: 'context-compaction::compact_session',
     description: 'compact a session now',
   },
-  {
-    id: 'context-compaction::prune_tool_outputs',
-    description: 'prune old tool outputs',
-  },
   {
     id: 'session::messages',
     description: 'read a session transcript',
diff --git a/context-manager/Cargo.lock b/context-manager/Cargo.lock
new file mode 100644
index 00000000..65b486ab
--- /dev/null
+++ b/context-manager/Cargo.lock
@@ -0,0 +1,3161 @@
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 4
+
+[[package]]
+name = "aho-corasick"
+version = "1.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ddd31a130427c27518df266943a5308ed92d4b226cc639f5a8f1002816174301"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "anstream"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "824a212faf96e9acacdbd09febd34438f8f711fb84e09a8916013cd7815ca28d"
+dependencies = [
+ "anstyle",
+ "anstyle-parse",
+ "anstyle-query",
+ "anstyle-wincon",
+ "colorchoice",
+ "is_terminal_polyfill",
+ "utf8parse",
+]
+
+[[package]]
+name = "anstyle"
+version = "1.0.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "940b3a0ca603d1eade50a4846a2afffd5ef57a9feac2c0e2ec2e14f9ead76000"
+
+[[package]]
+name = "anstyle-parse"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52ce7f38b242319f7cabaa6813055467063ecdc9d355bbb4ce0c68908cd8130e"
+dependencies = [
+ "utf8parse",
+]
+
+[[package]]
+name = "anstyle-query"
+version = "1.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "40c48f72fd53cd289104fc64099abca73db4166ad86ea0b4341abe65af83dadc"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "anstyle-wincon"
+version = "3.0.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "291e6a250ff86cd4a820112fb8898808a366d8f9f58ce16d1f538353ad55747d"
+dependencies = [
+ "anstyle",
+ "once_cell_polyfill",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "anyhow"
+version = "1.0.102"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
+
+[[package]]
+name = "async-trait"
+version = "0.1.89"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9035ad2d096bed7955a320ee7e2230574d28fd3c3a0f186cbea1ff3c7eed5dbb"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "atomic-waker"
+version = "1.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"
+
+[[package]]
+name = "base64"
+version = "0.22.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
+
+[[package]]
+name = "bitflags"
+version = "2.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b4388bee8683e3d04af747c73422af53102d2bd24d9eadb6cbc100baef4b43f8"
+
+[[package]]
+name = "block-buffer"
+version = "0.10.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3078c7629b62d3f0439517fa394996acacc5cbc91c5a20d8c658e77abd503a71"
+dependencies = [
+ "generic-array",
+]
+
+[[package]]
+name = "bstr"
+version = "1.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63044e1ae8e69f3b5a92c736ca6269b8d12fa7efe39bf34ddb06d102cf0e2cab"
+dependencies = [
+ "memchr",
+ "serde",
+]
+
+[[package]]
+name = "bumpalo"
+version = "3.20.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "72f5acc6cb2ba439de613abc23857ec3d78374d8ed5ac84e9d11336e87da8649"
+
+[[package]]
+name = "bytecount"
+version = "0.6.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "175812e0be2bccb6abe50bb8d566126198344f707e304f45c648fd8f2cc0365e"
+
+[[package]]
+name = "bytes"
+version = "1.11.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e748733b7cbc798e1434b6ac524f0c1ff2ab456fe201501e6497c8417a4fc33"
+
+[[package]]
+name = "cc"
+version = "1.2.64"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dad887fd958be91b5098c0248def011f4523ab786cd411be668777e55063501f"
+dependencies = [
+ "find-msvc-tools",
+ "shlex",
+]
+
+[[package]]
+name = "cfg-if"
+version = "1.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
+
+[[package]]
+name = "cfg_aliases"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
+
+[[package]]
+name = "clap"
+version = "4.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ddb117e43bbf7dacf0a4190fef4d345b9bad68dfc649cb349e7d17d28428e51"
+dependencies = [
+ "clap_builder",
+ "clap_derive",
+]
+
+[[package]]
+name = "clap_builder"
+version = "4.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "714a53001bf66416adb0e2ef5ac857140e7dc3a0c48fb28b2f10762fc4b5069f"
+dependencies = [
+ "anstream",
+ "anstyle",
+ "clap_lex",
+ "strsim",
+ "terminal_size",
+]
+
+[[package]]
+name = "clap_derive"
+version = "4.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2ce8604710f6733aa641a2b3731eaa1e8b3d9973d5e3565da11800813f997a9"
+dependencies = [
+ "heck",
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "clap_lex"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c8d4a3bb8b1e0c1050499d1815f5ab16d04f0959b233085fb31653fbfc9d98f9"
+
+[[package]]
+name = "colorchoice"
+version = "1.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
+
+[[package]]
+name = "console"
+version = "0.16.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d64e8af5551369d19cf50138de61f1c42074ab970f74e99be916646777f8fc87"
+dependencies = [
+ "encode_unicode",
+ "libc",
+ "unicode-width",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "context-manager"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "clap",
+ "cucumber",
+ "dirs",
+ "futures",
+ "iii-sdk",
+ "schemars",
+ "serde",
+ "serde_json",
+ "serde_yaml",
+ "sha2",
+ "tempfile",
+ "thiserror 2.0.18",
+ "tokio",
+ "tracing",
+ "tracing-subscriber",
+ "uuid",
+ "which",
+]
+
+[[package]]
+name = "convert_case"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "633458d4ef8c78b72454de2d54fd6ab2e60f9e02be22f3c6104cdc8a4e0fceb9"
+dependencies = [
+ "unicode-segmentation",
+]
+
+[[package]]
+name = "core-foundation"
+version = "0.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b2a6cd9ae233e7f62ba4e9353e81a88df7fc8a5987b8d445b4d90c879bd156f6"
+dependencies = [
+ "core-foundation-sys",
+ "libc",
+]
+
+[[package]]
+name = "core-foundation-sys"
+version = "0.8.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
+
+[[package]]
+name = "cpufeatures"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "59ed5838eebb26a2bb2e58f6d5b5316989ae9d08bab10e0e6d103e656d1b0280"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "crossbeam-deque"
+version = "0.8.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9dd111b7b7f7d55b72c0a6ae361660ee5853c9af73f70c3c2ef6858b950e2e51"
+dependencies = [
+ "crossbeam-epoch",
+ "crossbeam-utils",
+]
+
+[[package]]
+name = "crossbeam-epoch"
+version = "0.9.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5b82ac4a3c2ca9c3460964f020e1402edd5753411d7737aa39c3714ad1b5420e"
+dependencies = [
+ "crossbeam-utils",
+]
+
+[[package]]
+name = "crossbeam-utils"
+version = "0.8.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
+
+[[package]]
+name = "crypto-common"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "78c8292055d1c1df0cce5d180393dc8cce0abec0a7102adb6c7b1eef6016d60a"
+dependencies = [
+ "generic-array",
+ "typenum",
+]
+
+[[package]]
+name = "cucumber"
+version = "0.22.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "16cbb27bc2064274afa3a3d8bc9a0e71333589850573aa632ec4520e4af14d94"
+dependencies = [
+ "anyhow",
+ "clap",
+ "console",
+ "cucumber-codegen",
+ "cucumber-expressions",
+ "derive_more",
+ "either",
+ "futures",
+ "gherkin",
+ "globwalk",
+ "humantime",
+ "inventory",
+ "itertools",
+ "linked-hash-map",
+ "pin-project",
+ "ref-cast",
+ "regex",
+ "sealed",
+ "smart-default",
+]
+
+[[package]]
+name = "cucumber-codegen"
+version = "0.22.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5a1afaf9c422380861111c6be56f39b324e351fd9efc07a1486268798bf79cfd"
+dependencies = [
+ "cucumber-expressions",
+ "inflections",
+ "itertools",
+ "proc-macro2",
+ "quote",
+ "regex",
+ "syn",
+ "synthez",
+]
+
+[[package]]
+name = "cucumber-expressions"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6401038de3af44fe74e6fccdb8a5b7db7ba418f480c8e9ad584c6f65c05a27a6"
+dependencies = [
+ "derive_more",
+ "either",
+ "nom",
+ "nom_locate",
+ "regex",
+ "regex-syntax",
+]
+
+[[package]]
+name = "data-encoding"
+version = "2.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4ae5f15dda3c708c0ade84bfee31ccab44a3da4f88015ed22f63732abe300c8"
+
+[[package]]
+name = "derive_more"
+version = "2.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d751e9e49156b02b44f9c1815bcb94b984cdcc4396ecc32521c739452808b134"
+dependencies = [
+ "derive_more-impl",
+]
+
+[[package]]
+name = "derive_more-impl"
+version = "2.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "799a97264921d8623a957f6c3b9011f3b5492f557bbb7a5a19b7fa6d06ba8dcb"
+dependencies = [
+ "convert_case",
+ "proc-macro2",
+ "quote",
+ "rustc_version",
+ "syn",
+ "unicode-xid",
+]
+
+[[package]]
+name = "digest"
+version = "0.10.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
+dependencies = [
+ "block-buffer",
+ "crypto-common",
+]
+
+[[package]]
+name = "dirs"
+version = "5.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "44c45a9d03d6676652bcb5e724c7e988de1acad23a711b5217ab9cbecbec2225"
+dependencies = [
+ "dirs-sys",
+]
+
+[[package]]
+name = "dirs-sys"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "520f05a5cbd335fae5a99ff7a6ab8627577660ee5cfd6a94a6a929b52ff0321c"
+dependencies = [
+ "libc",
+ "option-ext",
+ "redox_users",
+ "windows-sys 0.48.0",
+]
+
+[[package]]
+name = "displaydoc"
+version = "0.2.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ac70aa55017e108007fbaf5aa0f54b021c98f92ff8af59d42eda9da96e3dd4f"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "dyn-clone"
+version = "1.0.20"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d0881ea181b1df73ff77ffaaf9c7544ecc11e82fba9b5f27b262a3c73a332555"
+
+[[package]]
+name = "either"
+version = "1.16.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "91622ff5e7162018101f2fea40d6ebf4a78bbe5a49736a2020649edf9693679e"
+
+[[package]]
+name = "encode_unicode"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "34aa73646ffb006b8f5147f3dc182bd4bcb190227ce861fc4a4844bf8e3cb2c0"
+
+[[package]]
+name = "equivalent"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f"
+
+[[package]]
+name = "errno"
+version = "0.3.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
+dependencies = [
+ "libc",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "fastrand"
+version = "2.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f1f227452a390804cdb637b74a86990f2a7d7ba4b7d5693aac9b4dd6defd8d6"
+
+[[package]]
+name = "find-msvc-tools"
+version = "0.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5baebc0774151f905a1a2cc41989300b1e6fbb29aff0ceffa1064fdd3088d582"
+
+[[package]]
+name = "foldhash"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2"
+
+[[package]]
+name = "form_urlencoded"
+version = "1.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb4cb245038516f5f85277875cdaa4f7d2c9a0fa0468de06ed190163b1581fcf"
+dependencies = [
+ "percent-encoding",
+]
+
+[[package]]
+name = "futures"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8b147ee9d1f6d097cef9ce628cd2ee62288d963e16fb287bd9286455b241382d"
+dependencies = [
+ "futures-channel",
+ "futures-core",
+ "futures-executor",
+ "futures-io",
+ "futures-sink",
+ "futures-task",
+ "futures-util",
+]
+
+[[package]]
+name = "futures-channel"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07bbe89c50d7a535e539b8c17bc0b49bdb77747034daa8087407d655f3f7cc1d"
+dependencies = [
+ "futures-core",
+ "futures-sink",
+]
+
+[[package]]
+name = "futures-core"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7e3450815272ef58cec6d564423f6e755e25379b217b0bc688e295ba24df6b1d"
+
+[[package]]
+name = "futures-executor"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "baf29c38818342a3b26b5b923639e7b1f4a61fc5e76102d4b1981c6dc7a7579d"
+dependencies = [
+ "futures-core",
+ "futures-task",
+ "futures-util",
+]
+
+[[package]]
+name = "futures-io"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cecba35d7ad927e23624b22ad55235f2239cfa44fd10428eecbeba6d6a717718"
+
+[[package]]
+name = "futures-macro"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e835b70203e41293343137df5c0664546da5745f82ec9b84d40be8336958447b"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "futures-sink"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c39754e157331b013978ec91992bde1ac089843443c49cbc7f46150b0fad0893"
+
+[[package]]
+name = "futures-task"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "037711b3d59c33004d3856fbdc83b99d4ff37a24768fa1be9ce3538a1cde4393"
+
+[[package]]
+name = "futures-util"
+version = "0.3.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "389ca41296e6190b48053de0321d02a77f32f8a5d2461dd38762c0593805c6d6"
+dependencies = [
+ "futures-channel",
+ "futures-core",
+ "futures-io",
+ "futures-macro",
+ "futures-sink",
+ "futures-task",
+ "memchr",
+ "pin-project-lite",
+ "slab",
+]
+
+[[package]]
+name = "generic-array"
+version = "0.14.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a"
+dependencies = [
+ "typenum",
+ "version_check",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ff2abc00be7fca6ebc474524697ae276ad847ad0a6b3faa4bcb027e9a4614ad0"
+dependencies = [
+ "cfg-if",
+ "js-sys",
+ "libc",
+ "wasi",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.3.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "899def5c37c4fd7b2664648c28120ecec138e4d395b459e5ca34f9cce2dd77fd"
+dependencies = [
+ "cfg-if",
+ "js-sys",
+ "libc",
+ "r-efi 5.3.0",
+ "wasip2",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0de51e6874e94e7bf76d726fc5d13ba782deca734ff60d5bb2fb2607c7406555"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "r-efi 6.0.0",
+ "wasip2",
+ "wasip3",
+]
+
+[[package]]
+name = "gherkin"
+version = "0.15.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "70197ce7751bfe8bc828e3a855502d3a869a1e9416b58b10c4bde5cf8a0a3cb3"
+dependencies = [
+ "heck",
+ "peg",
+ "quote",
+ "serde",
+ "serde_json",
+ "syn",
+ "textwrap",
+ "thiserror 2.0.18",
+ "typed-builder",
+]
+
+[[package]]
+name = "globset"
+version = "0.4.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52dfc19153a48bde0cbd630453615c8151bce3a5adfac7a0aebfbf0a1e1f57e3"
+dependencies = [
+ "aho-corasick",
+ "bstr",
+ "log",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "globwalk"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0bf760ebf69878d9fd8f110c89703d90ce35095324d1f1edcb595c63945ee757"
+dependencies = [
+ "bitflags",
+ "ignore",
+ "walkdir",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.15.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9229cfe53dfd69f0609a49f65461bd93001ea1ef889cd5529dd176593f5338a1"
+dependencies = [
+ "foldhash",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+
+[[package]]
+name = "heck"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
+
+[[package]]
+name = "hostname"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "617aaa3557aef3810a6369d0a99fac8a080891b68bd9f9812a1eeda0c0730cbd"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "windows-link",
+]
+
+[[package]]
+name = "http"
+version = "1.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6970f50e31d6fc17d3fa27329444bfa74e196cf62e95052a3f6fee181dba6425"
+dependencies = [
+ "bytes",
+ "itoa",
+]
+
+[[package]]
+name = "http-body"
+version = "1.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1efedce1fb8e6913f23e0c92de8e62cd5b772a67e7b3946df930a62566c93184"
+dependencies = [
+ "bytes",
+ "http",
+]
+
+[[package]]
+name = "http-body-util"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b021d93e26becf5dc7e1b75b1bed1fd93124b374ceb73f43d4d4eafec896a64a"
+dependencies = [
+ "bytes",
+ "futures-core",
+ "http",
+ "http-body",
+ "pin-project-lite",
+]
+
+[[package]]
+name = "httparse"
+version = "1.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
+
+[[package]]
+name = "humantime"
+version = "2.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "135b12329e5e3ce057a9f972339ea52bc954fe1e9358ef27f95e89716fbc5424"
+
+[[package]]
+name = "hyper"
+version = "1.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "55281c53a1894c864990125767da440a4e630446785086f52523b20033b74498"
+dependencies = [
+ "atomic-waker",
+ "bytes",
+ "futures-channel",
+ "futures-core",
+ "http",
+ "http-body",
+ "httparse",
+ "itoa",
+ "pin-project-lite",
+ "smallvec",
+ "tokio",
+ "want",
+]
+
+[[package]]
+name = "hyper-rustls"
+version = "0.27.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "33ca68d021ef39cf6463ab54c1d0f5daf03377b70561305bb89a8f83aab66e0f"
+dependencies = [
+ "http",
+ "hyper",
+ "hyper-util",
+ "rustls",
+ "tokio",
+ "tokio-rustls",
+ "tower-service",
+ "webpki-roots",
+]
+
+[[package]]
+name = "hyper-util"
+version = "0.1.20"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
+dependencies = [
+ "base64",
+ "bytes",
+ "futures-channel",
+ "futures-util",
+ "http",
+ "http-body",
+ "hyper",
+ "ipnet",
+ "libc",
+ "percent-encoding",
+ "pin-project-lite",
+ "socket2",
+ "tokio",
+ "tower-service",
+ "tracing",
+]
+
+[[package]]
+name = "icu_collections"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2984d1cd16c883d7935b9e07e44071dca8d917fd52ecc02c04d5fa0b5a3f191c"
+dependencies = [
+ "displaydoc",
+ "potential_utf",
+ "utf8_iter",
+ "yoke",
+ "zerofrom",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_locale_core"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92219b62b3e2b4d88ac5119f8904c10f8f61bf7e95b640d25ba3075e6cac2c29"
+dependencies = [
+ "displaydoc",
+ "litemap",
+ "tinystr",
+ "writeable",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_normalizer"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c56e5ee99d6e3d33bd91c5d85458b6005a22140021cc324cea84dd0e72cff3b4"
+dependencies = [
+ "icu_collections",
+ "icu_normalizer_data",
+ "icu_properties",
+ "icu_provider",
+ "smallvec",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_normalizer_data"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "da3be0ae77ea334f4da67c12f149704f19f81d1adf7c51cf482943e84a2bad38"
+
+[[package]]
+name = "icu_properties"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bee3b67d0ea5c2cca5003417989af8996f8604e34fb9ddf96208a033901e70de"
+dependencies = [
+ "icu_collections",
+ "icu_locale_core",
+ "icu_properties_data",
+ "icu_provider",
+ "zerotrie",
+ "zerovec",
+]
+
+[[package]]
+name = "icu_properties_data"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8e2bbb201e0c04f7b4b3e14382af113e17ba4f63e2c9d2ee626b720cbce54a14"
+
+[[package]]
+name = "icu_provider"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "139c4cf31c8b5f33d7e199446eff9c1e02decfc2f0eec2c8d71f65befa45b421"
+dependencies = [
+ "displaydoc",
+ "icu_locale_core",
+ "writeable",
+ "yoke",
+ "zerofrom",
+ "zerotrie",
+ "zerovec",
+]
+
+[[package]]
+name = "id-arena"
+version = "2.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3d3067d79b975e8844ca9eb072e16b31c3c1c36928edf9c6789548c524d0d954"
+
+[[package]]
+name = "idna"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3b0875f23caa03898994f6ddc501886a45c7d3d62d04d2d90788d47be1b1e4de"
+dependencies = [
+ "idna_adapter",
+ "smallvec",
+ "utf8_iter",
+]
+
+[[package]]
+name = "idna_adapter"
+version = "1.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb68373c0d6620ef8105e855e7745e18b0d00d3bdb07fb532e434244cdb9a714"
+dependencies = [
+ "icu_normalizer",
+ "icu_properties",
+]
+
+[[package]]
+name = "ignore"
+version = "0.4.26"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b915661dd01db3f05050265b2477bcc6527b3792388e2749b41623cc592be67d"
+dependencies = [
+ "crossbeam-deque",
+ "globset",
+ "log",
+ "memchr",
+ "regex-automata",
+ "same-file",
+ "walkdir",
+ "winapi-util",
+]
+
+[[package]]
+name = "iii-observability"
+version = "0.19.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e11586fcd304a563c143837f67b2e5f3cb73d23f6c8452c9734ff18e4a1402bf"
+dependencies = [
+ "futures-util",
+ "opentelemetry",
+ "opentelemetry-http",
+ "opentelemetry_sdk",
+ "reqwest",
+ "serde_json",
+ "sysinfo",
+ "tokio",
+ "tokio-tungstenite",
+ "tracing",
+ "uuid",
+]
+
+[[package]]
+name = "iii-sdk"
+version = "0.19.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8490f2ad470d54cf7e0bc2f4105aca71bdb4c24752fcb406183f24b8dd328f80"
+dependencies = [
+ "async-trait",
+ "futures-util",
+ "hostname",
+ "iii-observability",
+ "reqwest",
+ "schemars",
+ "serde",
+ "serde_json",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-tungstenite",
+ "tracing",
+ "uuid",
+]
+
+[[package]]
+name = "indexmap"
+version = "2.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d466e9454f08e4a911e14806c24e16fba1b4c121d1ea474396f396069cf949d9"
+dependencies = [
+ "equivalent",
+ "hashbrown 0.17.1",
+ "serde",
+ "serde_core",
+]
+
+[[package]]
+name = "inflections"
+version = "1.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a257582fdcde896fd96463bf2d40eefea0580021c0712a0e2b028b60b47a837a"
+
+[[package]]
+name = "inventory"
+version = "0.3.24"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4f0c30c76f2f4ccee3fe55a2435f691ca00c0e4bd87abe4f4a851b1d4dac39b"
+dependencies = [
+ "rustversion",
+]
+
+[[package]]
+name = "ipnet"
+version = "2.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d98f6fed1fde3f8c21bc40a1abb88dd75e67924f9cffc3ef95607bad8017f8e2"
+
+[[package]]
+name = "is_terminal_polyfill"
+version = "1.70.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6cb138bb79a146c1bd460005623e142ef0181e3d0219cb493e02f7d08a35695"
+
+[[package]]
+name = "itertools"
+version = "0.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2b192c782037fadd9cfa75548310488aabdbf3d2da73885b31bd0abd03351285"
+dependencies = [
+ "either",
+]
+
+[[package]]
+name = "itoa"
+version = "1.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
+
+[[package]]
+name = "js-sys"
+version = "0.3.100"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2025f20d7a4fa7785846e7b63d10a76d3f1cee98ee5cb79ea59703f95e42162"
+dependencies = [
+ "cfg-if",
+ "futures-util",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "lazy_static"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
+
+[[package]]
+name = "leb128fmt"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09edd9e8b54e49e587e4f6295a7d29c3ea94d469cb40ab8ca70b288248a81db2"
+
+[[package]]
+name = "libc"
+version = "0.2.186"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66"
+
+[[package]]
+name = "libredox"
+version = "0.1.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f02ab6bace2054fb888a3c16f990117b579d14a3088e472d63c6011fa185c9d3"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "linked-hash-map"
+version = "0.5.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0717cef1bc8b636c6e1c1bbdefc09e6322da8a9321966e8928ef80d20f7f770f"
+
+[[package]]
+name = "linux-raw-sys"
+version = "0.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32a66949e030da00e8c7d4434b251670a91556f4144941d37452769c25d58a53"
+
+[[package]]
+name = "litemap"
+version = "0.8.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92daf443525c4cce67b150400bc2316076100ce0b3686209eb8cf3c31612e6f0"
+
+[[package]]
+name = "log"
+version = "0.4.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "953f07c43838f8e6f9758cab68bf5bed85465e7587ebe0b823f1bcd81978ad3a"
+
+[[package]]
+name = "lru-slab"
+version = "0.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "112b39cec0b298b6c1999fee3e31427f74f676e4cb9879ed1a121b43661a4154"
+
+[[package]]
+name = "matchers"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d1525a2a28c7f4fa0fc98bb91ae755d1e2d1505079e05539e35bc876b5d65ae9"
+dependencies = [
+ "regex-automata",
+]
+
+[[package]]
+name = "memchr"
+version = "2.8.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "88904434abc2901f197fe8cc55f0445e7ded921dba5911dad2e2b39b48e663c4"
+
+[[package]]
+name = "mio"
+version = "1.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "02bd0af71c67b473010cbbc60715ee815645a4dc942899111f494b4b737d6fda"
+dependencies = [
+ "libc",
+ "wasi",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "nom"
+version = "8.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df9761775871bdef83bee530e60050f7e54b1105350d6884eb0fb4f46c2f9405"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "nom_locate"
+version = "5.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b577e2d69827c4740cba2b52efaad1c4cc7c73042860b199710b3575c68438d"
+dependencies = [
+ "bytecount",
+ "memchr",
+ "nom",
+]
+
+[[package]]
+name = "ntapi"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3b335231dfd352ffb0f8017f3b6027a4917f7df785ea2143d8af2adc66980ae"
+dependencies = [
+ "winapi",
+]
+
+[[package]]
+name = "nu-ansi-term"
+version = "0.50.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "objc2-core-foundation"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2a180dd8642fa45cdb7dd721cd4c11b1cadd4929ce112ebd8b9f5803cc79d536"
+dependencies = [
+ "bitflags",
+]
+
+[[package]]
+name = "objc2-io-kit"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "33fafba39597d6dc1fb709123dfa8289d39406734be322956a69f0931c73bb15"
+dependencies = [
+ "libc",
+ "objc2-core-foundation",
+]
+
+[[package]]
+name = "once_cell"
+version = "1.21.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
+
+[[package]]
+name = "once_cell_polyfill"
+version = "1.70.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
+
+[[package]]
+name = "openssl-probe"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7c87def4c32ab89d880effc9e097653c8da5d6ef28e6b539d313baaacfbafcbe"
+
+[[package]]
+name = "opentelemetry"
+version = "0.31.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b84bcd6ae87133e903af7ef497404dda70c60d0ea14895fc8a5e6722754fc2a0"
+dependencies = [
+ "futures-core",
+ "futures-sink",
+ "js-sys",
+ "pin-project-lite",
+ "thiserror 2.0.18",
+ "tracing",
+]
+
+[[package]]
+name = "opentelemetry-http"
+version = "0.31.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d7a6d09a73194e6b66df7c8f1b680f156d916a1a942abf2de06823dd02b7855d"
+dependencies = [
+ "async-trait",
+ "bytes",
+ "http",
+ "opentelemetry",
+ "reqwest",
+]
+
+[[package]]
+name = "opentelemetry_sdk"
+version = "0.31.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e14ae4f5991976fd48df6d843de219ca6d31b01daaab2dad5af2badeded372bd"
+dependencies = [
+ "futures-channel",
+ "futures-executor",
+ "futures-util",
+ "opentelemetry",
+ "percent-encoding",
+ "rand",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-stream",
+]
+
+[[package]]
+name = "option-ext"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "04744f49eae99ab78e0d5c0b603ab218f515ea8cfe5a456d7629ad883a3b6e7d"
+
+[[package]]
+name = "peg"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f76678828272f177ac33b7e2ac2e3e73cc6c1cd1e3e387928aa69562fa51367"
+dependencies = [
+ "peg-macros",
+ "peg-runtime",
+]
+
+[[package]]
+name = "peg-macros"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "636d60acf97633e48d266d7415a9355d4389cea327a193f87df395d88cd2b14d"
+dependencies = [
+ "peg-runtime",
+ "proc-macro2",
+ "quote",
+]
+
+[[package]]
+name = "peg-runtime"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9555b1514d2d99d78150d3c799d4c357a3e2c2a8062cd108e93a06d9057629c5"
+
+[[package]]
+name = "percent-encoding"
+version = "2.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9b4f627cb1b25917193a259e49bdad08f671f8d9708acfd5fe0a8c1455d87220"
+
+[[package]]
+name = "pin-project"
+version = "1.1.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2466b2336ed02bcdca6b294417127b90ec92038d1d5c4fbeac971a922e0e0924"
+dependencies = [
+ "pin-project-internal",
+]
+
+[[package]]
+name = "pin-project-internal"
+version = "1.1.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c96395f0a926bc13b1c17622aaddda1ecb55d49c8f1bf9777e4d877800a43f8b"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "pin-project-lite"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"
+
+[[package]]
+name = "potential_utf"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0103b1cef7ec0cf76490e969665504990193874ea05c85ff9bab8b911d0a0564"
+dependencies = [
+ "zerovec",
+]
+
+[[package]]
+name = "ppv-lite86"
+version = "0.2.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85eae3c4ed2f50dcfe72643da4befc30deadb458a9b590d720cde2f2b1e97da9"
+dependencies = [
+ "zerocopy",
+]
+
+[[package]]
+name = "prettyplease"
+version = "0.2.37"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
+dependencies = [
+ "proc-macro2",
+ "syn",
+]
+
+[[package]]
+name = "proc-macro2"
+version = "1.0.106"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "quinn"
+version = "0.11.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9e20a958963c291dc322d98411f541009df2ced7b5a4f2bd52337638cfccf20"
+dependencies = [
+ "bytes",
+ "cfg_aliases",
+ "pin-project-lite",
+ "quinn-proto",
+ "quinn-udp",
+ "rustc-hash",
+ "rustls",
+ "socket2",
+ "thiserror 2.0.18",
+ "tokio",
+ "tracing",
+ "web-time",
+]
+
+[[package]]
+name = "quinn-proto"
+version = "0.11.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "434b42fec591c96ef50e21e886936e66d3cc3f737104fdb9b737c40ffb94c098"
+dependencies = [
+ "bytes",
+ "getrandom 0.3.4",
+ "lru-slab",
+ "rand",
+ "ring",
+ "rustc-hash",
+ "rustls",
+ "rustls-pki-types",
+ "slab",
+ "thiserror 2.0.18",
+ "tinyvec",
+ "tracing",
+ "web-time",
+]
+
+[[package]]
+name = "quinn-udp"
+version = "0.5.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "addec6a0dcad8a8d96a771f815f0eaf55f9d1805756410b39f5fa81332574cbd"
+dependencies = [
+ "cfg_aliases",
+ "libc",
+ "once_cell",
+ "socket2",
+ "tracing",
+ "windows-sys 0.60.2",
+]
+
+[[package]]
+name = "quote"
+version = "1.0.45"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "41f2619966050689382d2b44f664f4bc593e129785a36d6ee376ddf37259b924"
+dependencies = [
+ "proc-macro2",
+]
+
+[[package]]
+name = "r-efi"
+version = "5.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "69cdb34c158ceb288df11e18b4bd39de994f6657d83847bdffdbd7f346754b0f"
+
+[[package]]
+name = "r-efi"
+version = "6.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
+
+[[package]]
+name = "rand"
+version = "0.9.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "44c5af06bb1b7d3216d91932aed5265164bf384dc89cd6ba05cf59a35f5f76ea"
+dependencies = [
+ "rand_chacha",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_chacha"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
+dependencies = [
+ "ppv-lite86",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_core"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "76afc826de14238e6e8c374ddcc1fa19e374fd8dd986b0d2af0d02377261d83c"
+dependencies = [
+ "getrandom 0.3.4",
+]
+
+[[package]]
+name = "redox_users"
+version = "0.4.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ba009ff324d1fc1b900bd1fdb31564febe58a8ccc8a6fdbb93b543d33b13ca43"
+dependencies = [
+ "getrandom 0.2.17",
+ "libredox",
+ "thiserror 1.0.69",
+]
+
+[[package]]
+name = "ref-cast"
+version = "1.0.25"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f354300ae66f76f1c85c5f84693f0ce81d747e2c3f21a45fef496d89c960bf7d"
+dependencies = [
+ "ref-cast-impl",
+]
+
+[[package]]
+name = "ref-cast-impl"
+version = "1.0.25"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b7186006dcb21920990093f30e3dea63b7d6e977bf1256be20c3563a5db070da"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "regex"
+version = "1.12.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f1292b7759ae1cb9ec195452d1390a074f0cd8541ab7a5a8c31cd6db45d4a6ba"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-automata"
+version = "0.4.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e1dd4122fc1595e8162618945476892eefca7b88c52820e74af6262213cae8f"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-syntax"
+version = "0.8.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
+
+[[package]]
+name = "reqwest"
+version = "0.12.28"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
+dependencies = [
+ "base64",
+ "bytes",
+ "futures-core",
+ "http",
+ "http-body",
+ "http-body-util",
+ "hyper",
+ "hyper-rustls",
+ "hyper-util",
+ "js-sys",
+ "log",
+ "percent-encoding",
+ "pin-project-lite",
+ "quinn",
+ "rustls",
+ "rustls-pki-types",
+ "serde",
+ "serde_json",
+ "serde_urlencoded",
+ "sync_wrapper",
+ "tokio",
+ "tokio-rustls",
+ "tower",
+ "tower-http",
+ "tower-service",
+ "url",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+ "webpki-roots",
+]
+
+[[package]]
+name = "ring"
+version = "0.17.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4689e6c2294d81e88dc6261c768b63bc4fcdb852be6d1352498b114f61383b7"
+dependencies = [
+ "cc",
+ "cfg-if",
+ "getrandom 0.2.17",
+ "libc",
+ "untrusted",
+ "windows-sys 0.52.0",
+]
+
+[[package]]
+name = "rustc-hash"
+version = "2.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "94300abf3f1ae2e2b8ffb7b58043de3d399c73fa6f4b73826402a5c457614dbe"
+
+[[package]]
+name = "rustc_version"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cfcb3a22ef46e85b45de6ee7e79d063319ebb6594faafcf1c225ea92ab6e9b92"
+dependencies = [
+ "semver",
+]
+
+[[package]]
+name = "rustix"
+version = "1.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6fe4565b9518b83ef4f91bb47ce29620ca828bd32cb7e408f0062e9930ba190"
+dependencies = [
+ "bitflags",
+ "errno",
+ "libc",
+ "linux-raw-sys",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "rustls"
+version = "0.23.40"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ef86cd5876211988985292b91c96a8f2d298df24e75989a43a3c73f2d4d8168b"
+dependencies = [
+ "once_cell",
+ "ring",
+ "rustls-pki-types",
+ "rustls-webpki",
+ "subtle",
+ "zeroize",
+]
+
+[[package]]
+name = "rustls-native-certs"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dab5152771c58876a2146916e53e35057e1a4dfa2b9df0f0305b07f611fdea4d"
+dependencies = [
+ "openssl-probe",
+ "rustls-pki-types",
+ "schannel",
+ "security-framework",
+]
+
+[[package]]
+name = "rustls-pki-types"
+version = "1.14.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "30a7197ae7eb376e574fe940d068c30fe0462554a3ddbe4eca7838e049c937a9"
+dependencies = [
+ "web-time",
+ "zeroize",
+]
+
+[[package]]
+name = "rustls-webpki"
+version = "0.103.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
+dependencies = [
+ "ring",
+ "rustls-pki-types",
+ "untrusted",
+]
+
+[[package]]
+name = "rustversion"
+version = "1.0.22"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b39cdef0fa800fc44525c84ccb54a029961a8215f9619753635a9c0d2538d46d"
+
+[[package]]
+name = "ryu"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
+
+[[package]]
+name = "same-file"
+version = "1.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93fc1dc3aaa9bfed95e02e6eadabb4baf7e3078b0bd1b4d7b6b0b68378900502"
+dependencies = [
+ "winapi-util",
+]
+
+[[package]]
+name = "schannel"
+version = "0.1.29"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "91c1b7e4904c873ef0710c1f407dde2e6287de2bebc1bbbf7d430bb7cbffd939"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "schemars"
+version = "0.8.22"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3fbf2ae1b8bc8e02df939598064d22402220cd5bbcca1c76f7d6a310974d5615"
+dependencies = [
+ "dyn-clone",
+ "schemars_derive",
+ "serde",
+ "serde_json",
+]
+
+[[package]]
+name = "schemars_derive"
+version = "0.8.22"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32e265784ad618884abaea0600a9adf15393368d840e0222d101a072f3f7534d"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "serde_derive_internals",
+ "syn",
+]
+
+[[package]]
+name = "sealed"
+version = "0.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "22f968c5ea23d555e670b449c1c5e7b2fc399fdaec1d304a17cd48e288abc107"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "security-framework"
+version = "3.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b7f4bc775c73d9a02cde8bf7b2ec4c9d12743edf609006c7facc23998404cd1d"
+dependencies = [
+ "bitflags",
+ "core-foundation",
+ "core-foundation-sys",
+ "libc",
+ "security-framework-sys",
+]
+
+[[package]]
+name = "security-framework-sys"
+version = "2.17.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6ce2691df843ecc5d231c0b14ece2acc3efb62c0a398c7e1d875f3983ce020e3"
+dependencies = [
+ "core-foundation-sys",
+ "libc",
+]
+
+[[package]]
+name = "semver"
+version = "1.0.28"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8a7852d02fc848982e0c167ef163aaff9cd91dc640ba85e263cb1ce46fae51cd"
+
+[[package]]
+name = "serde"
+version = "1.0.228"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9a8e94ea7f378bd32cbbd37198a4a91436180c5bb472411e48b5ec2e2124ae9e"
+dependencies = [
+ "serde_core",
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_core"
+version = "1.0.228"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "41d385c7d4ca58e59fc732af25c3983b67ac852c1a25000afe1175de458b67ad"
+dependencies = [
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_derive"
+version = "1.0.228"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "serde_derive_internals"
+version = "0.29.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "18d26a20a969b9e3fdf2fc2d9f21eda6c40e2de84c9408bb5d3b05d499aae711"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "serde_json"
+version = "1.0.150"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e8014e44b4736ed0538adeecded0fce2a272f22dc9578a7eb6b2d9993c74cfb9"
+dependencies = [
+ "itoa",
+ "memchr",
+ "serde",
+ "serde_core",
+ "zmij",
+]
+
+[[package]]
+name = "serde_urlencoded"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3491c14715ca2294c4d6a88f15e84739788c1d030eed8c110436aafdaa2f3fd"
+dependencies = [
+ "form_urlencoded",
+ "itoa",
+ "ryu",
+ "serde",
+]
+
+[[package]]
+name = "serde_yaml"
+version = "0.9.34+deprecated"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6a8b1a1a2ebf674015cc02edccce75287f1a0130d394307b36743c2f5d504b47"
+dependencies = [
+ "indexmap",
+ "itoa",
+ "ryu",
+ "serde",
+ "unsafe-libyaml",
+]
+
+[[package]]
+name = "sha1"
+version = "0.10.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e3bf829a2d51ab4a5ddf1352d8470c140cadc8301b2ae1789db023f01cedd6ba"
+dependencies = [
+ "cfg-if",
+ "cpufeatures",
+ "digest",
+]
+
+[[package]]
+name = "sha2"
+version = "0.10.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
+dependencies = [
+ "cfg-if",
+ "cpufeatures",
+ "digest",
+]
+
+[[package]]
+name = "sharded-slab"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f40ca3c46823713e0d4209592e8d6e826aa57e928f09752619fc696c499637f6"
+dependencies = [
+ "lazy_static",
+]
+
+[[package]]
+name = "shlex"
+version = "2.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8fadd59c855ef2080decdef8ff161eb6661b86933c9d82e5ba29dc602a55aba"
+
+[[package]]
+name = "signal-hook-registry"
+version = "1.4.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c4db69cba1110affc0e9f7bcd48bbf87b3f4fc7c61fc9155afd4c469eb3d6c1b"
+dependencies = [
+ "errno",
+ "libc",
+]
+
+[[package]]
+name = "slab"
+version = "0.4.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c790de23124f9ab44544d7ac05d60440adc586479ce501c1d6d7da3cd8c9cf5"
+
+[[package]]
+name = "smallvec"
+version = "1.15.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ed6a63f02c8539c91a8685a86f4099661ba3da017932f6ebbea6de3f0fa7c90"
+
+[[package]]
+name = "smart-default"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0eb01866308440fc64d6c44d9e86c5cc17adfe33c4d6eed55da9145044d0ffc1"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "smawk"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b7c388c1b5e93756d0c740965c41e8822f866621d41acbdf6336a6a168f8840c"
+
+[[package]]
+name = "socket2"
+version = "0.6.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52d1cfed4120b4d927bf7c0f86d2087a4a7d6027c906d9f9d525a80573b9be51"
+dependencies = [
+ "libc",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "stable_deref_trait"
+version = "1.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6ce2be8dc25455e1f91df71bfa12ad37d7af1092ae736f3a6cd0e37bc7810596"
+
+[[package]]
+name = "strsim"
+version = "0.11.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7da8b5736845d9f2fcb837ea5d9e2628564b3b043a70948a3f0b778838c5fb4f"
+
+[[package]]
+name = "subtle"
+version = "2.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
+
+[[package]]
+name = "syn"
+version = "2.0.117"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e665b8803e7b1d2a727f4023456bbbbe74da67099c585258af0ad9c5013b9b99"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "sync_wrapper"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0bf256ce5efdfa370213c1dabab5935a12e49f2c58d15e9eac2870d3b4f27263"
+dependencies = [
+ "futures-core",
+]
+
+[[package]]
+name = "synstructure"
+version = "0.13.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "728a70f3dbaf5bab7f0c4b1ac8d7ae5ea60a4b5549c8a5914361c99147a709d2"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "synthez"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d8a928f38f1bc873f28e0d2ba8298ad65374a6ac2241dabd297271531a736cd"
+dependencies = [
+ "syn",
+ "synthez-codegen",
+ "synthez-core",
+]
+
+[[package]]
+name = "synthez-codegen"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8fb83b8df4238e11746984dfb3819b155cd270de0e25847f45abad56b3671047"
+dependencies = [
+ "syn",
+ "synthez-core",
+]
+
+[[package]]
+name = "synthez-core"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "906fba967105d822e7c7ed60477b5e76116724d33de68a585681fb253fc30d5c"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "sealed",
+ "syn",
+]
+
+[[package]]
+name = "sysinfo"
+version = "0.38.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "92ab6a2f8bfe508deb3c6406578252e491d299cbbf3bc0529ecc3313aee4a52f"
+dependencies = [
+ "libc",
+ "memchr",
+ "ntapi",
+ "objc2-core-foundation",
+ "objc2-io-kit",
+ "windows",
+]
+
+[[package]]
+name = "tempfile"
+version = "3.27.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
+dependencies = [
+ "fastrand",
+ "getrandom 0.4.2",
+ "once_cell",
+ "rustix",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "terminal_size"
+version = "0.4.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "230a1b821ccbd75b185820a1f1ff7b14d21da1e442e22c0863ea5f08771a8874"
+dependencies = [
+ "rustix",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "textwrap"
+version = "0.16.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c13547615a44dc9c452a8a534638acdf07120d4b6847c8178705da06306a3057"
+dependencies = [
+ "smawk",
+ "unicode-linebreak",
+ "unicode-width",
+]
+
+[[package]]
+name = "thiserror"
+version = "1.0.69"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6aaf5339b578ea85b50e080feb250a3e8ae8cfcdff9a461c9ec2904bc923f52"
+dependencies = [
+ "thiserror-impl 1.0.69",
+]
+
+[[package]]
+name = "thiserror"
+version = "2.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4288b5bcbc7920c07a1149a35cf9590a2aa808e0bc1eafaade0b80947865fbc4"
+dependencies = [
+ "thiserror-impl 2.0.18",
+]
+
+[[package]]
+name = "thiserror-impl"
+version = "1.0.69"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4fee6c4efc90059e10f81e6d42c60a18f76588c3d74cb83a0b242a2b6c7504c1"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "thiserror-impl"
+version = "2.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "thread_local"
+version = "1.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f60246a4944f24f6e018aa17cdeffb7818b76356965d03b07d6a9886e8962185"
+dependencies = [
+ "cfg-if",
+]
+
+[[package]]
+name = "tinystr"
+version = "0.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c8323304221c2a851516f22236c5722a72eaa19749016521d6dff0824447d96d"
+dependencies = [
+ "displaydoc",
+ "zerovec",
+]
+
+[[package]]
+name = "tinyvec"
+version = "1.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3e61e67053d25a4e82c844e8424039d9745781b3fc4f32b8d55ed50f5f667ef3"
+dependencies = [
+ "tinyvec_macros",
+]
+
+[[package]]
+name = "tinyvec_macros"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
+
+[[package]]
+name = "tokio"
+version = "1.52.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8fc7f01b389ac15039e4dc9531aa973a135d7a4135281b12d7c1bc79fd57fffe"
+dependencies = [
+ "bytes",
+ "libc",
+ "mio",
+ "pin-project-lite",
+ "signal-hook-registry",
+ "socket2",
+ "tokio-macros",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "tokio-macros"
+version = "2.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "385a6cb71ab9ab790c5fe8d67f1645e6c450a7ce006a33de03daa956cf70a496"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "tokio-rustls"
+version = "0.26.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1729aa945f29d91ba541258c8df89027d5792d85a8841fb65e8bf0f4ede4ef61"
+dependencies = [
+ "rustls",
+ "tokio",
+]
+
+[[package]]
+name = "tokio-stream"
+version = "0.1.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32da49809aab5c3bc678af03902d4ccddea2a87d028d86392a4b1560c6906c70"
+dependencies = [
+ "futures-core",
+ "pin-project-lite",
+ "tokio",
+]
+
+[[package]]
+name = "tokio-tungstenite"
+version = "0.28.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d25a406cddcc431a75d3d9afc6a7c0f7428d4891dd973e4d54c56b46127bf857"
+dependencies = [
+ "futures-util",
+ "log",
+ "rustls",
+ "rustls-native-certs",
+ "rustls-pki-types",
+ "tokio",
+ "tokio-rustls",
+ "tungstenite",
+]
+
+[[package]]
+name = "tower"
+version = "0.5.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebe5ef63511595f1344e2d5cfa636d973292adc0eec1f0ad45fae9f0851ab1d4"
+dependencies = [
+ "futures-core",
+ "futures-util",
+ "pin-project-lite",
+ "sync_wrapper",
+ "tokio",
+ "tower-layer",
+ "tower-service",
+]
+
+[[package]]
+name = "tower-http"
+version = "0.6.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4cfcf7e2740e6fc6d4d688b4ef00650406bb94adf4731e43c096c3a19fe40840"
+dependencies = [
+ "bitflags",
+ "bytes",
+ "futures-util",
+ "http",
+ "http-body",
+ "pin-project-lite",
+ "tower",
+ "tower-layer",
+ "tower-service",
+ "url",
+]
+
+[[package]]
+name = "tower-layer"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "121c2a6cda46980bb0fcd1647ffaf6cd3fc79a013de288782836f6df9c48780e"
+
+[[package]]
+name = "tower-service"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
+
+[[package]]
+name = "tracing"
+version = "0.1.44"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63e71662fa4b2a2c3a26f570f037eb95bb1f85397f3cd8076caed2f026a6d100"
+dependencies = [
+ "pin-project-lite",
+ "tracing-attributes",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-attributes"
+version = "0.1.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "tracing-core"
+version = "0.1.36"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
+dependencies = [
+ "once_cell",
+ "valuable",
+]
+
+[[package]]
+name = "tracing-log"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ee855f1f400bd0e5c02d150ae5de3840039a3f54b025156404e34c23c03f47c3"
+dependencies = [
+ "log",
+ "once_cell",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-subscriber"
+version = "0.3.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb7f578e5945fb242538965c2d0b04418d38ec25c79d160cd279bf0731c8d319"
+dependencies = [
+ "matchers",
+ "nu-ansi-term",
+ "once_cell",
+ "regex-automata",
+ "sharded-slab",
+ "smallvec",
+ "thread_local",
+ "tracing",
+ "tracing-core",
+ "tracing-log",
+]
+
+[[package]]
+name = "try-lock"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b"
+
+[[package]]
+name = "tungstenite"
+version = "0.28.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8628dcc84e5a09eb3d8423d6cb682965dea9133204e8fb3efee74c2a0c259442"
+dependencies = [
+ "bytes",
+ "data-encoding",
+ "http",
+ "httparse",
+ "log",
+ "rand",
+ "rustls",
+ "rustls-pki-types",
+ "sha1",
+ "thiserror 2.0.18",
+ "utf-8",
+]
+
+[[package]]
+name = "typed-builder"
+version = "0.23.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "31aa81521b70f94402501d848ccc0ecaa8f93c8eb6999eb9747e72287757ffda"
+dependencies = [
+ "typed-builder-macro",
+]
+
+[[package]]
+name = "typed-builder-macro"
+version = "0.23.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "076a02dc54dd46795c2e9c8282ed40bcfb1e22747e955de9389a1de28190fb26"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "typenum"
+version = "1.20.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
+
+[[package]]
+name = "unicode-ident"
+version = "1.0.24"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6e4313cd5fcd3dad5cafa179702e2b244f760991f45397d14d4ebf38247da75"
+
+[[package]]
+name = "unicode-linebreak"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3b09c83c3c29d37506a3e260c08c03743a6bb66a9cd432c6934ab501a190571f"
+
+[[package]]
+name = "unicode-segmentation"
+version = "1.13.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c6f5d3c3b1bf09027a88a6bc961fc00497d651009560b5463668dc81b0fa87a8"
+
+[[package]]
+name = "unicode-width"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b4ac048d71ede7ee76d585517add45da530660ef4390e49b098733c6e897f254"
+
+[[package]]
+name = "unicode-xid"
+version = "0.2.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebc1c04c71510c7f702b52b7c350734c9ff1295c464a03335b00bb84fc54f853"
+
+[[package]]
+name = "unsafe-libyaml"
+version = "0.2.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "673aac59facbab8a9007c7f6108d11f63b603f7cabff99fabf650fea5c32b861"
+
+[[package]]
+name = "untrusted"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"
+
+[[package]]
+name = "url"
+version = "2.5.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ff67a8a4397373c3ef660812acab3268222035010ab8680ec4215f38ba3d0eed"
+dependencies = [
+ "form_urlencoded",
+ "idna",
+ "percent-encoding",
+ "serde",
+]
+
+[[package]]
+name = "utf-8"
+version = "0.7.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"
+
+[[package]]
+name = "utf8_iter"
+version = "1.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6c140620e7ffbb22c2dee59cafe6084a59b5ffc27a8859a5f0d494b5d52b6be"
+
+[[package]]
+name = "utf8parse"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
+
+[[package]]
+name = "uuid"
+version = "1.23.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "144d6b123cef80b301b8f72a9e2ca4370ddec21950d0a103dd22c437006d2db7"
+dependencies = [
+ "getrandom 0.4.2",
+ "js-sys",
+ "serde_core",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "valuable"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ba73ea9cf16a25df0c8caa16c51acb937d5712a8429db78a3ee29d5dcacd3a65"
+
+[[package]]
+name = "version_check"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
+
+[[package]]
+name = "walkdir"
+version = "2.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29790946404f91d9c5d06f9874efddea1dc06c5efe94541a7d6863108e3a5e4b"
+dependencies = [
+ "same-file",
+ "winapi-util",
+]
+
+[[package]]
+name = "want"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bfa7760aed19e106de2c7c0b581b509f2f25d3dacaf737cb82ac61bc6d760b0e"
+dependencies = [
+ "try-lock",
+]
+
+[[package]]
+name = "wasi"
+version = "0.11.1+wasi-snapshot-preview1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
+
+[[package]]
+name = "wasip2"
+version = "1.0.4+wasi-0.2.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b67efb37e106e55ce722a510d6b5f9c17f083e5fc79afc2badeb12cc313d9487"
+dependencies = [
+ "wit-bindgen 0.57.1",
+]
+
+[[package]]
+name = "wasip3"
+version = "0.4.0+wasi-0.3.0-rc-2026-01-06"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5428f8bf88ea5ddc08faddef2ac4a67e390b88186c703ce6dbd955e1c145aca5"
+dependencies = [
+ "wit-bindgen 0.51.0",
+]
+
+[[package]]
+name = "wasm-bindgen"
+version = "0.2.123"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a254a4b10c19a76f09a27640e7ffbf9bc30bf67e16a3bf28aaefa4920fe81563"
+dependencies = [
+ "cfg-if",
+ "once_cell",
+ "rustversion",
+ "wasm-bindgen-macro",
+ "wasm-bindgen-shared",
+]
+
+[[package]]
+name = "wasm-bindgen-futures"
+version = "0.4.73"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "54568702fabf5d4849ce2b90fadfa64168a097eaf4b351ce9df8b687a0086aaf"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "wasm-bindgen-macro"
+version = "0.2.123"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "24a40fc75b0ec6f3746ceb10d36f53a93dcd68a93b11b6445983945d79eba0dc"
+dependencies = [
+ "quote",
+ "wasm-bindgen-macro-support",
+]
+
+[[package]]
+name = "wasm-bindgen-macro-support"
+version = "0.2.123"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "908f34bd9b9ce3d4caf07b72dfab63d61504d156856c6bd3cd87fa350cf3985b"
+dependencies = [
+ "bumpalo",
+ "proc-macro2",
+ "quote",
+ "syn",
+ "wasm-bindgen-shared",
+]
+
+[[package]]
+name = "wasm-bindgen-shared"
+version = "0.2.123"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7acbf7616c27b194bbb550bf77ed0c2c3e5b7fd1260a93082b95fb7f47959b92"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "wasm-encoder"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "990065f2fe63003fe337b932cfb5e3b80e0b4d0f5ff650e6985b1048f62c8319"
+dependencies = [
+ "leb128fmt",
+ "wasmparser",
+]
+
+[[package]]
+name = "wasm-metadata"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bb0e353e6a2fbdc176932bbaab493762eb1255a7900fe0fea1a2f96c296cc909"
+dependencies = [
+ "anyhow",
+ "indexmap",
+ "wasm-encoder",
+ "wasmparser",
+]
+
+[[package]]
+name = "wasmparser"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
+dependencies = [
+ "bitflags",
+ "hashbrown 0.15.5",
+ "indexmap",
+ "semver",
+]
+
+[[package]]
+name = "web-sys"
+version = "0.3.100"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e0871acf327f283dc6da28a1696cdc64fb355ba9f935d052021fa77f35cce69"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "web-time"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5a6580f308b1fad9207618087a65c04e7a10bc77e02c8e84e9b00dd4b12fa0bb"
+dependencies = [
+ "js-sys",
+ "wasm-bindgen",
+]
+
+[[package]]
+name = "webpki-roots"
+version = "1.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52f5ee44c96cf55f1b349600768e3ece3a8f26010c05265ab73f945bb1a2eb9d"
+dependencies = [
+ "rustls-pki-types",
+]
+
+[[package]]
+name = "which"
+version = "8.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c789537cf2f7f55be8e6192f92e464174ee55f91af622777f7f1ceb0dbccd03e"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "winapi"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5c839a674fcd7a98952e593242ea400abe93992746761e38641405d28b00f419"
+dependencies = [
+ "winapi-i686-pc-windows-gnu",
+ "winapi-x86_64-pc-windows-gnu",
+]
+
+[[package]]
+name = "winapi-i686-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac3b87c63620426dd9b991e5ce0329eff545bccbbb34f3be09ff6fb6ab51b7b6"
+
+[[package]]
+name = "winapi-util"
+version = "0.1.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "winapi-x86_64-pc-windows-gnu"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
+
+[[package]]
+name = "windows"
+version = "0.62.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "527fadee13e0c05939a6a05d5bd6eec6cd2e3dbd648b9f8e447c6518133d8580"
+dependencies = [
+ "windows-collections",
+ "windows-core",
+ "windows-future",
+ "windows-numerics",
+]
+
+[[package]]
+name = "windows-collections"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "23b2d95af1a8a14a3c7367e1ed4fc9c20e0a26e79551b1454d72583c97cc6610"
+dependencies = [
+ "windows-core",
+]
+
+[[package]]
+name = "windows-core"
+version = "0.62.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
+dependencies = [
+ "windows-implement",
+ "windows-interface",
+ "windows-link",
+ "windows-result",
+ "windows-strings",
+]
+
+[[package]]
+name = "windows-future"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1d6f90251fe18a279739e78025bd6ddc52a7e22f921070ccdc67dde84c605cb"
+dependencies = [
+ "windows-core",
+ "windows-link",
+ "windows-threading",
+]
+
+[[package]]
+name = "windows-implement"
+version = "0.60.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "windows-interface"
+version = "0.59.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "windows-link"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5"
+
+[[package]]
+name = "windows-numerics"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e2e40844ac143cdb44aead537bbf727de9b044e107a0f1220392177d15b0f26"
+dependencies = [
+ "windows-core",
+ "windows-link",
+]
+
+[[package]]
+name = "windows-result"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7781fa89eaf60850ac3d2da7af8e5242a5ea78d1a11c49bf2910bb5a73853eb5"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-strings"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.48.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "677d2418bec65e3338edb076e806bc1ec15693c5d0104683f2efe857f61056a9"
+dependencies = [
+ "windows-targets 0.48.5",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.52.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "282be5f36a8ce781fad8c8ae18fa3f9beff57ec1b52cb3de0789201425d9a33d"
+dependencies = [
+ "windows-targets 0.52.6",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.60.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2f500e4d28234f72040990ec9d39e3a6b950f9f22d3dba18416c35882612bcb"
+dependencies = [
+ "windows-targets 0.53.5",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.61.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ae137229bcbd6cdf0f7b80a31df61766145077ddf49416a728b02cb3921ff3fc"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-targets"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9a2fa6e2155d7247be68c096456083145c183cbbbc2764150dda45a87197940c"
+dependencies = [
+ "windows_aarch64_gnullvm 0.48.5",
+ "windows_aarch64_msvc 0.48.5",
+ "windows_i686_gnu 0.48.5",
+ "windows_i686_msvc 0.48.5",
+ "windows_x86_64_gnu 0.48.5",
+ "windows_x86_64_gnullvm 0.48.5",
+ "windows_x86_64_msvc 0.48.5",
+]
+
+[[package]]
+name = "windows-targets"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973"
+dependencies = [
+ "windows_aarch64_gnullvm 0.52.6",
+ "windows_aarch64_msvc 0.52.6",
+ "windows_i686_gnu 0.52.6",
+ "windows_i686_gnullvm 0.52.6",
+ "windows_i686_msvc 0.52.6",
+ "windows_x86_64_gnu 0.52.6",
+ "windows_x86_64_gnullvm 0.52.6",
+ "windows_x86_64_msvc 0.52.6",
+]
+
+[[package]]
+name = "windows-targets"
+version = "0.53.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4945f9f551b88e0d65f3db0bc25c33b8acea4d9e41163edf90dcd0b19f9069f3"
+dependencies = [
+ "windows-link",
+ "windows_aarch64_gnullvm 0.53.1",
+ "windows_aarch64_msvc 0.53.1",
+ "windows_i686_gnu 0.53.1",
+ "windows_i686_gnullvm 0.53.1",
+ "windows_i686_msvc 0.53.1",
+ "windows_x86_64_gnu 0.53.1",
+ "windows_x86_64_gnullvm 0.53.1",
+ "windows_x86_64_msvc 0.53.1",
+]
+
+[[package]]
+name = "windows-threading"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3949bd5b99cafdf1c7ca86b43ca564028dfe27d66958f2470940f73d86d75b37"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2b38e32f0abccf9987a4e3079dfb67dcd799fb61361e53e2882c3cbaf0d905d8"
+
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
+
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a9d8416fa8b42f5c947f8482c43e7d89e73a173cead56d044f6a56104a6d1b53"
+
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dc35310971f3b2dbbf3f0690a219f40e2d9afcf64f9ab7cc1be722937c26b4bc"
+
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
+
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9d782e804c2f632e395708e99a94275910eb9100b2114651e04744e9b125006"
+
+[[package]]
+name = "windows_i686_gnu"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a75915e7def60c94dcef72200b9a8e58e5091744960da64ec734a6c6e9b3743e"
+
+[[package]]
+name = "windows_i686_gnu"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
+
+[[package]]
+name = "windows_i686_gnu"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "960e6da069d81e09becb0ca57a65220ddff016ff2d6af6a223cf372a506593a3"
+
+[[package]]
+name = "windows_i686_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
+
+[[package]]
+name = "windows_i686_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fa7359d10048f68ab8b09fa71c3daccfb0e9b559aed648a8f95469c27057180c"
+
+[[package]]
+name = "windows_i686_msvc"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f55c233f70c4b27f66c523580f78f1004e8b5a8b659e05a4eb49d4166cca406"
+
+[[package]]
+name = "windows_i686_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
+
+[[package]]
+name = "windows_i686_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e7ac75179f18232fe9c285163565a57ef8d3c89254a30685b57d83a38d326c2"
+
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "53d40abd2583d23e4718fddf1ebec84dbff8381c07cae67ff7768bbf19c6718e"
+
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
+
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9c3842cdd74a865a8066ab39c8a7a473c0778a3f29370b5fd6b4b9aa7df4a499"
+
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b7b52767868a23d5bab768e390dc5f5c55825b6d30b86c844ff2dc7414044cc"
+
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
+
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ffa179e2d07eee8ad8f57493436566c7cc30ac536a3379fdf008f47f6bb7ae1"
+
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed94fce61571a4006852b7389a063ab983c02eb1bb37b47f8272ce92d06d9538"
+
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
+
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
+
+[[package]]
+name = "wit-bindgen"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d7249219f66ced02969388cf2bb044a09756a083d0fab1e566056b04d9fbcaa5"
+dependencies = [
+ "wit-bindgen-rust-macro",
+]
+
+[[package]]
+name = "wit-bindgen"
+version = "0.57.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ebf944e87a7c253233ad6766e082e3cd714b5d03812acc24c318f549614536e"
+
+[[package]]
+name = "wit-bindgen-core"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ea61de684c3ea68cb082b7a88508a8b27fcc8b797d738bfc99a82facf1d752dc"
+dependencies = [
+ "anyhow",
+ "heck",
+ "wit-parser",
+]
+
+[[package]]
+name = "wit-bindgen-rust"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b7c566e0f4b284dd6561c786d9cb0142da491f46a9fbed79ea69cdad5db17f21"
+dependencies = [
+ "anyhow",
+ "heck",
+ "indexmap",
+ "prettyplease",
+ "syn",
+ "wasm-metadata",
+ "wit-bindgen-core",
+ "wit-component",
+]
+
+[[package]]
+name = "wit-bindgen-rust-macro"
+version = "0.51.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c0f9bfd77e6a48eccf51359e3ae77140a7f50b1e2ebfe62422d8afdaffab17a"
+dependencies = [
+ "anyhow",
+ "prettyplease",
+ "proc-macro2",
+ "quote",
+ "syn",
+ "wit-bindgen-core",
+ "wit-bindgen-rust",
+]
+
+[[package]]
+name = "wit-component"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
+dependencies = [
+ "anyhow",
+ "bitflags",
+ "indexmap",
+ "log",
+ "serde",
+ "serde_derive",
+ "serde_json",
+ "wasm-encoder",
+ "wasm-metadata",
+ "wasmparser",
+ "wit-parser",
+]
+
+[[package]]
+name = "wit-parser"
+version = "0.244.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ecc8ac4bc1dc3381b7f59c34f00b67e18f910c2c0f50015669dde7def656a736"
+dependencies = [
+ "anyhow",
+ "id-arena",
+ "indexmap",
+ "log",
+ "semver",
+ "serde",
+ "serde_derive",
+ "serde_json",
+ "unicode-xid",
+ "wasmparser",
+]
+
+[[package]]
+name = "writeable"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ffae5123b2d3fc086436f8834ae3ab053a283cfac8fe0a0b8eaae044768a4c4"
+
+[[package]]
+name = "yoke"
+version = "0.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "709fe23a0424b6a435d82152b1bd3fdfb0833487d5fa90d05d42762a9891fef5"
+dependencies = [
+ "stable_deref_trait",
+ "yoke-derive",
+ "zerofrom",
+]
+
+[[package]]
+name = "yoke-derive"
+version = "0.8.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "de844c262c8848816172cef550288e7dc6c7b7814b4ee56b3e1553f275f1858e"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+ "synstructure",
+]
+
+[[package]]
+name = "zerocopy"
+version = "0.8.52"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ce1022995ff5ff5d841ad7d994facc23098cd40152f2c1d11cd607c6f530653f"
+dependencies = [
+ "zerocopy-derive",
+]
+
+[[package]]
+name = "zerocopy-derive"
+version = "0.8.52"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ae7f38b72ec2a254e2b87ef277cf2cd4fb97cbebf944faa6f33354da0867930"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "zerofrom"
+version = "0.1.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ec05a11813ea801ff6d75110ad09cd0824ddba17dfe17128ea0d5f68e6c5272"
+dependencies = [
+ "zerofrom-derive",
+]
+
+[[package]]
+name = "zerofrom-derive"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "11532158c46691caf0f2593ea8358fed6bbf68a0315e80aae9bd41fbade684a1"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+ "synstructure",
+]
+
+[[package]]
+name = "zeroize"
+version = "1.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e13c156562582aa81c60cb29407084cdb54c4164760106ab78e6c5b0858cf64e"
+
+[[package]]
+name = "zerotrie"
+version = "0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0f9152d31db0792fa83f70fb2f83148effb5c1f5b8c7686c3459e361d9bc20bf"
+dependencies = [
+ "displaydoc",
+ "yoke",
+ "zerofrom",
+]
+
+[[package]]
+name = "zerovec"
+version = "0.11.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "90f911cbc359ab6af17377d242225f4d75119aec87ea711a880987b18cd7b239"
+dependencies = [
+ "yoke",
+ "zerofrom",
+ "zerovec-derive",
+]
+
+[[package]]
+name = "zerovec-derive"
+version = "0.11.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "625dc425cab0dca6dc3c3319506e6593dcb08a9f387ea3b284dbd52a92c40555"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "zmij"
+version = "1.0.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8848ee67ecc8aedbaf3e4122217aff892639231befc6a1b58d29fff4c2cabaa"
diff --git a/context-manager/Cargo.toml b/context-manager/Cargo.toml
new file mode 100644
index 00000000..ebee31f9
--- /dev/null
+++ b/context-manager/Cargo.toml
@@ -0,0 +1,60 @@
+[workspace]
+
+[package]
+name = "context-manager"
+version = "0.1.0"
+edition = "2021"
+publish = false
+
+[[bin]]
+name = "context-manager"
+path = "src/main.rs"
+
+[lib]
+name = "context_manager"
+path = "src/lib.rs"
+
+[dependencies]
+iii-sdk = "=0.19.2"
+tokio = { version = "1", features = ["rt-multi-thread", "macros", "sync", "signal", "time"] }
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+serde_yaml = "0.9"
+anyhow = "1"
+thiserror = "2"
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["fmt", "env-filter"] }
+clap = { version = "4", features = ["derive"] }
+async-trait = "0.1"
+# Must stay on the same schemars major as iii-sdk so the derived
+# `JsonSchema` impls satisfy `RegisterFunction::new_async`.
+schemars = "0.8"
+uuid = { version = "1", features = ["v4"] }
+# Default compaction lease key = sha256 of the message set; must be
+# stable across processes, so no std hasher.
+sha2 = "0.10"
+# Home-dir resolution for the `~/`-prefixed lease_dir (same as
+# session-manager's data_dir).
+dirs = "5"
+
+[dev-dependencies]
+serde_json = "1"
+tempfile = "3"
+which = "8"
+# BDD test runner. Drives Gherkin .feature files under tests/features/.
+cucumber = "0.22"
+futures = "0.3"
+uuid = { version = "1", features = ["v4"] }
+tokio = { version = "1", features = ["rt-multi-thread", "macros", "sync", "signal", "time"] }
+
+# Single BDD entry point. cucumber needs full control of the runner, so
+# `harness = false` is required. All .feature files under
+# `tests/features/**` are discovered at runtime; tags (`@pure`,
+# `@engine`, ...) let you filter to a subset:
+#
+#   cargo test --test bdd                  # everything (engine scenarios
+#                                          # soft-skip without an engine)
+#   cargo test --test bdd -- --tags @pure  # no engine required
+[[test]]
+name = "bdd"
+harness = false
diff --git a/context-manager/README.md b/context-manager/README.md
new file mode 100644
index 00000000..b87294b2
--- /dev/null
+++ b/context-manager/README.md
@@ -0,0 +1,98 @@
+# context-manager
+
+Turns a raw conversation history plus a target model into a **model-ready
+context**: a system prompt and an ordered message list that fits inside the
+model's usable token budget. It owns *what the model sees this turn* — token
+counting, function-result pruning, and history compaction (summarisation) —
+and nothing else. It is stateless with respect to conversation storage: pass
+messages in, get results back, persist them yourself (or with
+[`session-manager`](../session-manager/)). Summarisation and model-limit
+lookups go through `llm-router` when installed; token counting and pruning
+work standalone.
+
+## Install
+
+```bash
+iii worker add context-manager
+```
+
+`iii worker add` fetches the binary and the engine starts the worker on the
+next `iii start`. Runtime configuration lives in the `configuration` worker
+(enabled by default in the engine): at boot the worker registers its schema and
+fetches the authoritative value, and it hot-reloads on change (the `lease_dir`
+and `summarizer_timeout_ms` fields are read once at boot and need a restart). An
+optional `--config <file>.yaml` only seeds the first registration.
+
+## Quickstart
+
+Build a budgeted context before every model call:
+
+```rust
+use iii_sdk::{register_worker, InitOptions, TriggerRequest};
+use serde_json::json;
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    let iii = register_worker("ws://localhost:49134", InitOptions::default());
+
+    let result = iii.trigger(TriggerRequest {
+        function_id: "context::assemble".into(),
+        payload: json!({
+            "messages": [
+                { "role": "user", "content": [{ "type": "text", "text": "hi" }], "timestamp": 1 }
+            ],
+            // Inline limits keep the call standalone; pass only `id`/`provider`
+            // to resolve limits through llm-router instead.
+            "model": { "id": "claude-sonnet-4",
+                       "limits": { "context_window": 200000, "max_output_tokens": 8000 } },
+            "system_prompt": "You are a helpful assistant."
+        }),
+        action: None,
+        timeout_ms: Some(30_000),
+    }).await?;
+
+    // -> { system_prompt, messages, token_count, usable, model_resolved,
+    //      applied: { pruned, pruned_tokens, compacted, summary?, tail_start_index? } }
+    println!("{result:#?}");
+    Ok(())
+}
+```
+
+The compaction round trip: when `applied.compacted` is true, **persist
+`applied.summary`** and the boundary your storage maps `applied.tail_start_index`
+to. On later calls pass only the post-compaction window as `messages` plus the
+stored summary as `options.previous_summary` — the summary is rendered into the
+system prompt under `# Conversation summary`, and any further compaction
+*updates* it instead of starting over. Callers that skip persistence stay
+correct at the cost of one summariser call per over-budget request.
+
+The other three functions: `context::count-tokens` (estimate messages + tools +
+system prompt vs a model), `context::prune` (replace verbose function outputs
+with `[output pruned: was ~N tokens]` placeholders, no LLM involved), and
+`context::compact` (summarise the head, keep a recent tail verbatim — returns
+`ok | busy | empty | overflow`).
+
+## Configuration
+
+```yaml
+reserved_tokens_cap: 20000     # default reserve = min(cap, pct% of context window)
+reserved_pct: 10
+tail_turns: 2                  # user+assistant pairs kept verbatim by compaction
+protect_recent_tokens: 40000   # newest function-output tokens never pruned
+min_free_tokens: 20000         # skip pruning when it would free less
+max_output_chars: 2000         # outputs at or under this size are never pruned
+lease_ttl_secs: 300            # compaction mutual-exclusion lease TTL
+allow_fallback_limits: true    # conservative 8192/1024 when limits can't resolve
+summarizer_timeout_ms: 320000  # outer budget for one router::chat summariser call
+```
+
+Other details (and the defaults' single source of truth) live in
+[`src/config.rs`](src/config.rs).
+
+## Local development & testing
+
+```bash
+cargo test                          # unit + BDD (engine scenarios soft-skip)
+cargo test --test bdd -- --tags @pure    # no engine required
+UPDATE_GOLDENS=1 cargo test --test schemas   # regenerate wire-schema goldens
+```
diff --git a/context-manager/architecture/README.md b/context-manager/architecture/README.md
new file mode 100644
index 00000000..f735b8bc
--- /dev/null
+++ b/context-manager/architecture/README.md
@@ -0,0 +1,87 @@
+# context-manager architecture
+
+Reference documentation for the `context-manager` worker — the stateless,
+model-aware engine that turns a raw conversation history into a model-ready
+context, specified in
+[tech-specs/2026-06-agentic/context-manager.md](../../tech-specs/2026-06-agentic/context-manager.md).
+These documents are written to be sufficient on their own: a reader (human or
+LLM) should be able to maintain the worker or integrate against it without
+opening the source.
+
+## Document map
+
+| Document | Audience | Read it when |
+|---|---|---|
+| [internals.md](internals.md) | Maintainers of this worker | You are changing context-manager itself: fixing the budget math, the prune/compaction pipeline, the lease protocol, the summariser adapter, or adding a function. |
+| [integration.md](integration.md) | Authors of other workers / clients | You are building something that calls `context::*` — the harness pre-flight, a batch summariser, a cost-estimating gate, a RAG pipeline. This file is the handoff contract. |
+
+The BDD suite under [../tests/features/](../tests/features) is the executable
+companion to both: every behavioural claim made here is pinned by a scenario,
+each annotated (`# Prevents:`) with the regression it guards against.
+
+## The system in one paragraph
+
+context-manager is a **pure transform over caller-supplied messages**. You
+hand it an `AgentMessage[]` plus a target model; it returns a system prompt
+and a budgeted `AgentMessage[]` that fits the model's usable token window.
+It owns exactly three policies — token counting, function-result pruning, and
+history compaction (LLM summarisation) — and nothing else. It is **stateless
+with respect to conversation storage**: it never reads or writes a session,
+never decides *when* to compact a live conversation, and never talks to a
+provider directly. The only state it keeps is operational, not conversational:
+short-lived compaction **leases** written as files under its own `lease_dir`
+(the same on-disk strategy as `session-manager`), so two callers can't
+summarise the same logical history at once. Summarisation
+and model-limit lookups go through `llm-router` when installed; token counting
+and pruning are fully standalone. Its own runtime configuration is registered
+with and fetched from the `configuration` worker and hot-reloads on change
+(except `lease_dir` / `summarizer_timeout_ms`, which are restart-required).
+That deliberate statelessness is what makes
+it reusable — a chat harness, a document summariser, a RAG pre-flight, or
+another team's bespoke agent can all call `context::assemble` without adopting
+`session-manager` or any particular storage model.
+
+## The system in one diagram
+
+```mermaid
+flowchart LR
+  subgraph callers [Callers]
+    harness[harness / batch summariser / RAG pre-flight]
+  end
+  subgraph worker [context-manager]
+    fns["4 context::* functions"]
+    pipe["assemble pipeline: count -> prune -> compact"]
+    core["pure core: budget · estimate · prune · selection · summary · lease"]
+    ports["ports: ModelResolver · Summarizer · LeaseStore · Clock"]
+  end
+  subgraph deps [Soft dependency]
+    router["llm-router: router::models::get, router::chat"]
+  end
+  leases[("lease_dir: one file per lease")]
+  cfg["configuration worker: schema + value"]
+  harness -->|"iii.trigger(context::*)"| fns
+  fns --> pipe --> core
+  pipe --> ports
+  ports -->|model limits + summariser| router
+  ports -->|compaction leases only| leases
+  cfg -->|"authoritative config + configuration:updated"| worker
+  fns -->|"system_prompt + budgeted messages + applied{}"| harness
+```
+
+## Vocabulary
+
+| Term | Meaning |
+|---|---|
+| **Model-ready context** | The output of `context::assemble`: a `system_prompt` string plus an ordered `AgentMessage[]` that fits the model's `usable` budget, ready to hand to `router::chat`. |
+| **`usable` budget** | The token ceiling one call may fill: `max(0, (input_limit ?? context_window - max_output_tokens) - reserved - thinking_budget)`. Model-adaptive, not a flat constant. |
+| **`reserved`** | Headroom held back from the input budget for response framing; defaults to `min(20000, 10% of context_window)`, overridable per call. |
+| **Prune** | The cheap first pass: replace verbose `function_result` outputs with `[output pruned: was ~N tokens]` placeholders. No LLM, no removal — content is rewritten in place. |
+| **Compaction** | The expensive pass: summarise the **head** of the history into one Markdown summary via the summariser LLM, keeping a recent **tail** verbatim. |
+| **Head / tail** | Compaction splits the history at a boundary: everything before it (the head) is summarised; everything from it on (the tail) is kept verbatim. |
+| **Safe cut** | A boundary the tail may start at without orphaning a `function_result` from its `function_call`: a user or assistant message, never a result (see structural invariants). |
+| **Summary anchor / round trip** | A prior summary passed back as `previous_summary`. The summariser *updates* it in place instead of starting over, so summaries converge instead of growing. The caller persists the summary; the worker never does. |
+| **`tail_start_index`** | Index into the **request** `messages` array where the verbatim tail begins (`null` = everything was summarised). The worker never sees storage ids; the caller maps this onto its own ids. |
+| **Compaction lease** | A `{ nonce, ts }` claim stored as a file under `lease_dir` (scope `context_lease`), keyed by `lease_key` (e.g. a session id) or a hash of the message set. Mutual exclusion so one logical history is summarised by one caller at a time; TTL-expiring so a crash never deadlocks it. |
+| **`model_resolved`** | How limits were obtained: `inline` (caller supplied), `router` (`router::models::get`), or `fallback` (conservative 8192/1024 default). Echoed so a silent fallback is detectable. |
+| **Estimator** | The token counter behind a trait. v1 ships the `chars/4` heuristic for every model; responses report `estimator: "heuristic"` so a future per-model tokenizer is a visible swap. |
+| **`custom` message** | A `role: "custom"` transcript item (app-facing: UI markers, notices). It has no provider wire mapping, so `assemble` excludes it from the model-facing list and its token count — but `count-tokens` still counts what it is given. |
diff --git a/context-manager/architecture/integration.md b/context-manager/architecture/integration.md
new file mode 100644
index 00000000..cf9c4799
--- /dev/null
+++ b/context-manager/architecture/integration.md
@@ -0,0 +1,413 @@
+# Integrating with context-manager
+
+The handoff contract for workers and clients that build on context-manager —
+the harness pre-flight, batch summarisers, cost-estimating gates, RAG
+pipelines, bespoke agents. It is self-contained: everything needed to integrate
+is here, with the
+[spec](../../tech-specs/2026-06-agentic/context-manager.md) as the design
+rationale and [internals.md](internals.md) as the implementation deep-dive.
+
+Contents: [mental model](#1-mental-model) · [conventions](#2-conventions) ·
+[data types](#3-data-types) · [functions](#4-function-catalog) ·
+[the compaction round trip](#5-the-compaction-round-trip) ·
+[structural invariants](#6-structural-invariants-what-you-can-rely-on) ·
+[errors](#7-error-contract) · [patterns](#8-canonical-patterns) ·
+[dependencies & degraded modes](#9-dependencies-and-degraded-modes) ·
+[boundaries](#10-boundaries) · [harness notes](#11-notes-for-the-harness)
+
+## 1. Mental model
+
+context-manager is a **pure function from `(messages, model)` to a model-ready
+context**. You pass the full candidate history and a target model; it returns a
+system prompt and a budgeted `AgentMessage[]` that fits the model's usable
+token window, plus an `applied` report of what it did to get there (pruned,
+compacted, neither). It holds no conversation state, so **you own everything
+durable**: deciding *when* to call it, persisting any compaction summary it
+produces, and storing the transcript itself.
+
+Integration is always some subset of four moves:
+
+1. **Fit a context** with `context::assemble` before a model call — the main
+   entry point. It prunes and/or compacts as needed.
+2. **Persist the round trip** when `applied.compacted` is true (store the
+   summary + boundary, pass them back next time) so summarisation stays cheap
+   and convergent.
+3. **Estimate** with `context::count-tokens` to gate model choice or cost with
+   no LLM and no `llm-router`.
+4. **Use a single pass directly** — `context::prune` (no LLM) or
+   `context::compact` (LLM) — when you want only that step, not the whole
+   pipeline.
+
+## 2. Conventions
+
+- **Invocation**: every function is a **sync** `iii.trigger({ function_id,
+  payload, timeout_ms })` call; responses are the JSON shapes in §4. Give
+  `assemble`/`compact` a generous `timeout_ms` (they may make a summariser LLM
+  call — default outer budget is 320s).
+- **Model input**: functions that need limits take a `ModelInput`. Supply
+  inline `limits` to stay fully standalone (no `llm-router` needed); supply
+  only `id`/`provider` to have the worker resolve limits via
+  `router::models::get`. Resolution order: inline → router → conservative
+  fallback (`8192`/`1024`); the response's `model_resolved` tells you which ran.
+- **Tokens are estimates.** v1 uses a `chars/4` heuristic for every model
+  (`estimator: "heuristic"`). Treat `token_count`/`tokens` as approximate and
+  rely on the built-in `reserved` cushion rather than counting to the byte.
+- **Timestamps** inside `AgentMessage` are caller-supplied integer ms since
+  epoch. The worker reads them only incidentally — order is array order
+  (oldest first), never timestamp.
+- **Errors** are strings beginning with a stable code: `context/<snake>:
+  message`. Match on the code substring. Only `assemble`/`compact` validation
+  and unresolved-model throw; pipeline degradations do **not** error (§7).
+- **Indices, not ids.** `tail_start_index` is an index into the `messages`
+  array *you sent*. The worker never sees your storage ids; you map the index
+  onto your own (§5).
+- **Agent exposure is cost-gated, not secret-gated.** Every function is a pure
+  transform of caller input — nothing to leak — but `assemble`/`compact` can
+  spend a summariser call. Deny those two to agents in cost-sensitive
+  deployments; `count-tokens`/`prune` are always safe (§9).
+
+## 3. Data types
+
+The cross-cutting agentic contracts (TypeScript notation; the wire is plain
+JSON, byte-compatible with session-manager and llm-router):
+
+```typescript
+type Role = "user" | "assistant" | "function_result" | "custom";
+type ThinkingLevel = "minimal" | "low" | "medium" | "high" | "xhigh";
+
+type ContentBlock =
+  | { type: "text"; text: string }
+  | { type: "image"; mime: string; data: string }              // base64
+  | { type: "thinking"; text: string; signature?: string }
+  | { type: "function_call"; id: string; function_id: string; arguments: unknown }
+  | { type: "function_result"; function_call_id: string; content: ContentBlock[]; is_error?: boolean };
+
+type AgentMessage =
+  | { role: "user"; content: ContentBlock[]; timestamp: number }
+  | { role: "assistant"; content: ContentBlock[];
+      stop_reason: "end" | "length" | "function_call" | "aborted" | "error";
+      native_stop_reason?: string; error_message?: string;
+      error_kind?: "auth_expired" | "rate_limited" | "context_overflow" | "transient" | "permanent";
+      warnings?: string[]; usage?: Usage; model: string; provider: string; timestamp: number }
+  | { role: "function_result"; function_call_id: string; function_id: string;
+      content: ContentBlock[]; details: unknown; is_error: boolean; timestamp: number }
+  | { role: "custom"; custom_type: string; content: ContentBlock[];
+      display?: string; details?: unknown; timestamp: number };
+
+// How you name the target model. Inline limits = standalone; id/provider = router lookup.
+type ModelInput = {
+  id: string;
+  provider?: string;
+  limits?: { context_window: number; max_output_tokens: number; input_limit?: number };
+};
+```
+
+`role: "custom"` messages are app-facing transcript items (UI markers, system
+notices, bookkeeping). They have **no provider wire mapping**, so `assemble`
+drops them from the model-facing list and its token count — but `count-tokens`
+still counts what you give it, including customs (use its `by_role.custom`
+bucket to see or subtract that share).
+
+## 4. Function catalog
+
+All four are registered with JSON Schemas (`iii worker info context-manager` /
+`get function info`); the shapes below are the contract. All are **sync**.
+
+### `context::assemble` — build the model-ready context
+
+The main entry point. Pipeline: count → (if over budget) prune function
+outputs → (if still over) compact the head → return the budgeted list.
+
+```typescript
+{
+  messages: AgentMessage[];        // full candidate history, oldest first (required)
+  model: ModelInput;
+  system_prompt?: string;          // base prompt; the summary is merged under it
+  options?: {
+    reserved_tokens?: number;      // override the default reserve
+    tail_turns?: number;           // user+assistant pairs kept verbatim (default 2)
+    allow_compaction?: boolean;    // default true
+    allow_prune?: boolean;         // default true
+    protected_functions?: string[];   // function_ids whose outputs are never pruned
+    thinking_level?: ThinkingLevel;   // reserve the model's thinking budget for this tier
+    lease_key?: string;            // compaction mutual-exclusion key (e.g. a session id)
+    previous_summary?: string;     // persisted summary from a prior compaction (round trip)
+  };
+}
+-> {
+  system_prompt: string;           // base + "# Conversation summary" section when applicable
+  messages: AgentMessage[];        // budgeted, ready for router::chat (no custom roles)
+  token_count: number;             // estimated tokens of the returned context
+  usable: number;                  // the budget it was fit into
+  model_resolved: "inline" | "router" | "fallback";
+  applied: {
+    pruned: boolean; pruned_tokens: number;
+    compacted: boolean;
+    summary?: string;              // present iff compacted — PERSIST THIS (§5)
+    tail_start_index?: number | null;  // index into REQUEST messages where the tail begins
+    tokens_before?: number;        // estimated tokens of the summarised head
+  };
+}
+```
+
+Throws only: `context/invalid_request` (`messages is required`),
+`context/model_unresolved` (`could not resolve model limits` — only when no
+inline limits, no router, and `allow_fallback_limits` is off). Everything else
+— busy lease, failed/absent summariser, disabled steps — is **best effort**:
+the context still returns, possibly with `token_count > usable`.
+
+### `context::count-tokens` — estimate usage
+
+Pure and router-free; safe for cost-sensitive callers with no `llm-router`.
+
+```typescript
+{
+  messages: AgentMessage[];        // required
+  system_prompt?: string;          // counted on top of messages
+  tools?: AgentFunction[];         // invocation schema(s), typically [agent_trigger]
+  model: ModelInput;               // tokenizer selection (v1: always heuristic)
+}
+-> {
+  tokens: number;                  // messages + system_prompt + tools
+  by_role?: { user; assistant; function_result; custom };  // message buckets only
+  estimator: "tokenizer" | "heuristic";
+}
+```
+
+`tokens` equals the sum of `by_role` plus the system-prompt and tools tokens
+(which belong to no role bucket). Counts customs, unlike `assemble`.
+
+### `context::prune` — placeholder verbose function outputs
+
+The cheap pass alone: rewrite verbose `function_result` outputs to
+`[output pruned: was ~N tokens]`. No LLM, no state, no removal.
+
+```typescript
+{
+  messages: AgentMessage[];        // required
+  model?: ModelInput;              // only for token math; optional (heuristic needs none)
+  options?: {
+    protect_recent_tokens?: number;   // newest output tokens never pruned (default 40000)
+    min_free_tokens?: number;         // skip the pass if it frees less (default 20000)
+    max_output_chars?: number;        // outputs at/under this are not "verbose" (default 2000)
+    protected_functions?: string[];   // function_ids never pruned
+  };
+}
+-> { messages: AgentMessage[]; pruned_tokens; pruned_parts; scanned_parts }
+```
+
+The two most recent user turns are always exempt (a hard guard, independent of
+the window). The pass is idempotent.
+
+### `context::compact` — summarise the head, keep a tail
+
+The expensive pass alone. Transient and storage-agnostic: it returns the
+summary for *you* to persist; no session is touched. Most callers never call
+this directly — `assemble` applies compaction inline.
+
+```typescript
+{
+  messages: AgentMessage[];        // required
+  model: ModelInput;
+  options?: {
+    tail_turns?: number;             // default 2
+    previous_summary?: string;       // anchor so summaries converge instead of growing
+    preserve_recent_tokens?: number; // override the adaptive verbatim-tail budget
+    lease_key?: string;              // mutual-exclusion key; default: hash of the message set
+  };
+}
+-> // discriminated on `status`:
+  | { status: "ok"; summary: string; tail_start_index: number | null;
+      tokens_before: number; tokens_after: number; used_prior_summary: boolean }
+  | { status: "busy" }       // a compaction lease is held; retry later
+  | { status: "empty" }      // nothing to compact (empty history / empty summary)
+  | { status: "overflow" };  // summariser unavailable (no llm-router) or it failed
+```
+
+The summary follows a fixed Markdown template (Goal / Constraints & Preferences
+/ Progress / Key Decisions / Actions Taken / Next Steps / Critical Context /
+Relevant Files). With `previous_summary` it is **updated in place**, not
+restarted. Requires `llm-router`; without it you get `overflow`.
+
+## 5. The compaction round trip
+
+The single most important integration contract. **context-manager never
+persists a summary — you must, or every over-budget call re-summarises from
+scratch** (one extra LLM call per request) and summaries never converge.
+
+```mermaid
+sequenceDiagram
+  participant You as your worker
+  participant CM as context::assemble
+  participant Store as your store
+  You->>CM: assemble(messages, model, system_prompt)
+  CM-->>You: applied.compacted=true, summary, tail_start_index, tokens_before
+  You->>Store: persist {summary, boundary = your_id_at(tail_start_index)}
+  Note over You,Store: a later turn
+  You->>CM: assemble(window_from_boundary, model, options.previous_summary=stored)
+  CM-->>You: summary rendered into system_prompt; if over again, summary UPDATED
+```
+
+The contract, step by step:
+
+1. When `applied.compacted` is true, **persist `applied.summary`** and whatever
+   your storage maps `applied.tail_start_index` to. `tail_start_index` is an
+   index into the `messages` array you sent (customs included); resolve it to
+   your own id (e.g. the entry id at that position).
+2. On later calls, pass **only the post-compaction window** as `messages` (the
+   verbatim tail and everything after it) **plus** the stored summary as
+   `options.previous_summary`.
+3. `assemble` renders `previous_summary` into the returned `system_prompt`
+   under a `# Conversation summary` heading. If compaction triggers again, the
+   summariser **updates** that summary instead of starting over — so it
+   converges instead of growing.
+
+`tail_start_index` is `null` when everything was summarised (no verbatim tail).
+A caller that skips step 1 stays correct but pays one summariser call per
+over-budget request.
+
+## 6. Structural invariants — what you can rely on
+
+Whatever pruning or compaction does, the returned context is always
+provider-legal. Build on these:
+
+- **Call/result pairing is never split.** A `function_call` and its
+  `function_result` always land on the same side of any boundary; the
+  compaction tail only ever starts at a user/assistant turn boundary (never
+  between a call and its result, never at a user message carrying an inline
+  result block whose call sits earlier). Orphaned results — which providers
+  reject — cannot appear.
+- **Prune replaces, never removes.** A pruned output's content becomes a single
+  `[output pruned: was ~N tokens]` text block; the message, its
+  `function_call_id`, and the message ordering all survive. Message counts are
+  stable across a prune.
+- **`custom` messages never reach the model.** `assemble` excludes
+  `role: "custom"` from the returned `messages` and from `token_count`. A huge
+  custom entry can't trigger a phantom overflow, and customs never leak to a
+  provider with no wire mapping for them.
+- **`tail_start_index` indexes the request array** you sent (customs included),
+  so it maps cleanly onto your storage even though the model-facing list
+  dropped customs.
+
+## 7. Error contract
+
+| Code | Meaning / trigger | Functions |
+|---|---|---|
+| `context/invalid_request` | `messages` missing or `null` (`messages is required`); a `model` missing where required; malformed shapes serde can't coerce. | all |
+| `context/model_unresolved` | No inline limits, router can't resolve, and `allow_fallback_limits` is off (`could not resolve model limits`). | assemble, compact, count-tokens |
+| `context/state` | A backing lease filesystem write hard-failed (rare; lease problems usually degrade to `busy`). | compact, assemble |
+
+**Not errors — degradations you must read, not catch:**
+
+- `assemble` over budget with prune/compaction disabled, a busy lease, or an
+  unavailable summariser → returns normally with `applied.compacted: false` and
+  `token_count > usable`. Inspect `token_count` vs `usable` to know it didn't fit.
+- `compact` → `{ status: "busy" | "empty" | "overflow" }` are normal outcomes,
+  not thrown errors. `overflow` specifically means "compaction unavailable"
+  (no `llm-router`, or the summariser failed) — treat it as such.
+- Unknown **extra** request fields are tolerated (ignored), so additive API
+  evolution never breaks older callers.
+
+## 8. Canonical patterns
+
+### Pre-flight before every model call (the driver loop)
+
+```mermaid
+sequenceDiagram
+  participant D as driver (e.g. harness)
+  participant CM as context-manager
+  participant R as llm-router
+  D->>CM: assemble(history, model, system_prompt, options{lease_key, previous_summary})
+  CM-->>D: {system_prompt, messages, applied}
+  alt applied.compacted
+    D->>D: persist applied.summary + boundary at tail_start_index
+  end
+  D->>R: router::chat(system_prompt, messages)
+```
+
+Pass `options.lease_key` = your session id so concurrent turns of the *same*
+session serialise their compaction; pass `previous_summary` from your last
+persisted summary; send only the post-compaction window as `messages`.
+
+### Standalone cost gate (no llm-router)
+
+Call `context::count-tokens` with inline `model.limits` (or just an id — the
+heuristic ignores it) to decide whether a cheaper/larger model is needed before
+committing. Include the `agent_trigger` schema in `tools` so tool tokens count.
+No LLM, no router, no state — works on a bare engine.
+
+### Prune-only first pass
+
+If you maintain your own compaction elsewhere and only want to reclaim verbose
+tool output, call `context::prune` directly and persist the rewritten
+`messages`. Idempotent, so re-running is safe.
+
+### Direct compaction with explicit mutual exclusion
+
+A batch summariser that isn't going through `assemble` calls `context::compact`
+with an explicit `lease_key` (so parallel workers on the same logical input see
+`busy` instead of double-summarising), handles the `ok | busy | empty |
+overflow` union, and persists `summary` + `tail_start_index` itself.
+
+## 9. Dependencies and degraded modes
+
+| Dependency | Used for | Without it |
+|---|---|---|
+| `configuration` (required) | the worker's own runtime config: schema registration + the authoritative value, hot-reloaded on change | the worker **cannot boot** — `register`/`get` run at startup and a failure aborts it. Not a per-request dependency: once booted, `context::*` calls never touch it. |
+| local filesystem (`lease_dir`) | compaction leases only | a filesystem error makes `compact`/`assemble` treat the lease as busy → compaction is skipped (best effort); `count-tokens`/`prune` unaffected. |
+| `llm-router` (soft) | model-limit resolution (`router::models::get`) + the summariser (`router::chat`) | Limits fall back to `8192`/`1024` (`model_resolved: "fallback"`) unless you pass inline `limits`; `compact` returns `overflow`; `assemble` can prune but not summarise. |
+
+The fully standalone *request* path — inline `limits` + `count-tokens`/`prune`,
+or `assemble` with compaction off — needs no `llm-router` and writes no lease
+files (the `configuration` dependency is a one-time boot cost, not per call).
+Cost note: only `assemble` and `compact` can trigger a summariser LLM call.
+
+## 10. Boundaries
+
+context-manager does **not**:
+
+- store conversations — pass messages in, persist results yourself (or use
+  [session-manager](../../session-manager/architecture/integration.md));
+- decide *when* to compact a live session — that is your policy (a pre-flight,
+  or the optional reactive trigger in §11);
+- talk to LLM providers directly — summarisation goes through `llm-router`;
+- guarantee exact token counts — v1 is a heuristic estimator (§2);
+- implement long-term / vector memory — that belongs in a dedicated sibling
+  worker (not in v1).
+
+## 11. Notes for the harness
+
+The integration the spec was designed around — context-manager replaces a
+bespoke in-harness compaction side-car with a reusable worker.
+
+- **Hot-path pre-flight.** On each turn, before calling `router::chat`, call
+  `context::assemble` with the session's candidate history, the target model
+  (id + provider so limits resolve through the router; or inline limits), the
+  base `system_prompt`, `options.lease_key` = the session id, and
+  `options.previous_summary` = the last persisted summary. Send the returned
+  `system_prompt` + `messages` straight to the model.
+- **Compaction persistence (the round trip, §5).** When `applied.compacted`,
+  write a session bookkeeping record — the established pattern is a
+  `session::append` with `custom: { custom_type: "compaction", data: { summary,
+  tail_start_entry_id, ... } }`, mapping `applied.tail_start_index` onto the
+  entry id at that position in the history you sent. On the next turn, read it
+  back (`session::messages { include_custom: true }`, scan for the latest
+  `compaction` entry), pass its `summary` as `previous_summary` and the
+  messages from `tail_start_entry_id` onward as `messages`.
+- **Why a lease.** Two harness instances (or a retried turn) on the same
+  session must not double-summarise. The lease keyed on the session id makes
+  the second caller skip compaction (`assemble`) or see `busy` (`compact`)
+  without burning a model call. Crashed holders expire after `lease_ttl_secs`.
+- **Optional reactive pre-warm.** To pre-warm or surface a token-usage metric
+  off the hot path, bind a handler to `session::message-added` and call
+  `context::count-tokens` (cheap, no LLM) there. This is opt-in and lives in
+  the harness — context-manager binds no triggers and never reaches into a
+  session itself, which is exactly what keeps it store-agnostic.
+- **Agent exposure.** All functions are pure transforms (nothing to leak), but
+  deny `context::assemble` and `context::compact` to in-run agents in
+  cost-sensitive deployments — they can trigger a summariser call.
+  `context::count-tokens` and `context::prune` are always safe.
+- **Degraded engine.** With `llm-router` absent the harness still gets budgeted
+  output: limits fall back (or use inline `limits`), prune runs, and an
+  over-budget context returns visibly over (`token_count > usable`) instead of
+  erroring the turn.
diff --git a/context-manager/architecture/internals.md b/context-manager/architecture/internals.md
new file mode 100644
index 00000000..58cb5942
--- /dev/null
+++ b/context-manager/architecture/internals.md
@@ -0,0 +1,527 @@
+# context-manager internals
+
+Maintainer documentation. Everything here describes the implementation as it
+is; the consumer-facing contract lives in [integration.md](integration.md).
+Spec of record:
+[tech-specs/2026-06-agentic/context-manager.md](../../tech-specs/2026-06-agentic/context-manager.md).
+
+## 1. Crate layout
+
+One cargo crate, one `[[bin]]`, plus a `[lib]` so tests drive production code
+in-process. The cut that keeps everything testable: **handlers are thin, the
+`core` modules are pure logic with no I/O, and every outside dependency is a
+port** wired to a production adapter in the binary and to a deterministic fake
+in tests.
+
+| Path | Responsibility |
+|---|---|
+| [src/main.rs](../src/main.rs) | Boot: CLI (`--config` seed, `--url`, `--manifest`), engine connect, **register the config schema (+ optional seed) with the `configuration` worker and fetch the authoritative value** (boot-fatal on failure), build adapters from it, `register_all`, then bind the `configuration` hot-reload trigger; Ctrl+C → `shutdown_async`. |
+| [src/lib.rs](../src/lib.rs) | Module tree only. |
+| [src/types.rs](../src/types.rs) | Wire contracts shared with the agentic family: `Role`, `ContentBlock` (5 variants), `AgentMessage` (4 roles), `ModelInput`, `ModelLimits`, `Model`, `ThinkingLevel`, `AgentFunction`. Serde renames keep the JSON byte-compatible with the TypeScript spec and session-manager's Rust copy. |
+| [src/config.rs](../src/config.rs) | `WorkerConfig` (10 budget/prune/lease knobs incl. `lease_dir`, `~/`-expanded via `resolved_lease_dir`), each with a serde default; `deny_unknown_fields` so a typo'd key fails loudly. Also the JSON-Schema source (`json_schema`/`to_json`/`from_json`, derived `JsonSchema`) and the env-expanding seed parser (`from_file`/`from_yaml`); `boot_signature` names the restart-required fields. |
+| [src/configuration.rs](../src/configuration.rs) | The `configuration` worker client: `register_config` (schema + seed), `fetch_config` (authoritative, env-expanded), the `ConfigCell` snapshot + `apply_config`, the `reloadable` restart-required gate, and the `context::on-config-change` trigger handler. |
+| [src/error.rs](../src/error.rs) | `ContextError` → `code: message` on the bus (`context/invalid_request`, `context/model_unresolved`, `context/state`). The two spec strings are kept verbatim. |
+| [src/ports.rs](../src/ports.rs) | The four seams: `ModelResolver`, `Summarizer`, `LeaseStore`, `Clock`, plus the `Deps` struct every handler receives. |
+| [src/manifest.rs](../src/manifest.rs) | `--manifest` JSON for the registry publish pipeline; `default_config` mirrors `WorkerConfig::default()` field-for-field (unit-tested). |
+| [src/functions/mod.rs](../src/functions/mod.rs) | Function ids + descriptions, `resolve_model` (the spec's resolution order), the generic typed `register` helper, `register_all`, and the schema `catalog()` (golden-tested). |
+| [src/functions/&lt;verb&gt;.rs](../src/functions) | One file per function: request/response structs (serde + `JsonSchema`, doc comments become schema descriptions) and a `pub async fn handle(deps, req)`. BDD calls these `handle` fns directly, so engine-free tests exercise the exact production path. |
+| [src/core/budget.rs](../src/core/budget.rs) | `ResolvedModel`, the `usable` math, `default_reserved`, `preserve_recent_budget`, `fallback_model`. |
+| [src/core/estimate.rs](../src/core/estimate.rs) | `Estimator` trait + `HeuristicEstimator` (`chars/4`), `estimator_for_model`, per-role tallies. |
+| [src/core/prune.rs](../src/core/prune.rs) | The prune algorithm (newest-first scan, protected window, `min_free_tokens` guard, in-place placeholder rewrite). |
+| [src/core/selection.rs](../src/core/selection.rs) | Turn partitioning and token-aware verbatim-tail selection with the safe-cut invariant. |
+| [src/core/summary.rs](../src/core/summary.rs) | Summariser prompt construction (template, previous-summary anchoring), `strip_media`, and the `# Conversation summary` system-prompt rendering. |
+| [src/core/lease.rs](../src/core/lease.rs) | Compaction lease acquire/release protocol + the default sha256 lease key. |
+| [src/adapters/router.rs](../src/adapters/router.rs) | `RouterModelResolver` (`router::models::get`) and `RouterSummarizer` (`router::chat` over an SDK channel). |
+| [src/adapters/fs_lease.rs](../src/adapters/fs_lease.rs) | `FsLeaseStore`: one JSON file per lease key under `lease_dir/<scope>/`, a process-local `Mutex` cache, and atomic `tmp + rename` writes (session-manager's `FsStore` strategy). |
+| [tests/](../tests) | Cucumber BDD (`tests/bdd.rs`, `harness = false`) + schema goldens + manifest subprocess test. See §13. |
+
+## 2. Request lifecycle
+
+```mermaid
+flowchart LR
+  bus[engine bus] -->|"typed request (serde rejects malformed)"| handler["functions::&lt;verb&gt;::handle"]
+  handler -->|resolve_model| resolver[ModelResolver port]
+  handler --> core["core:: budget / estimate / prune / selection / summary"]
+  core -->|"compaction only"| lease[LeaseStore + Clock ports]
+  core -->|"compaction only"| summarizer[Summarizer port]
+  handler -->|response or ContextError| bus
+```
+
+- **Input validation is two-layered.** Serde shape validation at the boundary
+  (unknown enum values, wrong types) fails before any logic runs; then the
+  handler checks the one thing serde can't (`messages` present and non-null →
+  else `context/invalid_request: messages is required`).
+- **Handlers orchestrate; `core` decides.** A handler resolves the model,
+  computes the budget, calls the pure `core` functions, and (for compaction)
+  acquires/releases the lease and invokes the summariser port. No `core`
+  module performs I/O.
+- **`Deps` is the whole world.** `Deps { config, resolver, summarizer, leases,
+  clock }` — every handler takes `&Deps`, so swapping the four ports for fakes
+  reproduces the exact production code path without an engine.
+- **Config is a hot-swappable snapshot.** `Deps.config` is a `ConfigCell`
+  (`Arc<RwLock<Arc<WorkerConfig>>>`) sourced from the `configuration` worker;
+  handlers call `deps.config().await` once per request. A `configuration:updated`
+  trigger ([configuration.rs](../src/configuration.rs)) re-fetches and swaps it
+  live for the per-call tuning knobs, but refuses changes to `lease_dir` /
+  `summarizer_timeout_ms` (consumed once at boot for the `FsLeaseStore` /
+  `RouterSummarizer`) with a "restart required" log.
+
+## 3. The assemble pipeline
+
+`context::assemble` is the worker's reason to exist; the other three functions
+are either steps of it or probes into it. The pipeline, all in
+[functions/assemble.rs](../src/functions/assemble.rs):
+
+```mermaid
+flowchart TD
+  A[resolve model limits] --> B[compute usable budget]
+  B --> C["build model-facing view: drop role:custom, record view_to_orig"]
+  C --> D["render system prompt: base + optional previous_summary"]
+  D --> E[count tokens]
+  E --> F{over usable AND allow_prune?}
+  F -->|yes| G[prune verbose function outputs] --> H[recount]
+  F -->|no| H
+  H --> I{still over AND allow_compaction?}
+  I -->|yes| J["try_compact under lease: select tail, summarise head"]
+  I -->|no| K[assemble response]
+  J -->|summary produced| L["replace system prompt with summary, drop head, recount"] --> K
+  J -->|busy / failed / empty| K
+```
+
+Load-bearing details:
+
+1. **Order is fixed: count → prune → compact.** Prune is cheap (no LLM) and
+   often enough, so it runs first; compaction is the expensive fallback. Each
+   step re-counts, and each is gated on *still being over budget*, so a context
+   that already fits passes through byte-identical with `applied` all-false and
+   no summariser cost (`assemble.feature` "under budget passes through
+   untouched").
+2. **The model-facing view excludes `role: "custom"`.** Custom messages have
+   no provider wire mapping, so they are filtered out of `working` *before*
+   counting (a huge custom entry must not trigger a phantom overflow) and never
+   appear in the returned `messages`. `view_to_orig: Vec<usize>` records, for
+   each surviving message, its index in the original request array — so
+   `tail_start_index` can be reported against the **request** array the caller
+   holds, customs included (`invariants.feature` "tail_start_index accounts for
+   excluded custom messages").
+3. **Degradation is best effort and visible, never an error.** A busy lease, a
+   failed/unavailable summariser, or disabled steps all leave the turn alive:
+   the response still returns, `applied.compacted` is false, and
+   `token_count > usable` is the visible signal that the context didn't fit.
+   `assemble` only throws for `messages is required` or `could not resolve
+   model limits` (fallback disabled).
+4. **`applied` is the audit trail.** `{ pruned, pruned_tokens, compacted,
+   summary?, tail_start_index?, tokens_before? }` reports exactly what ran.
+   `summary`/`tail_start_index`/`tokens_before` appear only when
+   `compacted` — `tail_start_index` is `Some(Some(i))` for a real cut and
+   `Some(None)` for "everything summarised", serialised as a number or `null`.
+5. **The compaction lease key defaults to a hash of the *request* messages** —
+   the same derivation `context::compact` uses — so a caller hitting both
+   functions with the same history contends on the same claim.
+
+## 4. Token budget model
+
+[core/budget.rs](../src/core/budget.rs). The usable input budget is
+model-adaptive:
+
+```text
+base   = input_limit ?? (context_window - max_output_tokens)
+usable = max(0, base - reserved - thinking_budget)
+```
+
+- **All arithmetic saturates.** A model whose output budget exceeds its window,
+  or a reserve larger than the base, clamps `usable` to 0 — never wraps around
+  `u64` (`model_and_budget.feature` "a model smaller than its own output budget
+  clamps to zero").
+- **`reserved`** defaults to `default_reserved = min(reserved_tokens_cap,
+  context_window * reserved_pct / 100)` = `min(20000, 10%)` with stock config.
+  This scales the reserve down for small models (a flat 20k would eat a third
+  of a 32k window) while capping it for huge ones. Overridable per call via
+  `options.reserved_tokens`.
+- **`thinking_budget`** is `thinking_budgets[thinking_level]` *only* when the
+  caller passes `options.thinking_level` **and** the resolved model declares a
+  budget for that tier; otherwise 0. Inline limits carry no budgets, so a
+  thinking level is a no-op for standalone callers. This is how assemble leaves
+  room for the reasoning tokens a thinking tier will consume.
+- Worked: a 200k/8k model with defaults → `200000 - 8000 - 20000 = 172000`
+  usable; a 32k/16k model → `32000 - 16000 - 3200 = 12800`; the conservative
+  fallback (8192/1024) → `8192 - 1024 - 819 = 6349`.
+
+`preserve_recent_budget(usable, override)` is the **separate** budget that caps
+the verbatim tail during compaction: `clamp(usable / 4, 2000, 8000)`, or the
+caller's `preserve_recent_tokens` verbatim. It is intentionally small — the
+tail is "the last little bit kept raw", not a second copy of the window.
+
+## 5. Token estimation
+
+[core/estimate.rs](../src/core/estimate.rs). Estimation sits behind the
+`Estimator` trait with three methods (`message`, `text`, `function`) and a
+`kind()` the response echoes.
+
+- **v1 ships one implementation: `HeuristicEstimator` = serialized-JSON
+  `chars / 4`.** It counts the *full serialized message*, so structure and
+  metadata weigh in roughly as they do on the wire, and it is deterministic and
+  model-independent. `estimator_for_model` ignores the model id today and always
+  returns the heuristic; `count-tokens` reports `estimator: "heuristic"`.
+- The trait is the seam for a real per-model tokenizer later: slot it into
+  `estimator_for_model`, return `EstimatorKind::Tokenizer`, and every count —
+  budget math, prune sizing, tail selection — picks it up with no caller
+  change. The reported `estimator` field makes the swap observable.
+- `estimate_by_role` partitions a message set into `{ user, assistant,
+  function_result, custom }`; the buckets sum exactly to the message total
+  (`count_tokens.feature`), which is the contract that lets callers subtract
+  the custom share.
+
+## 6. Pruning
+
+[core/prune.rs](../src/core/prune.rs). One pass, newest-to-oldest, that
+rewrites verbose `function_result` outputs to `[output pruned: was ~N tokens]`
+and **never removes anything** — the message, its `function_call_id`, and the
+ordering all survive (the structural invariant providers depend on).
+
+Eligibility, applied while scanning from the newest message backward:
+
+1. **The two most recent user turns are always exempt** (`PROTECTED_USER_TURNS
+   = 2`, a hard prior-art constant, not operator-tunable): the scan counts
+   `user` messages and skips everything until it has passed two of them.
+2. **Protected token window:** accumulate each scanned output's tokens into
+   `window_tokens`; while `window_tokens <= protect_recent_tokens` the output
+   is inside the newest window and kept. Because the scan is newest-first, the
+   freshest outputs fill the window and push older ones out of it.
+3. **`protected_functions`** (by `function_id`) are never pruned and never even
+   counted as scanned.
+4. **Verbosity threshold:** an output whose text is `<= max_output_chars` is
+   not "verbose" — pruning it would free almost nothing — so it stays.
+5. **`min_free_tokens` guard:** sum what the eligible outputs *would* free; if
+   that total is below `min_free_tokens`, **nothing is rewritten at all**. A
+   no-op beats a destroyed-but-still-over context. This guard fires before any
+   mutation, so a skipped pass leaves `messages` untouched.
+
+`PruneStats { pruned_tokens, pruned_parts, scanned_parts }` distinguishes
+"examined" from "rewritten". The pass is idempotent: a second run sees only the
+tiny placeholders, which are under `max_output_chars`, so nothing is verbose
+(`prune.feature` "pruning twice is idempotent"). On the assemble path the same
+`core::prune::prune` runs with config-derived params plus the call's
+`protected_functions`.
+
+## 7. Compaction — tail selection
+
+[core/selection.rs](../src/core/selection.rs) decides the head/tail boundary;
+it is pure and the most invariant-sensitive code in the worker.
+
+- **Turns** partition the history: a turn starts at each `user` message and
+  runs to the next one. Messages before the first user message belong to no
+  turn and are always head material.
+- **`select(messages, budget, tail_turns, estimator)`** keeps up to the last
+  `tail_turns` whole turns that fit `budget`, newest-first; when a whole turn
+  won't fit, it falls back to a **safe partial cut** inside that turn via
+  `split_turn`. Everything before the kept tail is the head to summarise.
+- **The safe-cut invariant** is the whole point. A `function_call` and its
+  `function_result` must land on the same side of the boundary, or providers
+  reject the orphaned result. So the tail may only *start* at:
+  - an `assistant` message, or
+  - a `user` message that carries **no inline `function_result` block**
+    (Anthropic-style tool results travel inside user messages; a user message
+    that opens with one has its call in the previous turn).
+
+  A `function_result` message, or such a user message, is **not** a safe cut.
+  When the newest turn starts unsafely, selection accumulates it and defers the
+  boundary to an older safe turn, carrying the unsafe turn tail-side with its
+  call (`selection.rs` tests + `invariants.feature`).
+- `Selection { head_len, tail_start_index }`. `tail_start_index` is `None` when
+  nothing could be safely kept verbatim (whole history is head); the handler
+  maps a `Some(view_index)` back through `view_to_orig` before returning it.
+- `tail_turns: 0` summarises everything (`tail_start_index: null`).
+
+## 8. Compaction — summariser prompts and the round trip
+
+[core/summary.rs](../src/core/summary.rs) builds what the summariser sees and
+renders what the caller gets back.
+
+- **`build_system_prompt(previous_summary)`** — instructs either "Create a new
+  anchored summary" (no prior) or "Update the anchored summary below …"
+  wrapping the prior in `<previous-summary>` tags (anchored). Either way it
+  appends `SUMMARY_TEMPLATE`, the fixed Markdown structure (Goal / Constraints
+  & Preferences / Progress {Done, In Progress, Blocked} / Key Decisions /
+  Actions Taken / Next Steps / Critical Context / Relevant Files). Update mode
+  is how summaries **converge instead of growing**.
+- **`render_user_prompt(head)`** — the head messages inside a `<conversation>`
+  block: text verbatim, `function_call`s as terse `[tool_call] <id> <args>`
+  one-liners; other block kinds dropped.
+- **`strip_media(head, max_output_chars)`** — a copy fed to the summariser with
+  images replaced by `[image stripped]` everywhere and `function_result` text
+  truncated to `max_output_chars` (+ `... [truncated]`). The original
+  transcript the caller holds is never mutated. This keeps base64 blobs and
+  one giant tool dump from dominating the summariser's input.
+- **`render_system_prompt(base, summary)`** — what assemble returns: the base
+  prompt, then a `# Conversation summary` section with the summary. Empty base
+  → just the section; no summary → just the base.
+
+The compaction round trip, the worker's only stateful-feeling contract (the
+state is the *caller's*, not the worker's):
+
+```mermaid
+sequenceDiagram
+  participant C as caller (e.g. harness)
+  participant A as context::assemble
+  participant Store as caller's store
+  C->>A: assemble(messages, model)
+  A->>A: over budget -> compact head
+  A-->>C: applied.compacted, summary, tail_start_index
+  C->>Store: persist summary + boundary (caller maps index -> own id)
+  Note over C,Store: later turn
+  C->>A: assemble(post-compaction window, previous_summary = stored summary)
+  A->>A: render summary into system prompt; if over again, UPDATE the anchor
+  A-->>C: new summary (converged), new tail_start_index
+```
+
+The worker **never persists**. A caller that skips persistence stays correct
+but pays one summariser call per over-budget request, because every call
+re-summarises from scratch (`assemble_round_trip.feature` "skipping persistence
+costs one summariser call per request"). Persisting `summary` +
+`tail_start_index` and passing them back as `previous_summary` + a trimmed
+`messages` window is what makes it cheap and convergent.
+
+## 9. Compaction leases and concurrency
+
+[core/lease.rs](../src/core/lease.rs) over the `LeaseStore` port. The only
+state the worker writes: a `{ nonce, ts }` claim per lease, stored as one JSON
+file at `<lease_dir>/<scope>/<key>.json` (scope `context_lease`), keyed by
+`options.lease_key` or `default_lease_key` (sha256 over the serialized message
+set, separator byte between messages so `["ab"]` and `["a","b"]` never
+collide). The key derivation is deliberately `sha2`, not a std hasher, so it is
+**stable across processes and Rust versions** — two callers of the same logical
+history always derive the same file.
+
+Storage is the [`FsLeaseStore`](../src/adapters/fs_lease.rs) adapter, which
+copies session-manager's `FsStore` strategy: a process-local `Mutex` cache plus
+atomic `tmp + rename` writes. Atomicity is therefore **per-process** — the
+single-writer-per-`lease_dir` assumption session-manager also makes — which is
+sufficient for one context-manager instance. The protocol below holds within
+that process: `swap` runs the whole read-modify-write under the cache mutex, so
+concurrent acquirers serialise.
+
+The protocol (ported from harness `runtime/lease.ts`):
+
+- **TTL is enforced by readers.** A claim whose `ts` is older than `ttl_ms`
+  reads as free, folding crash recovery into acquisition — a dead holder's
+  lease simply expires (`lease_ttl_secs`, default 300). The boundary is strict:
+  299999 ms old still wins, 300001 ms is taken over.
+- **`acquire`**: read the key; if a live claim exists, return `None` (busy).
+  Otherwise write a fresh `{ nonce, ts }` via the atomic **`swap`** and inspect
+  the prior value it returns; if we clobbered a *still-live* claim (a racing
+  acquirer beat us), restore it and bow out. Exactly one concurrent acquirer
+  sees a free/expired prior and wins.
+- **`release`** only clears the key **if it still holds our nonce** — a holder
+  that lost the lease to TTL takeover must never wipe the new holder's claim.
+- **Store failures read as busy, never as a win.** A `get`/`swap` error returns
+  `None` (busy) so a filesystem outage can't let every contender summarise at
+  once; a transient busy is retryable (`compact_lease.feature` "a state outage
+  reads as busy, never as a win").
+
+`compact` returns `status: "busy"` on a held lease; `assemble` silently skips
+compaction (best effort) and leaves the holder's claim intact. The cache mutex
+serialises acquirers within the process, and the per-history key keeps unrelated
+histories from interfering; the worker is otherwise stateless.
+
+## 10. Ports and adapters
+
+The four ports in [ports.rs](../src/ports.rs), production adapters in
+[adapters/](../src/adapters):
+
+| Port | Production adapter | Backed by | Failure posture |
+|---|---|---|---|
+| `ModelResolver` | `RouterModelResolver` | `router::models::get` (5s timeout) | `Ok(None)` = router up but model unknown; `Err` = router absent/unreachable. Both → fallback when allowed. |
+| `Summarizer` | `RouterSummarizer` | `router::chat` over an SDK channel | `Unavailable` (router not routable) vs `Failed` (provider/stream error) vs `Empty`. |
+| `LeaseStore` | `FsLeaseStore` | one JSON file per key under `lease_dir` (atomic `tmp + rename`; process-local `Mutex` for `swap`) | Errors surface as busy via `core::lease`. |
+| `Clock` | `SystemClock` | wall clock (ms since epoch) | — |
+
+`RouterSummarizer` details worth knowing before touching it:
+
+- It creates an SDK channel, passes `writer_ref` to `router::chat`, and reads
+  the streamed `AssistantMessageEvent` frames. Frames may arrive as WS text
+  (callback) or binary (`read_all`); it collects both because the router's
+  framing is provider-defined.
+- `extract_summary` folds frames: the terminal `done` event's message text
+  wins; accumulated `text_delta`s are the fallback; a terminal `error` (or a
+  `stop` with `stop_reason: "error"`) fails. Empty result → `SummarizeError::Empty`.
+- `summarizer_timeout_ms` (default 320000) is the **outer** budget for the
+  whole call and must exceed the router's own stream budget; `READER_DRAIN_MS`
+  is a 2s grace for the socket to drain after the trigger returns.
+- Unroutable detection is string-sniffing the error (`function_not_found` /
+  `not found` / `no function`) → `Unavailable`, which `compact` maps to
+  `status: "overflow"` with `error_kind: "permanent"` logged. This is the
+  defined degraded mode when `llm-router` is absent.
+
+`FsLeaseStore::swap` holds the cache `Mutex` across the whole read-modify-write:
+it reads the current claim (replaying the key's file on a cold cache), writes the
+new `{ nonce, ts }` atomically (`tmp + rename`), and returns the prior — so two
+acquirers in the process serialise and exactly one sees a free/expired prior. A
+missing file, or a file whose contents aren't a `{ nonce, ts }` record, reads as
+`None` (free) — the same tolerant posture as the harness lease. All filesystem
+I/O is synchronous and never crosses an `.await` while the guard is held.
+
+## 11. Model resolution
+
+`functions::resolve_model` implements the spec order exactly:
+
+1. `input.limits` present → `ResolvedModel::from_inline` (`model_resolved:
+   "inline"`); no router call. This is the standalone path.
+2. else `router::models::get(provider, id)`:
+   - `Ok(Some(model))` → `from_router` (`"router"`), carrying the model's
+     `thinking_budgets`.
+   - `Ok(None)` (unknown) or `Err` (router down/absent) → fall through.
+3. fallback: if `allow_fallback_limits` (default true), `fallback_model()`
+   (8192/1024, `"fallback"`, logged with the reason). Else
+   `ContextError::ModelUnresolved` → `context/model_unresolved: could not
+   resolve model limits (...)`.
+
+A dead router and an unknown model are deliberately indistinguishable to the
+caller — both yield the same detectable `fallback`. Inline limits never carry
+`thinking_budgets`, so inline callers can't accidentally pay a thinking tax.
+
+## 12. Error model
+
+`ContextError` (in [error.rs](../src/error.rs)) renders as `code: message`;
+the code before the colon is the stable contract. One mapping point:
+`From<ContextError> for IIIError` (always `IIIError::Handler`).
+
+| Variant | Code | When |
+|---|---|---|
+| `InvalidRequest` | `context/invalid_request` | `messages` absent/null (`messages is required`); serde-survivable shape problems. |
+| `ModelUnresolved` | `context/model_unresolved` | No inline limits, router can't resolve, fallback disabled (`could not resolve model limits`). |
+| `State` | `context/state` | A filesystem call backing the lease failed (reserved; lease failures normally degrade to busy rather than throw). |
+
+`messages is required` and `could not resolve model limits` are kept word-for-word
+because callers match on them (`errors.feature`). Adding a variant means: add
+the code, a `Display` test in `error.rs`, and an `errors.feature` scenario.
+
+## 13. Configuration, boot, and the wire surface
+
+Runtime config is **registered with and fetched from the `configuration`
+worker** ([configuration.rs](../src/configuration.rs)); a `--config` YAML file
+only SEEDS the first registration. The shape below (shipped defaults, also the
+registered schema's `example`) is what `WorkerConfig::json_schema()` describes:
+
+```yaml
+reserved_tokens_cap: 20000     # default reserve = min(cap, reserved_pct% of context_window)
+reserved_pct: 10
+tail_turns: 2                  # user+assistant pairs kept verbatim by compaction
+protect_recent_tokens: 40000   # newest function-output tokens never pruned
+min_free_tokens: 20000         # skip pruning when it would free less
+max_output_chars: 2000         # outputs at/under this are not "verbose"; also the summariser truncation cap
+lease_ttl_secs: 300            # compaction mutual-exclusion lease TTL
+allow_fallback_limits: true    # conservative 8192/1024 when limits can't resolve
+summarizer_timeout_ms: 320000  # outer budget for one router::chat summariser call
+```
+
+Boot and reload rules (`main.rs` / `configuration.rs`):
+
+- `configuration` is a **required boot dependency**: `register_config` then
+  `fetch_config` run over the live connection, and a failure aborts startup. A
+  missing stored value seeds defaults (or the `--config` seed); a `null` value
+  reads as defaults.
+- The config **hot-reloads**: a `configuration:updated` trigger re-fetches the
+  authoritative value and swaps the snapshot. `lease_dir` and
+  `summarizer_timeout_ms` are consumed once at boot (the `FsLeaseStore` and
+  `RouterSummarizer` are built then and never rebuilt), so a change to either is
+  refused on reload (logged "restart required", previous snapshot kept) — every
+  other field applies live. The handler ignores its trigger payload and
+  re-fetches, so a forged call can at most reload the already-stored value.
+- A `--config` file is only a SEED for the first registration; an unparseable
+  seed WARNS and is skipped (the authoritative value comes from the worker).
+- Every config field is per-call-overridable where the spec allows
+  (`reserved_tokens`, `tail_turns`, the prune thresholds, `lease_key`,
+  `preserve_recent_tokens`); request options take precedence over config.
+- `llm-router` is soft: the worker serves `count-tokens` and `prune` without it;
+  only compaction and router-based model resolution degrade.
+
+**The wire surface is golden-tested.** `functions::catalog()` returns each
+function's id, description, and schemars-derived request/response schemas, in
+registration order; `tests/schemas.rs` diffs them against
+[tests/golden/schemas/](../tests/golden/schemas) (regenerate with
+`UPDATE_GOLDENS=1`). `schema_of` mirrors iii-sdk's internal generator
+(`SchemaSettings::draft07()`), so a golden pins exactly what `register_async`
+emits. Keep `catalog()` in lockstep with `register_all`.
+
+## 14. Determinism and testing
+
+Everything a test needs to pin behaviour is injectable: the four ports
+(`ModelResolver`, `Summarizer`, `LeaseStore`, `Clock`). The production binary
+is one composition of them; the BDD world is another.
+
+- **`tests/bdd.rs`** (cucumber, `harness = false`). Tags: `@pure` needs nothing
+  (calls `handle` fns directly with fake ports — see `tests/common/world.rs`,
+  `tests/common/fakes.rs`); `@engine` registers the production surface in-process
+  against a live engine and soft-skips when none is reachable.
+- The `@pure` world wires fake ports: a scriptable resolver (known models,
+  unavailable, declared thinking budgets), a recording summariser (returns /
+  fails / unavailable / empty, captures the prompts + model it ran on), an
+  in-memory lease store (+ "foreign lease", "unavailable"), and a manually
+  advanced `FakeClock` to drive TTL expiry.
+- **`engine_roundtrip.feature`** is the proof that the registered wire surface
+  matches the handlers: real triggers, production adapters, `llm-router` absent
+  — i.e. the spec's degraded mode (limits fall back, prune/count work, compact
+  overflows after a real filesystem lease acquire/release cycle).
+- Unit tests live next to what they pin: budget math in `budget.rs`, the
+  heuristic in `estimate.rs`, prune eligibility in `prune.rs`, the safe-cut
+  matrix in `selection.rs`, prompt construction in `summary.rs`, lease key
+  stability and TTL in `lease.rs`, frame folding in `adapters/router.rs`,
+  config defaults in `config.rs`.
+- Convention: every scenario carries a `# Prevents:` comment naming the
+  regression it catches.
+
+Verification commands (CI parity):
+
+```bash
+cargo fmt --all -- --check
+cargo clippy --all-targets --all-features -- -D warnings
+cargo test --all-features                       # @engine soft-skips without iii
+cargo test --test bdd -- --tags @pure           # no engine required
+UPDATE_GOLDENS=1 cargo test --test schemas      # regenerate wire-schema goldens
+./target/debug/context-manager --manifest | jq .
+```
+
+## 15. Sharp edges and known limitations
+
+- **The estimator is `chars/4`, not a real tokenizer.** Every budget decision,
+  prune sizing, and tail fit is an *estimate*; `token_count` can differ from
+  what the provider bills. The trait is ready for a per-model tokenizer; until
+  then treat all counts as approximate and keep `reserved` as the cushion.
+- **Compaction needs `llm-router`.** Without it, `compact` returns `overflow`
+  and `assemble` can prune but not summarise — an over-budget context comes
+  back over budget (visibly, via `token_count > usable`). Pure token/prune
+  calls are unaffected.
+- **One summariser call per over-budget request if the caller doesn't persist.**
+  The round trip (§8) is the fix; skipping it is correct but not cheap.
+- **`assemble`/`compact` can spend money.** They may trigger a summariser LLM
+  call; `count-tokens` and `prune` never do. Deny the former two in
+  cost-sensitive agent deployments (see [integration.md §9](integration.md)).
+- **Lease key is content-derived by default.** Two callers compacting the
+  *same* history contend even without a shared `lease_key`; conversely, a
+  single logical session whose message set changes between calls gets a
+  different default key, so pass an explicit `lease_key` (e.g. the session id)
+  to serialise a live session's compactions.
+- **No long-term / vector memory.** v1 compresses one history for one turn;
+  durable cross-session memory belongs in a dedicated sibling worker (spec
+  Boundaries).
+
+## 16. How to extend
+
+- **New function:** add `src/functions/<verb>.rs` with typed request/response
+  (+ doc comments → schema descriptions) and a `pub async fn handle(deps,
+  req)`; add the id/description consts, a `register(...)` line in
+  `register_all`, and a `spec::<Req, Resp>(...)` line in `catalog()` (then
+  regenerate the schema golden); add a dispatch arm in `tests/common/world.rs`
+  and a feature file.
+- **Real tokenizer:** implement `Estimator`, return `EstimatorKind::Tokenizer`,
+  and key it off the model in `estimator_for_model`. Nothing else changes; the
+  reported `estimator` field flips to `"tokenizer"` and pins the swap.
+- **New budget/prune knob:** add a field to `WorkerConfig` (with a `default_*`
+  fn and the `Default` impl), mirror it in `manifest.rs::default_config` (a
+  unit test enforces parity), thread it through the relevant `core` params, and
+  expose a per-call override in the function's `Options` if the spec allows it.
+- **New port-backed dependency:** add a trait to `ports.rs`, a production
+  adapter in `adapters/`, a field on `Deps`, a fake in `tests/common/fakes.rs`,
+  and wiring in `main.rs`. Keep `core` pure — it should depend on the trait,
+  never the adapter.
diff --git a/context-manager/build.rs b/context-manager/build.rs
new file mode 100644
index 00000000..81caa36d
--- /dev/null
+++ b/context-manager/build.rs
@@ -0,0 +1,6 @@
+fn main() {
+    println!(
+        "cargo:rustc-env=TARGET={}",
+        std::env::var("TARGET").unwrap()
+    );
+}
diff --git a/context-manager/config.yaml b/context-manager/config.yaml
new file mode 100644
index 00000000..86acad2b
--- /dev/null
+++ b/context-manager/config.yaml
@@ -0,0 +1,40 @@
+# context-manager runtime configuration.
+#
+# Token-budget knobs. The usable input budget per call is
+#   usable = (input_limit ?? context_window - max_output_tokens)
+#            - reserved - thinking_budget
+# where reserved defaults to min(reserved_tokens_cap,
+# context_window * reserved_pct / 100) and is overridable per call.
+
+# Cap on the default reserve (tokens).
+reserved_tokens_cap: 20000
+# Percentage of the context window reserved by default.
+reserved_pct: 10
+
+# user+assistant turn pairs always kept verbatim by compaction when the
+# request omits options.tail_turns.
+tail_turns: 2
+
+# Prune defaults (overridable per call via options).
+protect_recent_tokens: 40000   # newest function-output tokens never pruned
+min_free_tokens: 20000         # skip pruning when it would free less
+max_output_chars: 2000         # outputs at or under this size are not "verbose"
+
+# Compaction mutual-exclusion lease TTL (seconds). A crashed holder's
+# claim expires after this long.
+lease_ttl_secs: 300
+
+# Directory holding compaction-lease files (one `<scope>/<key>.json` per
+# lease, written atomically). A leading `~/` expands to the home
+# directory. Mirrors session-manager's data_dir; this worker is the
+# single writer of this directory.
+lease_dir: ~/.iii/data/context-manager
+
+# When the request has no inline limits and llm-router is absent, fall
+# back to a conservative model (context_window 8192 / max_output 1024)
+# instead of failing. Set false to make unresolvable models an error.
+allow_fallback_limits: true
+
+# Outer budget for one summariser call through router::chat (ms). Must
+# exceed the router's own stream budget.
+summarizer_timeout_ms: 320000
diff --git a/context-manager/iii.worker.yaml b/context-manager/iii.worker.yaml
new file mode 100644
index 00000000..f42bc6c7
--- /dev/null
+++ b/context-manager/iii.worker.yaml
@@ -0,0 +1,7 @@
+iii: v1
+name: context-manager
+language: rust
+deploy: binary
+manifest: Cargo.toml
+bin: context-manager
+description: Turns a raw conversation history plus a target model into a model-ready context — token counting, function-result pruning, and history compaction.
diff --git a/context-manager/src/adapters/fs_lease.rs b/context-manager/src/adapters/fs_lease.rs
new file mode 100644
index 00000000..c1506b2e
--- /dev/null
+++ b/context-manager/src/adapters/fs_lease.rs
@@ -0,0 +1,322 @@
+//! Filesystem-backed compaction-lease storage, mirroring
+//! `session-manager`'s `FsStore` strategy (`session-manager/src/store/fs.rs`):
+//! one file per lease key under `<lease_dir>/<scope>/`, a process-local
+//! `Mutex<HashMap>` cache, and atomic `tmp + rename` writes.
+//!
+//! Atomicity is **process-local** — the same single-writer-per-dir
+//! assumption session-manager makes — which is sufficient for a single
+//! context-manager instance. `swap` (the atomic read-modify-write
+//! `core::lease` relies on) holds the cache mutex across the whole
+//! read/decide/write, so two concurrent acquirers in this process
+//! serialize: exactly one sees a free/expired prior and wins.
+//!
+//! Each lease file holds exactly one `{nonce, ts}` value, so every write
+//! is a full O(1) rewrite (session-manager's `persist_snapshot`
+//! mechanism); there is no append/replay-merge to do. A missing file —
+//! or a file whose contents aren't a `{nonce, ts}` record — reads as free
+//! (`None`), matching the state-backed store's `parse_record` and
+//! session-manager's malformed-skip posture.
+
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::Mutex;
+
+use async_trait::async_trait;
+
+use crate::ports::{LeaseRecord, LeaseStore};
+
+/// Encode a lease key into a safe filename stem: `[A-Za-z0-9._-]` pass
+/// through, everything else becomes `%XX` (uppercase hex). Keys are
+/// caller-supplied (`options.lease_key`, e.g. a session id) or a sha256
+/// hex digest, so the encoding both keeps filenames portable and blocks
+/// path traversal. Ported from session-manager's `encode_session_id`.
+fn encode_key(key: &str) -> String {
+    let mut out = String::with_capacity(key.len());
+    for byte in key.bytes() {
+        match byte {
+            b'a'..=b'z' | b'A'..=b'Z' | b'0'..=b'9' | b'.' | b'_' | b'-' => out.push(byte as char),
+            other => out.push_str(&format!("%{other:02X}")),
+        }
+    }
+    out
+}
+
+/// Inverse of [`encode_key`]; `None` for malformed escapes. Only the
+/// encode direction is needed in production (keys never round-trip back
+/// from filenames), so this is the exact session-manager inverse kept for
+/// the encoding-safety test.
+#[cfg(test)]
+fn decode_key(stem: &str) -> Option<String> {
+    let bytes = stem.as_bytes();
+    let mut out = Vec::with_capacity(bytes.len());
+    let mut i = 0;
+    while i < bytes.len() {
+        if bytes[i] == b'%' {
+            let hex = bytes.get(i + 1..i + 3)?;
+            let hex = std::str::from_utf8(hex).ok()?;
+            out.push(u8::from_str_radix(hex, 16).ok()?);
+            i += 3;
+        } else {
+            out.push(bytes[i]);
+            i += 1;
+        }
+    }
+    String::from_utf8(out).ok()
+}
+
+/// `(scope, key) -> Option<LeaseRecord>` cache. A present `None` means
+/// "loaded and free"; an absent entry means "not yet read from disk"
+/// (cold), mirroring session-manager's `contains_key` cold check.
+type Cache = HashMap<(String, String), Option<LeaseRecord>>;
+
+pub struct FsLeaseStore {
+    dir: PathBuf,
+    cache: Mutex<Cache>,
+}
+
+impl FsLeaseStore {
+    /// Open (and create if needed) the lease directory.
+    pub fn new(dir: impl Into<PathBuf>) -> Result<Self, String> {
+        let dir = dir.into();
+        std::fs::create_dir_all(&dir)
+            .map_err(|e| format!("create lease_dir {}: {e}", dir.display()))?;
+        Ok(Self {
+            dir,
+            cache: Mutex::new(HashMap::new()),
+        })
+    }
+
+    fn lock(&self) -> std::sync::MutexGuard<'_, Cache> {
+        self.cache
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+    }
+
+    fn scope_dir(&self, scope: &str) -> PathBuf {
+        self.dir.join(encode_key(scope))
+    }
+
+    fn file_path(&self, scope: &str, key: &str) -> PathBuf {
+        self.scope_dir(scope)
+            .join(format!("{}.json", encode_key(key)))
+    }
+
+    /// Read the on-disk record for a key. Missing file -> `None` (free);
+    /// contents that don't parse as a `{nonce, ts}` record -> `None` too.
+    fn read_file(&self, scope: &str, key: &str) -> Result<Option<LeaseRecord>, String> {
+        let path = self.file_path(scope, key);
+        match std::fs::read_to_string(&path) {
+            Ok(contents) => Ok(serde_json::from_str::<LeaseRecord>(&contents).ok()),
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(None),
+            Err(e) => Err(format!("read {}: {e}", path.display())),
+        }
+    }
+
+    /// Atomically write one record: tmp + rename, creating the scope dir
+    /// on demand. Mirrors session-manager's `persist_snapshot`.
+    fn write_file(&self, scope: &str, key: &str, record: &LeaseRecord) -> Result<(), String> {
+        let dir = self.scope_dir(scope);
+        std::fs::create_dir_all(&dir)
+            .map_err(|e| format!("create lease scope dir {}: {e}", dir.display()))?;
+        let path = self.file_path(scope, key);
+        let body =
+            serde_json::to_string(record).map_err(|e| format!("serialize lease record: {e}"))?;
+        let tmp = path.with_extension("json.tmp");
+        std::fs::write(&tmp, body).map_err(|e| format!("write {}: {e}", tmp.display()))?;
+        std::fs::rename(&tmp, &path)
+            .map_err(|e| format!("rename {} -> {}: {e}", tmp.display(), path.display()))?;
+        Ok(())
+    }
+
+    /// Remove a key's file. A missing file is a no-op.
+    fn remove_file(&self, scope: &str, key: &str) -> Result<(), String> {
+        let path = self.file_path(scope, key);
+        match std::fs::remove_file(&path) {
+            Ok(()) => Ok(()),
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(()),
+            Err(e) => Err(format!("remove {}: {e}", path.display())),
+        }
+    }
+
+    /// Populate the cache entry for `(scope, key)` from disk on a cold
+    /// miss. Must be called with `cache` already locked.
+    fn ensure_loaded(&self, cache: &mut Cache, scope: &str, key: &str) -> Result<(), String> {
+        use std::collections::hash_map::Entry;
+        if let Entry::Vacant(slot) = cache.entry((scope.to_string(), key.to_string())) {
+            let record = self.read_file(scope, key)?;
+            slot.insert(record);
+        }
+        Ok(())
+    }
+}
+
+#[async_trait]
+impl LeaseStore for FsLeaseStore {
+    async fn get(&self, scope: &str, key: &str) -> Result<Option<LeaseRecord>, String> {
+        let mut cache = self.lock();
+        self.ensure_loaded(&mut cache, scope, key)?;
+        Ok(cache
+            .get(&(scope.to_string(), key.to_string()))
+            .cloned()
+            .flatten())
+    }
+
+    async fn set(&self, scope: &str, key: &str, value: Option<LeaseRecord>) -> Result<(), String> {
+        let mut cache = self.lock();
+        let k = (scope.to_string(), key.to_string());
+        // Disk first, cache second: a failed write must never leave the
+        // cache claiming state the file doesn't have.
+        match value {
+            Some(record) => {
+                self.write_file(scope, key, &record)?;
+                cache.insert(k, Some(record));
+            }
+            None => {
+                self.remove_file(scope, key)?;
+                cache.insert(k, None);
+            }
+        }
+        Ok(())
+    }
+
+    async fn swap(
+        &self,
+        scope: &str,
+        key: &str,
+        value: LeaseRecord,
+    ) -> Result<Option<LeaseRecord>, String> {
+        // The whole read-modify-write runs under the cache mutex, so two
+        // concurrent acquirers serialize and exactly one observes the
+        // free/expired prior. No `.await` is taken while the guard is
+        // held (all I/O below is synchronous), so the future stays Send.
+        let mut cache = self.lock();
+        self.ensure_loaded(&mut cache, scope, key)?;
+        let k = (scope.to_string(), key.to_string());
+        let prior = cache.get(&k).cloned().flatten();
+        self.write_file(scope, key, &value)?;
+        cache.insert(k, Some(value));
+        Ok(prior)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const SCOPE: &str = "context_lease";
+
+    fn rec(nonce: &str, ts: i64) -> LeaseRecord {
+        LeaseRecord {
+            nonce: nonce.into(),
+            ts,
+        }
+    }
+
+    #[tokio::test]
+    async fn set_get_roundtrip_and_restart_replay() {
+        let dir = tempfile::tempdir().unwrap();
+        {
+            let store = FsLeaseStore::new(dir.path()).unwrap();
+            store.set(SCOPE, "k1", Some(rec("n1", 1000))).await.unwrap();
+            assert_eq!(store.get(SCOPE, "k1").await.unwrap(), Some(rec("n1", 1000)));
+        }
+        // Fresh store over the same dir = worker restart: cache is cold,
+        // so the value must replay from disk.
+        let store = FsLeaseStore::new(dir.path()).unwrap();
+        assert_eq!(store.get(SCOPE, "k1").await.unwrap(), Some(rec("n1", 1000)));
+    }
+
+    #[tokio::test]
+    async fn set_none_removes_the_file() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = FsLeaseStore::new(dir.path()).unwrap();
+        store.set(SCOPE, "k1", Some(rec("n1", 1))).await.unwrap();
+        let path = store.file_path(SCOPE, "k1");
+        assert!(path.exists());
+
+        store.set(SCOPE, "k1", None).await.unwrap();
+        assert!(!path.exists());
+        assert_eq!(store.get(SCOPE, "k1").await.unwrap(), None);
+    }
+
+    #[tokio::test]
+    async fn swap_returns_prior_and_stores_new() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = FsLeaseStore::new(dir.path()).unwrap();
+
+        // First swap on a free key: prior is None.
+        assert_eq!(store.swap(SCOPE, "k1", rec("n1", 1)).await.unwrap(), None);
+        // Second swap: prior is the first claim, new is stored.
+        assert_eq!(
+            store.swap(SCOPE, "k1", rec("n2", 2)).await.unwrap(),
+            Some(rec("n1", 1))
+        );
+        assert_eq!(store.get(SCOPE, "k1").await.unwrap(), Some(rec("n2", 2)));
+    }
+
+    #[tokio::test]
+    async fn cold_swap_sees_the_on_disk_prior() {
+        let dir = tempfile::tempdir().unwrap();
+        {
+            let store = FsLeaseStore::new(dir.path()).unwrap();
+            store.set(SCOPE, "k1", Some(rec("n1", 1))).await.unwrap();
+        }
+        // A fresh store (cold cache) must read the prior from disk so a
+        // racing acquirer after restart can't be treated as a free win.
+        let store = FsLeaseStore::new(dir.path()).unwrap();
+        assert_eq!(
+            store.swap(SCOPE, "k1", rec("n2", 2)).await.unwrap(),
+            Some(rec("n1", 1))
+        );
+    }
+
+    #[test]
+    fn key_encoding_roundtrip_and_safety() {
+        let hostile = "../etc/passwd: weird key?";
+        let encoded = encode_key(hostile);
+        assert!(!encoded.contains('/'));
+        assert!(!encoded.contains(' '));
+        assert!(!encoded.contains(':'));
+        assert_eq!(decode_key(&encoded).as_deref(), Some(hostile));
+
+        // sha256-hex / session-id-style keys pass through unchanged.
+        assert_eq!(encode_key("s_abc-123.x"), "s_abc-123.x");
+        // Malformed escapes are rejected, not mangled.
+        assert_eq!(decode_key("%ZZ"), None);
+        assert_eq!(decode_key("%4"), None);
+    }
+
+    #[tokio::test]
+    async fn hostile_key_file_stays_inside_dir() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = FsLeaseStore::new(dir.path()).unwrap();
+        let key = "../escape/чат 42";
+        store.set(SCOPE, key, Some(rec("n", 1))).await.unwrap();
+
+        // Exactly one file landed under the scope dir, encoded — no
+        // traversal out of `lease_dir`.
+        let scope_dir = store.scope_dir(SCOPE);
+        let files: Vec<_> = std::fs::read_dir(&scope_dir).unwrap().collect();
+        assert_eq!(files.len(), 1);
+        assert_eq!(store.get(SCOPE, key).await.unwrap(), Some(rec("n", 1)));
+    }
+
+    #[tokio::test]
+    async fn malformed_or_foreign_value_reads_as_free() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = FsLeaseStore::new(dir.path()).unwrap();
+
+        // A stored value that isn't a `{nonce, ts}` record reads as free.
+        let path = store.file_path(SCOPE, "k1");
+        std::fs::create_dir_all(path.parent().unwrap()).unwrap();
+        std::fs::write(&path, "\"just a string\"").unwrap();
+        assert_eq!(store.get(SCOPE, "k1").await.unwrap(), None);
+    }
+
+    #[tokio::test]
+    async fn unknown_key_reads_as_none() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = FsLeaseStore::new(dir.path()).unwrap();
+        assert_eq!(store.get(SCOPE, "nope").await.unwrap(), None);
+    }
+}
diff --git a/context-manager/src/adapters/mod.rs b/context-manager/src/adapters/mod.rs
new file mode 100644
index 00000000..166ea80b
--- /dev/null
+++ b/context-manager/src/adapters/mod.rs
@@ -0,0 +1,5 @@
+//! Production adapters behind the ports: `llm-router` calls over the
+//! iii bus and filesystem-backed lease storage.
+
+pub mod fs_lease;
+pub mod router;
diff --git a/context-manager/src/adapters/router.rs b/context-manager/src/adapters/router.rs
new file mode 100644
index 00000000..e8b9c5a9
--- /dev/null
+++ b/context-manager/src/adapters/router.rs
@@ -0,0 +1,303 @@
+//! `llm-router` adapters: model-limit resolution via
+//! `router::models::get` and the summariser via `router::chat`.
+//!
+//! Both are soft dependencies — when the router is absent the resolver
+//! errors (callers degrade to the conservative fallback) and the
+//! summariser reports `Unavailable` (compaction maps it to
+//! `status: "overflow"`).
+
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+
+use async_trait::async_trait;
+use iii_sdk::helpers::create_channel;
+use iii_sdk::{IIIError, TriggerRequest, III};
+use serde_json::{json, Value};
+
+use crate::ports::{ModelResolver, SummarizeError, SummarizeRequest, Summarizer};
+use crate::types::Model;
+
+const MODELS_GET_TIMEOUT_MS: u64 = 5_000;
+/// Grace period for the stream reader to drain after `router::chat`
+/// returns; the terminal `done`/`error` event lands before or with the
+/// trigger response, so this only covers socket latency.
+const READER_DRAIN_MS: u64 = 2_000;
+
+fn is_unroutable(err: &IIIError) -> bool {
+    let msg = err.to_string().to_lowercase();
+    msg.contains("function_not_found") || msg.contains("not found") || msg.contains("no function")
+}
+
+/// `router::models::get` — `{ model: Model } | null`.
+pub struct RouterModelResolver {
+    iii: Arc<III>,
+}
+
+impl RouterModelResolver {
+    pub fn new(iii: Arc<III>) -> Self {
+        Self { iii }
+    }
+}
+
+#[async_trait]
+impl ModelResolver for RouterModelResolver {
+    async fn get_model(&self, provider: Option<&str>, id: &str) -> Result<Option<Model>, String> {
+        let mut payload = json!({ "id": id });
+        if let Some(p) = provider {
+            payload["provider"] = json!(p);
+        }
+        let resp = self
+            .iii
+            .trigger(TriggerRequest {
+                function_id: "router::models::get".to_string(),
+                payload,
+                action: None,
+                timeout_ms: Some(MODELS_GET_TIMEOUT_MS),
+            })
+            .await
+            .map_err(|e| e.to_string())?;
+
+        if resp.is_null() {
+            return Ok(None);
+        }
+        let model = resp
+            .get("model")
+            .cloned()
+            .ok_or_else(|| "router::models::get response has no `model` field".to_string())?;
+        if model.is_null() {
+            return Ok(None);
+        }
+        serde_json::from_value::<Model>(model)
+            .map(Some)
+            .map_err(|e| format!("router::models::get returned an unparseable model: {e}"))
+    }
+}
+
+/// `router::chat` — one summariser turn streamed over a channel the
+/// adapter creates; the summary text comes from the terminal `done`
+/// event (accumulated `text_delta`s as fallback).
+pub struct RouterSummarizer {
+    iii: Arc<III>,
+    timeout_ms: u64,
+}
+
+impl RouterSummarizer {
+    pub fn new(iii: Arc<III>, timeout_ms: u64) -> Self {
+        Self { iii, timeout_ms }
+    }
+}
+
+#[async_trait]
+impl Summarizer for RouterSummarizer {
+    async fn summarize(&self, req: SummarizeRequest) -> Result<String, SummarizeError> {
+        let channel = create_channel(&self.iii, None)
+            .await
+            .map_err(|e| SummarizeError::Unavailable(format!("create_channel failed: {e}")))?;
+
+        // Frames may arrive as WS text messages (dispatched to the
+        // callback) or as binary chunks (returned by read_all); collect
+        // both since the router's framing is provider-defined.
+        let text_frames: Arc<Mutex<Vec<String>>> = Arc::new(Mutex::new(Vec::new()));
+        let sink = text_frames.clone();
+        let reader = channel.reader;
+        reader
+            .on_message(move |frame| {
+                sink.lock()
+                    .unwrap_or_else(|poison| poison.into_inner())
+                    .push(frame);
+            })
+            .await;
+        let read_task = tokio::spawn(async move { reader.read_all().await });
+
+        let mut payload = json!({
+            "writer_ref": channel.writer_ref,
+            "model": req.model,
+            "system_prompt": req.system_prompt,
+            "messages": [{
+                "role": "user",
+                "content": [{ "type": "text", "text": req.user_prompt }],
+                "timestamp": std::time::SystemTime::now()
+                    .duration_since(std::time::UNIX_EPOCH)
+                    .map(|d| d.as_millis() as i64)
+                    .unwrap_or(0),
+            }],
+            "tools": [],
+        });
+        if let Some(p) = &req.provider {
+            payload["provider"] = json!(p);
+        }
+
+        let trigger_result = self
+            .iii
+            .trigger(TriggerRequest {
+                function_id: "router::chat".to_string(),
+                payload,
+                action: None,
+                timeout_ms: Some(self.timeout_ms),
+            })
+            .await;
+
+        // The terminal event precedes the trigger response; give the
+        // socket a short drain window, then stop reading either way.
+        let abort = read_task.abort_handle();
+        let binary =
+            match tokio::time::timeout(Duration::from_millis(READER_DRAIN_MS), read_task).await {
+                Ok(Ok(Ok(bytes))) => bytes,
+                _ => {
+                    abort.abort();
+                    Vec::new()
+                }
+            };
+
+        let response = match trigger_result {
+            Ok(v) => v,
+            Err(e) if is_unroutable(&e) => {
+                return Err(SummarizeError::Unavailable(e.to_string()));
+            }
+            Err(e) => return Err(SummarizeError::Failed(e.to_string())),
+        };
+        if response.get("ok").and_then(Value::as_bool) == Some(false) {
+            let detail = response
+                .get("error")
+                .map(|e| e.to_string())
+                .unwrap_or_else(|| "unknown router error".to_string());
+            return Err(SummarizeError::Failed(detail));
+        }
+
+        let mut frames: Vec<String> = text_frames
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .clone();
+        frames.extend(String::from_utf8_lossy(&binary).lines().map(str::to_string));
+
+        let summary = extract_summary(&frames)?;
+        if summary.is_empty() {
+            return Err(SummarizeError::Empty);
+        }
+        Ok(summary)
+    }
+}
+
+/// Fold a stream of `AssistantMessageEvent` JSON frames into the final
+/// summary text: the `done` event's message wins; accumulated
+/// `text_delta`s are the fallback; a terminal `error` event fails.
+fn extract_summary(frames: &[String]) -> Result<String, SummarizeError> {
+    let mut deltas = String::new();
+    let mut done_text: Option<String> = None;
+
+    for frame in frames {
+        let Ok(event) = serde_json::from_str::<Value>(frame) else {
+            continue;
+        };
+        match event.get("type").and_then(Value::as_str) {
+            Some("text_delta") => {
+                if let Some(d) = event.get("delta").and_then(Value::as_str) {
+                    deltas.push_str(d);
+                }
+            }
+            Some("done") => {
+                done_text = Some(text_of_message(event.get("message")));
+            }
+            Some("error") => {
+                return Err(SummarizeError::Failed(format!(
+                    "summariser stream error: {}",
+                    text_of_message(event.get("error"))
+                )));
+            }
+            Some("stop") if event.get("stop_reason").and_then(Value::as_str) == Some("error") => {
+                let detail = event
+                    .get("error_message")
+                    .and_then(Value::as_str)
+                    .unwrap_or("unknown provider error");
+                return Err(SummarizeError::Failed(format!(
+                    "summariser stream error: {detail}"
+                )));
+            }
+            _ => {}
+        }
+    }
+
+    Ok(done_text.filter(|t| !t.is_empty()).unwrap_or(deltas))
+}
+
+/// Concatenated text blocks of an `AssistantMessage`-shaped value.
+fn text_of_message(message: Option<&Value>) -> String {
+    let Some(blocks) = message
+        .and_then(|m| m.get("content"))
+        .and_then(Value::as_array)
+    else {
+        return String::new();
+    };
+    let mut out = String::new();
+    for block in blocks {
+        if block.get("type").and_then(Value::as_str) == Some("text") {
+            if let Some(text) = block.get("text").and_then(Value::as_str) {
+                if !out.is_empty() {
+                    out.push('\n');
+                }
+                out.push_str(text);
+            }
+        }
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn frame(v: Value) -> String {
+        v.to_string()
+    }
+
+    #[test]
+    fn done_event_text_wins_over_deltas() {
+        let frames = vec![
+            frame(json!({ "type": "text_delta", "partial": {}, "delta": "par" })),
+            frame(json!({ "type": "text_delta", "partial": {}, "delta": "tial" })),
+            frame(json!({ "type": "done", "message": {
+                "role": "assistant",
+                "content": [{ "type": "text", "text": "## Goal\n- finish" }]
+            }})),
+        ];
+        assert_eq!(extract_summary(&frames).unwrap(), "## Goal\n- finish");
+    }
+
+    #[test]
+    fn deltas_are_the_fallback_without_done() {
+        let frames = vec![
+            frame(json!({ "type": "text_delta", "delta": "a" })),
+            frame(json!({ "type": "text_delta", "delta": "b" })),
+            frame(json!({ "type": "ping" })),
+        ];
+        assert_eq!(extract_summary(&frames).unwrap(), "ab");
+    }
+
+    #[test]
+    fn terminal_error_event_fails() {
+        let frames = vec![frame(json!({ "type": "error", "error": {
+            "content": [{ "type": "text", "text": "boom" }]
+        }}))];
+        let err = extract_summary(&frames).unwrap_err();
+        assert!(matches!(err, SummarizeError::Failed(_)));
+        assert!(err.to_string().contains("boom"));
+    }
+
+    #[test]
+    fn stop_with_error_reason_fails() {
+        let frames = vec![frame(
+            json!({ "type": "stop", "stop_reason": "error", "error_message": "ctx overflow" }),
+        )];
+        let err = extract_summary(&frames).unwrap_err();
+        assert!(err.to_string().contains("ctx overflow"));
+    }
+
+    #[test]
+    fn unparseable_frames_are_skipped() {
+        let frames = vec![
+            "not-json".to_string(),
+            frame(json!({ "type": "text_delta", "delta": "ok" })),
+        ];
+        assert_eq!(extract_summary(&frames).unwrap(), "ok");
+    }
+}
diff --git a/context-manager/src/config.rs b/context-manager/src/config.rs
new file mode 100644
index 00000000..c110f54c
--- /dev/null
+++ b/context-manager/src/config.rs
@@ -0,0 +1,416 @@
+//! Operator-facing runtime configuration.
+//!
+//! The authoritative value comes from the `configuration` worker at boot
+//! (see [`crate::configuration`]); a `--config` YAML file, when passed,
+//! only SEEDS the initial registration. Every field has a serde default
+//! so an empty object yields a fully-populated config; per-call request
+//! options override these defaults where the spec allows it (reserved
+//! tokens, tail turns, prune thresholds).
+
+use std::path::PathBuf;
+
+use anyhow::Result;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+
+/// Root config shape. Unknown keys are rejected so a typo'd field
+/// (e.g. `tail_truns: 3`) fails loudly instead of silently running the
+/// default.
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema)]
+#[serde(deny_unknown_fields)]
+pub struct WorkerConfig {
+    /// Cap on the default reserve: `min(cap, context_window * pct/100)`.
+    #[serde(default = "default_reserved_tokens_cap")]
+    pub reserved_tokens_cap: u64,
+
+    /// Percentage of the context window reserved by default.
+    #[serde(default = "default_reserved_pct")]
+    pub reserved_pct: u64,
+
+    /// user+assistant turn pairs kept verbatim by compaction when the
+    /// request omits `options.tail_turns`.
+    #[serde(default = "default_tail_turns")]
+    pub tail_turns: usize,
+
+    /// Newest function-output tokens never pruned (prune default).
+    #[serde(default = "default_protect_recent_tokens")]
+    pub protect_recent_tokens: u64,
+
+    /// Skip pruning entirely when it would free fewer tokens than this.
+    #[serde(default = "default_min_free_tokens")]
+    pub min_free_tokens: u64,
+
+    /// Per-output verbosity threshold (chars): outputs at or under this
+    /// size are never considered verbose enough to prune.
+    #[serde(default = "default_max_output_chars")]
+    pub max_output_chars: usize,
+
+    /// Compaction lease TTL in seconds.
+    #[serde(default = "default_lease_ttl_secs")]
+    pub lease_ttl_secs: u64,
+
+    /// Fall back to conservative limits (8192/1024) when neither inline
+    /// limits nor `llm-router` are available. When false the same
+    /// situation errors with `could not resolve model limits`.
+    #[serde(default = "default_allow_fallback_limits")]
+    pub allow_fallback_limits: bool,
+
+    /// Outer budget for one summariser call through `router::chat` (ms).
+    #[serde(default = "default_summarizer_timeout_ms")]
+    pub summarizer_timeout_ms: u64,
+
+    /// Directory holding compaction-lease files (`<scope>/<key>.json`).
+    /// A leading `~/` expands to the home directory. Mirrors
+    /// session-manager's `data_dir`.
+    #[serde(default = "default_lease_dir")]
+    pub lease_dir: String,
+}
+
+impl WorkerConfig {
+    /// The lease directory with a leading `~/` expanded to `$HOME`.
+    pub fn resolved_lease_dir(&self) -> PathBuf {
+        expand_tilde(&self.lease_dir)
+    }
+
+    /// Parse a seed config from YAML, expanding `${NAME}` against the
+    /// process env FIRST (the seed file is the only path that needs
+    /// expansion — values fetched from `configuration::get` are already
+    /// env-expanded by the configuration worker), then deserializing.
+    pub fn from_yaml(yaml: &str) -> Result<Self, String> {
+        let expanded = expand_env(yaml);
+        serde_yaml::from_str(&expanded).map_err(|e| format!("yaml parse: {e}"))
+    }
+
+    /// Read and parse a YAML seed file (env-expanded — see [`Self::from_yaml`]).
+    pub fn from_file(path: &str) -> Result<Self, String> {
+        let raw = std::fs::read_to_string(path).map_err(|e| format!("read {path}: {e}"))?;
+        Self::from_yaml(&raw)
+    }
+
+    /// Parse a config from a JSON value already env-expanded by the
+    /// configuration worker. Does NOT run `expand_env` (double expansion
+    /// would be a bug) and tolerates a zero-field object (serde defaults
+    /// fill in).
+    pub fn from_json(value: &Value) -> Result<Self, String> {
+        serde_json::from_value(value.clone()).map_err(|e| format!("json parse: {e}"))
+    }
+
+    pub fn to_json(&self) -> Value {
+        serde_json::to_value(self).expect("WorkerConfig serializes")
+    }
+
+    /// The JSON Schema registered with the `configuration` worker. Field
+    /// doc-comments become property descriptions; the shipped defaults
+    /// are attached as a top-level `example`.
+    pub fn json_schema() -> Value {
+        let root = schemars::schema_for!(WorkerConfig);
+        let mut schema =
+            serde_json::to_value(&root.schema).expect("WorkerConfig JSON Schema serializes");
+        if let Some(obj) = schema.as_object_mut() {
+            if !root.definitions.is_empty() {
+                obj.insert(
+                    "definitions".into(),
+                    serde_json::to_value(&root.definitions).expect("definitions serialize"),
+                );
+            }
+            obj.insert("example".into(), WorkerConfig::default().to_json());
+        }
+        schema
+    }
+
+    /// The restart-required fields: everything consumed ONCE at boot
+    /// rather than per call. `summarizer_timeout_ms` is baked into the
+    /// `RouterSummarizer` and `lease_dir` into the `FsLeaseStore`; a live
+    /// config update that changes either is refused on hot-reload (logged
+    /// "restart required", the previous snapshot kept). Every OTHER field
+    /// is a per-call tuning knob that hot-applies.
+    pub fn boot_signature(&self) -> BootSignature {
+        BootSignature {
+            lease_dir: self.lease_dir.clone(),
+            summarizer_timeout_ms: self.summarizer_timeout_ms,
+        }
+    }
+}
+
+/// Signature of the config fields consumed at boot (see
+/// [`WorkerConfig::boot_signature`]). Two configs with an equal signature
+/// differ only in per-call tuning knobs that can be hot-applied; any
+/// other difference requires a worker restart.
+#[derive(Clone, PartialEq, Eq, Debug)]
+pub struct BootSignature {
+    pub lease_dir: String,
+    pub summarizer_timeout_ms: u64,
+}
+
+fn default_reserved_tokens_cap() -> u64 {
+    20_000
+}
+
+fn default_reserved_pct() -> u64 {
+    10
+}
+
+fn default_tail_turns() -> usize {
+    2
+}
+
+fn default_protect_recent_tokens() -> u64 {
+    40_000
+}
+
+fn default_min_free_tokens() -> u64 {
+    20_000
+}
+
+fn default_max_output_chars() -> usize {
+    2_000
+}
+
+fn default_lease_ttl_secs() -> u64 {
+    300
+}
+
+fn default_allow_fallback_limits() -> bool {
+    true
+}
+
+fn default_summarizer_timeout_ms() -> u64 {
+    320_000
+}
+
+fn default_lease_dir() -> String {
+    "~/.iii/data/context-manager".to_string()
+}
+
+fn expand_tilde(path: &str) -> PathBuf {
+    if path == "~" {
+        if let Some(home) = dirs::home_dir() {
+            return home;
+        }
+    } else if let Some(rest) = path.strip_prefix("~/") {
+        if let Some(home) = dirs::home_dir() {
+            return home.join(rest);
+        }
+    }
+    PathBuf::from(path)
+}
+
+/// Expand `${NAME}` occurrences against the process environment. Unknown
+/// variables expand to the empty string and emit a tracing warning. Only
+/// the `--config` seed path uses this — values from `configuration::get`
+/// are already expanded by the worker. An unterminated `${` is a literal.
+fn expand_env(input: &str) -> String {
+    let mut out = String::with_capacity(input.len());
+    let mut rest = input;
+    while let Some(start) = rest.find("${") {
+        out.push_str(&rest[..start]);
+        let after = &rest[start + 2..];
+        match after.find('}') {
+            Some(end) => {
+                let name = &after[..end];
+                match std::env::var(name) {
+                    Ok(v) => out.push_str(&v),
+                    Err(_) => tracing::warn!(var = %name, "config references undefined env var"),
+                }
+                rest = &after[end + 1..];
+            }
+            None => {
+                out.push_str("${");
+                rest = after;
+            }
+        }
+    }
+    out.push_str(rest);
+    out
+}
+
+impl Default for WorkerConfig {
+    fn default() -> Self {
+        Self {
+            reserved_tokens_cap: default_reserved_tokens_cap(),
+            reserved_pct: default_reserved_pct(),
+            tail_turns: default_tail_turns(),
+            protect_recent_tokens: default_protect_recent_tokens(),
+            min_free_tokens: default_min_free_tokens(),
+            max_output_chars: default_max_output_chars(),
+            lease_ttl_secs: default_lease_ttl_secs(),
+            allow_fallback_limits: default_allow_fallback_limits(),
+            summarizer_timeout_ms: default_summarizer_timeout_ms(),
+            lease_dir: default_lease_dir(),
+        }
+    }
+}
+
+pub fn load_config(path: &str) -> Result<WorkerConfig> {
+    let contents = std::fs::read_to_string(path)?;
+    let cfg: WorkerConfig = serde_yaml::from_str(&contents)?;
+    Ok(cfg)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn defaults_from_empty_yaml() {
+        let cfg: WorkerConfig = serde_yaml::from_str("{}").unwrap();
+        assert_eq!(cfg.reserved_tokens_cap, 20_000);
+        assert_eq!(cfg.reserved_pct, 10);
+        assert_eq!(cfg.tail_turns, 2);
+        assert_eq!(cfg.protect_recent_tokens, 40_000);
+        assert_eq!(cfg.min_free_tokens, 20_000);
+        assert_eq!(cfg.max_output_chars, 2_000);
+        assert_eq!(cfg.lease_ttl_secs, 300);
+        assert!(cfg.allow_fallback_limits);
+        assert_eq!(cfg.summarizer_timeout_ms, 320_000);
+        assert_eq!(cfg.lease_dir, "~/.iii/data/context-manager");
+    }
+
+    #[test]
+    fn custom_yaml_overrides_every_field() {
+        let yaml = "reserved_tokens_cap: 1\nreserved_pct: 2\ntail_turns: 3\n\
+                    protect_recent_tokens: 4\nmin_free_tokens: 5\nmax_output_chars: 6\n\
+                    lease_ttl_secs: 7\nallow_fallback_limits: false\nsummarizer_timeout_ms: 8\n\
+                    lease_dir: /tmp/leases";
+        let cfg: WorkerConfig = serde_yaml::from_str(yaml).unwrap();
+        assert_eq!(cfg.reserved_tokens_cap, 1);
+        assert_eq!(cfg.reserved_pct, 2);
+        assert_eq!(cfg.tail_turns, 3);
+        assert_eq!(cfg.protect_recent_tokens, 4);
+        assert_eq!(cfg.min_free_tokens, 5);
+        assert_eq!(cfg.max_output_chars, 6);
+        assert_eq!(cfg.lease_ttl_secs, 7);
+        assert!(!cfg.allow_fallback_limits);
+        assert_eq!(cfg.summarizer_timeout_ms, 8);
+        assert_eq!(cfg.lease_dir, "/tmp/leases");
+        assert_eq!(cfg.resolved_lease_dir(), PathBuf::from("/tmp/leases"));
+    }
+
+    #[test]
+    fn lease_dir_tilde_expands_to_home() {
+        let cfg = WorkerConfig::default();
+        let resolved = cfg.resolved_lease_dir();
+        if let Some(home) = dirs::home_dir() {
+            assert!(resolved.starts_with(home));
+            assert!(resolved.ends_with(".iii/data/context-manager"));
+        }
+    }
+
+    #[test]
+    fn impl_default_matches_yaml_defaults() {
+        let from_yaml: WorkerConfig = serde_yaml::from_str("{}").unwrap();
+        assert_eq!(from_yaml, WorkerConfig::default());
+    }
+
+    #[test]
+    fn unknown_root_key_is_rejected_at_parse() {
+        let err = serde_yaml::from_str::<WorkerConfig>("tail_truns: 3").unwrap_err();
+        assert!(err.to_string().contains("unknown field"), "got: {err}");
+    }
+
+    #[test]
+    fn committed_config_yaml_parses_to_defaults() {
+        let cfg = load_config(concat!(env!("CARGO_MANIFEST_DIR"), "/config.yaml")).unwrap();
+        assert_eq!(cfg, WorkerConfig::default());
+    }
+
+    #[test]
+    fn json_schema_has_every_property_with_descriptions_and_example() {
+        let schema = WorkerConfig::json_schema();
+        let props = schema
+            .get("properties")
+            .and_then(|p| p.as_object())
+            .expect("schema has a properties object");
+        for field in [
+            "reserved_tokens_cap",
+            "reserved_pct",
+            "tail_turns",
+            "protect_recent_tokens",
+            "min_free_tokens",
+            "max_output_chars",
+            "lease_ttl_secs",
+            "allow_fallback_limits",
+            "summarizer_timeout_ms",
+            "lease_dir",
+        ] {
+            assert!(
+                props.get(field).is_some(),
+                "missing schema property {field}"
+            );
+        }
+        // Field doc-comments survive as schema descriptions.
+        assert!(props["lease_dir"].get("description").is_some());
+        // The shipped defaults are attached as a top-level example.
+        assert_eq!(
+            schema.get("example"),
+            Some(&WorkerConfig::default().to_json())
+        );
+    }
+
+    #[test]
+    fn from_json_round_trips_from_default() {
+        let cfg = WorkerConfig::default();
+        let back = WorkerConfig::from_json(&cfg.to_json()).unwrap();
+        assert_eq!(back, cfg);
+    }
+
+    #[test]
+    fn from_json_tolerates_empty_object() {
+        let back = WorkerConfig::from_json(&serde_json::json!({})).unwrap();
+        assert_eq!(back, WorkerConfig::default());
+    }
+
+    #[test]
+    fn from_json_round_trips_custom_values() {
+        let json = serde_json::json!({ "tail_turns": 5, "lease_dir": "/tmp/x" });
+        let cfg = WorkerConfig::from_json(&json).unwrap();
+        assert_eq!(cfg.tail_turns, 5);
+        assert_eq!(cfg.lease_dir, "/tmp/x");
+        // Unspecified fields fall back to serde defaults.
+        assert_eq!(cfg.summarizer_timeout_ms, 320_000);
+    }
+
+    #[test]
+    fn from_json_rejects_garbage() {
+        let err = WorkerConfig::from_json(&serde_json::json!({ "tail_turns": "nan" })).unwrap_err();
+        assert!(err.contains("json parse"), "got: {err}");
+        let err = WorkerConfig::from_json(&serde_json::json!("garbage")).unwrap_err();
+        assert!(err.contains("json parse"), "got: {err}");
+    }
+
+    #[test]
+    fn from_yaml_expands_env_var() {
+        std::env::set_var("CTX_MGR_TEST_LEASE_DIR", "/tmp/expanded-leases");
+        let cfg = WorkerConfig::from_yaml("lease_dir: \"${CTX_MGR_TEST_LEASE_DIR}\"\n").unwrap();
+        assert_eq!(cfg.lease_dir, "/tmp/expanded-leases");
+        std::env::remove_var("CTX_MGR_TEST_LEASE_DIR");
+    }
+
+    #[test]
+    fn boot_signature_equal_when_only_tuning_knobs_differ() {
+        let base = WorkerConfig::default();
+        let tuned = WorkerConfig {
+            tail_turns: base.tail_turns + 1,
+            reserved_pct: base.reserved_pct + 1,
+            protect_recent_tokens: base.protect_recent_tokens + 1,
+            ..base.clone()
+        };
+        assert_eq!(base.boot_signature(), tuned.boot_signature());
+    }
+
+    #[test]
+    fn boot_signature_differs_on_restart_required_fields() {
+        let base = WorkerConfig::default();
+        let moved_dir = WorkerConfig {
+            lease_dir: "/tmp/other".to_string(),
+            ..base.clone()
+        };
+        let new_timeout = WorkerConfig {
+            summarizer_timeout_ms: base.summarizer_timeout_ms + 1,
+            ..base.clone()
+        };
+        assert_ne!(base.boot_signature(), moved_dir.boot_signature());
+        assert_ne!(base.boot_signature(), new_timeout.boot_signature());
+    }
+}
diff --git a/context-manager/src/configuration.rs b/context-manager/src/configuration.rs
new file mode 100644
index 00000000..e1b234fe
--- /dev/null
+++ b/context-manager/src/configuration.rs
@@ -0,0 +1,289 @@
+//! Integration with the `configuration` worker — register the schema,
+//! fetch the authoritative value at boot, and hot-reload it when it
+//! changes. Mirrors [`shell`](../../shell/src/configuration.rs) /
+//! [`database`](../../database/src/configuration.rs) /
+//! [`coder`](../../coder/src/configuration.rs).
+//!
+//! context-manager's config splits into two halves on a live update:
+//!
+//! - The BOOT SIGNATURE (`lease_dir` + `summarizer_timeout_ms`) is
+//!   everything consumed ONCE at startup: `lease_dir` builds the
+//!   `FsLeaseStore` and `summarizer_timeout_ms` the `RouterSummarizer`.
+//!   A config change that alters either is REFUSED on hot-reload (logged
+//!   "restart required", the previous snapshot kept) — those adapters are
+//!   built once at boot and never rebuilt.
+//! - Every OTHER field is a per-call tuning knob (token reserves, prune
+//!   thresholds, compaction tail, lease TTL). When a freshly-fetched
+//!   config's boot signature matches the boot-time signature, the
+//!   snapshot is swapped live; handlers read the current snapshot per
+//!   call via [`Deps::config`](crate::ports::Deps::config).
+
+use std::sync::Arc;
+use std::time::Duration;
+
+use iii_sdk::{IIIError, RegisterFunction, RegisterTriggerInput, TriggerRequest, III};
+use serde_json::{json, Value};
+use tokio::sync::RwLock;
+
+use crate::config::{BootSignature, WorkerConfig};
+
+/// Hot-swappable config snapshot shared with every handler. The
+/// `Arc<RwLock<Arc<WorkerConfig>>>` shape lets a handler take a
+/// `read().await` and `clone()` the inner `Arc` out (a cheap refcount
+/// bump) without holding the lock across its work, while `apply_config`
+/// whole-snapshot replaces the inner `Arc` under the write lock.
+pub type ConfigCell = Arc<RwLock<Arc<WorkerConfig>>>;
+
+pub const CONFIG_ID: &str = "context-manager";
+const CONFIG_FN_ID: &str = "context::on-config-change";
+const CONFIG_TIMEOUT_MS: u64 = 5_000;
+const CONFIG_RETRIES: u32 = 3;
+/// Base backoff between configuration RPC retries; multiplied by the
+/// attempt number for a linear backoff (250ms, 500ms, …).
+const CONFIG_RETRY_BACKOFF_MS: u64 = 250;
+
+/// Register the `context-manager` configuration schema with the
+/// configuration worker. When `seed` is present, its value is installed
+/// as `initial_value`. Otherwise, the built-in default is seeded only
+/// when no stored value exists yet (re-registration preserves the stored
+/// value, so this is safe to call every boot).
+pub async fn register_config(iii: &III, seed: Option<&WorkerConfig>) -> Result<(), String> {
+    let mut payload = json!({
+        "id": CONFIG_ID,
+        "name": "Context Manager",
+        "description": "Model-ready context assembly tuning: token-budget reserves, \
+                        function-result prune thresholds, compaction tail size, the \
+                        compaction lease TTL and directory, and the summariser timeout.",
+        "schema": WorkerConfig::json_schema(),
+    });
+    if let Some(seed) = seed {
+        payload["initial_value"] = seed.to_json();
+    } else if should_seed_default_value(iii).await? {
+        payload["initial_value"] = WorkerConfig::default().to_json();
+    }
+    trigger_with_retry(iii, "configuration::register", payload).await?;
+    Ok(())
+}
+
+/// Read the live `context-manager` configuration (env-expanded by the
+/// configuration worker — `from_json` does NOT re-expand).
+pub async fn fetch_config(iii: &III) -> Result<WorkerConfig, String> {
+    let value = get_config_value(iii).await?;
+    if value.is_null() {
+        tracing::info!("no configuration value found; using built-in default configuration");
+        return Ok(WorkerConfig::default());
+    }
+    WorkerConfig::from_json(&value)
+}
+
+async fn should_seed_default_value(iii: &III) -> Result<bool, String> {
+    match try_get_config_value(iii).await? {
+        None => Ok(true),
+        Some(value) if value.is_null() => Ok(true),
+        Some(_) => Ok(false),
+    }
+}
+
+async fn get_config_value(iii: &III) -> Result<Value, String> {
+    try_get_config_value(iii)
+        .await?
+        .ok_or_else(|| format!("configuration `{CONFIG_ID}` not found"))
+}
+
+/// Returns `Ok(None)` when the entry does not exist. The engine's
+/// missing-entry codes vary in case (`function_not_found`,
+/// `STATEMENT_NOT_FOUND`, `NOT_FOUND`), so match case-insensitively.
+async fn try_get_config_value(iii: &III) -> Result<Option<Value>, String> {
+    match trigger_with_retry(iii, "configuration::get", json!({ "id": CONFIG_ID })).await {
+        Ok(resp) => Ok(resp.get("value").cloned()),
+        Err(e) if e.to_ascii_uppercase().contains("NOT_FOUND") => Ok(None),
+        Err(e) => Err(e),
+    }
+}
+
+/// Swap the config snapshot under the write lock. No adapter rebuild —
+/// the boot signature is unchanged by construction (the caller has
+/// already passed [`reloadable`]).
+pub async fn apply_config(cell: &ConfigCell, cfg: WorkerConfig) {
+    *cell.write().await = Arc::new(cfg);
+}
+
+/// Decide whether a freshly-fetched config can be hot-applied. Returns
+/// the config when its boot signature matches the boot-time signature
+/// (only per-call tuning knobs changed), or an error describing the
+/// restart-required change.
+fn reloadable(cfg: WorkerConfig, boot_sig: &BootSignature) -> Result<WorkerConfig, String> {
+    if cfg.boot_signature() != *boot_sig {
+        return Err(
+            "configuration change alters lease_dir or summarizer_timeout_ms — these are \
+             consumed once at boot (the FsLeaseStore and RouterSummarizer are built then and \
+             never rebuilt); a worker restart is required to apply them"
+                .to_string(),
+        );
+    }
+    Ok(cfg)
+}
+
+/// Register the internal config-change handler and bind a `configuration`
+/// trigger. `boot_sig` is the signature captured at startup; any reload
+/// that would change it is refused (those require a worker restart).
+pub fn register_config_trigger(
+    iii: &III,
+    cell: ConfigCell,
+    boot_sig: BootSignature,
+) -> Result<(), IIIError> {
+    let cell_for_fn = cell.clone();
+    let engine = iii.clone();
+    iii.register_function(
+        CONFIG_FN_ID,
+        RegisterFunction::new_async(move |_payload: Value| {
+            let cell = cell_for_fn.clone();
+            let engine = engine.clone();
+            let boot_sig = boot_sig.clone();
+            async move {
+                on_config_change(&engine, &cell, &boot_sig).await;
+                Ok::<Value, IIIError>(json!({ "ok": true }))
+            }
+        })
+        .description(
+            "Internal: reload context-manager's tuning knobs from the authoritative \
+             configuration when it changes; lease_dir / summarizer_timeout_ms changes require \
+             a restart.",
+        ),
+    );
+
+    iii.register_trigger(RegisterTriggerInput {
+        trigger_type: "configuration".to_string(),
+        function_id: CONFIG_FN_ID.to_string(),
+        config: json!({
+            "configuration_id": CONFIG_ID,
+            "event_types": ["configuration:updated"],
+        }),
+        metadata: None,
+    })?;
+    Ok(())
+}
+
+/// Reload tuning knobs from the AUTHORITATIVE configuration.
+///
+/// The caller-supplied trigger payload is intentionally ignored:
+/// `context::on-config-change` is a discoverable bus function, so trusting
+/// `payload.new_value` would let any caller inject arbitrary config without
+/// updating persisted state. Re-fetch the stored value via
+/// `configuration::get` instead. A restart-required change is refused; the
+/// previous snapshot is always kept on any failure path.
+async fn on_config_change(iii: &III, cell: &ConfigCell, boot_sig: &BootSignature) {
+    let cfg = match fetch_config(iii).await {
+        Ok(cfg) => cfg,
+        Err(e) => {
+            tracing::error!(
+                error = %e,
+                "config-change: failed to fetch authoritative configuration; keeping previous config"
+            );
+            return;
+        }
+    };
+    let cfg = match reloadable(cfg, boot_sig) {
+        Ok(cfg) => cfg,
+        Err(reason) => {
+            tracing::warn!(
+                reason = %reason,
+                "config-change refused: restart required; keeping previous config"
+            );
+            return;
+        }
+    };
+    apply_config(cell, cfg).await;
+    tracing::info!("context-manager tuning knobs reloaded (boot signature unchanged)");
+}
+
+async fn trigger_with_retry(iii: &III, function_id: &str, payload: Value) -> Result<Value, String> {
+    let mut last_err = String::new();
+    for attempt in 1..=CONFIG_RETRIES {
+        match iii
+            .trigger(TriggerRequest {
+                function_id: function_id.to_string(),
+                payload: payload.clone(),
+                action: None,
+                timeout_ms: Some(CONFIG_TIMEOUT_MS),
+            })
+            .await
+        {
+            Ok(v) => return Ok(v),
+            Err(e) => {
+                last_err = e.to_string();
+                if attempt < CONFIG_RETRIES {
+                    tracing::warn!(
+                        function_id,
+                        attempt,
+                        error = %last_err,
+                        "configuration RPC failed; retrying"
+                    );
+                    tokio::time::sleep(Duration::from_millis(
+                        CONFIG_RETRY_BACKOFF_MS * u64::from(attempt),
+                    ))
+                    .await;
+                }
+            }
+        }
+    }
+    Err(format!(
+        "{function_id} failed after {CONFIG_RETRIES} attempts: {last_err}"
+    ))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn reloadable_allows_tuning_only_change() {
+        let boot = WorkerConfig::default();
+        let boot_sig = boot.boot_signature();
+        let next = WorkerConfig {
+            tail_turns: boot.tail_turns + 1,
+            reserved_pct: boot.reserved_pct + 1,
+            ..boot.clone()
+        };
+        let applied = reloadable(next, &boot_sig).expect("tuning-only change is reloadable");
+        assert_eq!(applied.tail_turns, boot.tail_turns + 1);
+    }
+
+    #[test]
+    fn reloadable_refuses_lease_dir_change() {
+        let boot = WorkerConfig::default();
+        let boot_sig = boot.boot_signature();
+        let moved = WorkerConfig {
+            lease_dir: "/tmp/somewhere-else".to_string(),
+            ..boot.clone()
+        };
+        assert!(reloadable(moved, &boot_sig).is_err());
+    }
+
+    #[test]
+    fn reloadable_refuses_summarizer_timeout_change() {
+        let boot = WorkerConfig::default();
+        let boot_sig = boot.boot_signature();
+        let retimed = WorkerConfig {
+            summarizer_timeout_ms: boot.summarizer_timeout_ms + 1,
+            ..boot.clone()
+        };
+        assert!(reloadable(retimed, &boot_sig).is_err());
+    }
+
+    #[tokio::test]
+    async fn apply_config_swaps_snapshot() {
+        let cell: ConfigCell = Arc::new(RwLock::new(Arc::new(WorkerConfig::default())));
+        assert_eq!(
+            cell.read().await.tail_turns,
+            WorkerConfig::default().tail_turns
+        );
+
+        let tuned = WorkerConfig {
+            tail_turns: 9,
+            ..WorkerConfig::default()
+        };
+        apply_config(&cell, tuned).await;
+        assert_eq!(cell.read().await.tail_turns, 9);
+    }
+}
diff --git a/context-manager/src/core/budget.rs b/context-manager/src/core/budget.rs
new file mode 100644
index 00000000..fd2b70af
--- /dev/null
+++ b/context-manager/src/core/budget.rs
@@ -0,0 +1,213 @@
+//! The model-adaptive token budget (context-manager.md § Token budget
+//! model):
+//!
+//! ```text
+//! usable = max(0, (input_limit ?? (context_window - max_output_tokens))
+//!              - reserved - thinking_budget)
+//! ```
+//!
+//! `reserved` defaults to `min(cap, pct% of context_window)`; the
+//! thinking budget comes from the model's declared budget for the
+//! requested tier (0 when the caller passes no tier or the model
+//! declares none).
+
+use crate::config::WorkerConfig;
+use crate::types::{Model, ModelLimits, ThinkingLevel};
+
+/// How the limits in a [`ResolvedModel`] were obtained — echoed to
+/// callers as `model_resolved` so a silent fallback is detectable.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ModelResolved {
+    Inline,
+    Router,
+    Fallback,
+}
+
+impl ModelResolved {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            ModelResolved::Inline => "inline",
+            ModelResolved::Router => "router",
+            ModelResolved::Fallback => "fallback",
+        }
+    }
+}
+
+/// Limits plus the metadata budget math needs, with provenance.
+#[derive(Debug, Clone, PartialEq)]
+pub struct ResolvedModel {
+    pub limits: ModelLimits,
+    /// Declared per-tier reasoning budgets (router-resolved models
+    /// only; inline limits carry none).
+    pub thinking_budgets: Option<std::collections::BTreeMap<ThinkingLevel, u64>>,
+    pub resolved: ModelResolved,
+}
+
+/// The conservative fallback used when neither inline limits nor the
+/// router can resolve the model.
+pub fn fallback_model() -> ResolvedModel {
+    ResolvedModel {
+        limits: ModelLimits {
+            context_window: 8_192,
+            max_output_tokens: 1_024,
+            input_limit: None,
+        },
+        thinking_budgets: None,
+        resolved: ModelResolved::Fallback,
+    }
+}
+
+impl ResolvedModel {
+    pub fn from_inline(limits: ModelLimits) -> Self {
+        Self {
+            limits,
+            thinking_budgets: None,
+            resolved: ModelResolved::Inline,
+        }
+    }
+
+    pub fn from_router(model: &Model) -> Self {
+        Self {
+            limits: ModelLimits {
+                context_window: model.context_window,
+                max_output_tokens: model.max_output_tokens,
+                input_limit: model.input_limit,
+            },
+            thinking_budgets: model.thinking_budgets.clone(),
+            resolved: ModelResolved::Router,
+        }
+    }
+
+    /// The model's declared budget for `level`, else 0.
+    pub fn thinking_budget(&self, level: Option<ThinkingLevel>) -> u64 {
+        match (level, &self.thinking_budgets) {
+            (Some(level), Some(budgets)) => budgets.get(&level).copied().unwrap_or(0),
+            _ => 0,
+        }
+    }
+}
+
+/// Default reserve: `min(cap, context_window * pct / 100)`.
+pub fn default_reserved(cfg: &WorkerConfig, context_window: u64) -> u64 {
+    cfg.reserved_tokens_cap
+        .min(context_window.saturating_mul(cfg.reserved_pct) / 100)
+}
+
+/// The usable input budget for one call. Saturating throughout — a
+/// tiny model with a large reserve clamps to 0 instead of wrapping.
+pub fn usable(limits: &ModelLimits, reserved: u64, thinking_budget: u64) -> u64 {
+    let base = limits.input_limit.unwrap_or_else(|| {
+        limits
+            .context_window
+            .saturating_sub(limits.max_output_tokens)
+    });
+    base.saturating_sub(reserved)
+        .saturating_sub(thinking_budget)
+}
+
+/// Adaptive verbatim-tail budget for compaction: 25% of `usable`,
+/// clamped to [2_000, 8_000]; the caller's `preserve_recent_tokens`
+/// overrides it (same policy as the harness prior art).
+pub fn preserve_recent_budget(usable: u64, override_tokens: Option<u64>) -> u64 {
+    const MIN_PRESERVE_RECENT_TOKENS: u64 = 2_000;
+    const MAX_PRESERVE_RECENT_TOKENS: u64 = 8_000;
+    if let Some(ovr) = override_tokens {
+        return ovr;
+    }
+    (usable / 4).clamp(MIN_PRESERVE_RECENT_TOKENS, MAX_PRESERVE_RECENT_TOKENS)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn limits(
+        context_window: u64,
+        max_output_tokens: u64,
+        input_limit: Option<u64>,
+    ) -> ModelLimits {
+        ModelLimits {
+            context_window,
+            max_output_tokens,
+            input_limit,
+        }
+    }
+
+    #[test]
+    fn reserved_defaults_to_ten_pct_capped_at_20k() {
+        let cfg = WorkerConfig::default();
+        // 10% of 200k = 20k, equal to the cap.
+        assert_eq!(default_reserved(&cfg, 200_000), 20_000);
+        // 10% of 32k = 3.2k, under the cap.
+        assert_eq!(default_reserved(&cfg, 32_000), 3_200);
+        // 10% of 1M = 100k, capped at 20k.
+        assert_eq!(default_reserved(&cfg, 1_000_000), 20_000);
+    }
+
+    #[test]
+    fn spec_examples_hold() {
+        let cfg = WorkerConfig::default();
+        // "A 200k model with defaults yields ~180k usable"
+        let l = limits(200_000, 8_000, None);
+        let u = usable(&l, default_reserved(&cfg, l.context_window), 0);
+        assert_eq!(u, 172_000); // 200k - 8k output - 20k reserved
+
+        // "a 32k model yields ~12k" (with a 16k output budget)
+        let l = limits(32_000, 16_000, None);
+        let u = usable(&l, default_reserved(&cfg, l.context_window), 0);
+        assert_eq!(u, 12_800); // 32k - 16k - 3.2k
+    }
+
+    #[test]
+    fn input_limit_takes_precedence_over_derivation() {
+        let l = limits(200_000, 8_000, Some(100_000));
+        assert_eq!(usable(&l, 20_000, 0), 80_000);
+    }
+
+    #[test]
+    fn thinking_budget_subtracts() {
+        let l = limits(200_000, 8_000, None);
+        assert_eq!(usable(&l, 20_000, 16_000), 156_000);
+    }
+
+    #[test]
+    fn tiny_models_clamp_to_zero() {
+        let l = limits(1_000, 2_000, None); // output exceeds window
+        assert_eq!(usable(&l, 0, 0), 0);
+        let l = limits(8_192, 1_024, None);
+        assert_eq!(usable(&l, 10_000, 0), 0); // reserve exceeds base
+    }
+
+    #[test]
+    fn fallback_is_8192_1024() {
+        let f = fallback_model();
+        assert_eq!(f.limits.context_window, 8_192);
+        assert_eq!(f.limits.max_output_tokens, 1_024);
+        assert_eq!(f.resolved, ModelResolved::Fallback);
+    }
+
+    #[test]
+    fn thinking_budget_requires_both_level_and_declaration() {
+        let inline = ResolvedModel::from_inline(limits(100, 10, None));
+        assert_eq!(inline.thinking_budget(Some(ThinkingLevel::High)), 0);
+
+        let mut budgets = std::collections::BTreeMap::new();
+        budgets.insert(ThinkingLevel::High, 9_000);
+        let routed = ResolvedModel {
+            limits: limits(100, 10, None),
+            thinking_budgets: Some(budgets),
+            resolved: ModelResolved::Router,
+        };
+        assert_eq!(routed.thinking_budget(Some(ThinkingLevel::High)), 9_000);
+        assert_eq!(routed.thinking_budget(Some(ThinkingLevel::Low)), 0);
+        assert_eq!(routed.thinking_budget(None), 0);
+    }
+
+    #[test]
+    fn preserve_recent_budget_clamps_and_overrides() {
+        assert_eq!(preserve_recent_budget(40_000, None), 8_000); // 10k capped
+        assert_eq!(preserve_recent_budget(20_000, None), 5_000); // within band
+        assert_eq!(preserve_recent_budget(1_000, None), 2_000); // floor
+        assert_eq!(preserve_recent_budget(40_000, Some(123)), 123);
+    }
+}
diff --git a/context-manager/src/core/estimate.rs b/context-manager/src/core/estimate.rs
new file mode 100644
index 00000000..f01b0e05
--- /dev/null
+++ b/context-manager/src/core/estimate.rs
@@ -0,0 +1,148 @@
+//! Token estimation. v1 ships the chars/4 heuristic (the harness prior
+//! art) behind a trait so a real tokenizer can slot in per model later;
+//! responses report which estimator ran.
+
+use crate::types::{AgentFunction, AgentMessage, Role};
+
+/// Which estimator produced a count — `count-tokens` echoes this.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum EstimatorKind {
+    Tokenizer,
+    Heuristic,
+}
+
+impl EstimatorKind {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            EstimatorKind::Tokenizer => "tokenizer",
+            EstimatorKind::Heuristic => "heuristic",
+        }
+    }
+}
+
+pub trait Estimator: Send + Sync {
+    fn kind(&self) -> EstimatorKind;
+
+    /// Tokens of one message (full serialized form, so structure and
+    /// metadata weigh in, matching what actually crosses the wire).
+    fn message(&self, message: &AgentMessage) -> u64;
+
+    /// Tokens of a bare string (system prompts, summaries).
+    fn text(&self, text: &str) -> u64;
+
+    /// Tokens of one invocation-schema entry.
+    fn function(&self, function: &AgentFunction) -> u64;
+}
+
+/// `serialized JSON chars / 4` — deterministic and model-independent.
+pub struct HeuristicEstimator;
+
+impl Estimator for HeuristicEstimator {
+    fn kind(&self) -> EstimatorKind {
+        EstimatorKind::Heuristic
+    }
+
+    fn message(&self, message: &AgentMessage) -> u64 {
+        let chars = serde_json::to_string(message).map(|s| s.len()).unwrap_or(0);
+        (chars / 4) as u64
+    }
+
+    fn text(&self, text: &str) -> u64 {
+        (text.len() / 4) as u64
+    }
+
+    fn function(&self, function: &AgentFunction) -> u64 {
+        let chars = serde_json::to_string(function)
+            .map(|s| s.len())
+            .unwrap_or(0);
+        (chars / 4) as u64
+    }
+}
+
+/// Estimator selection. v1: always the heuristic; a future tokenizer
+/// keys off the resolved model here.
+pub fn estimator_for_model(_model_id: &str) -> &'static dyn Estimator {
+    static HEURISTIC: HeuristicEstimator = HeuristicEstimator;
+    &HEURISTIC
+}
+
+/// Sum of message estimates.
+pub fn estimate_messages(est: &dyn Estimator, messages: &[AgentMessage]) -> u64 {
+    messages.iter().map(|m| est.message(m)).sum()
+}
+
+/// Per-role breakdown (`count-tokens.by_role`).
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct ByRole {
+    pub user: u64,
+    pub assistant: u64,
+    pub function_result: u64,
+    pub custom: u64,
+}
+
+pub fn estimate_by_role(est: &dyn Estimator, messages: &[AgentMessage]) -> ByRole {
+    let mut by_role = ByRole::default();
+    for m in messages {
+        let tokens = est.message(m);
+        match m.role() {
+            Role::User => by_role.user += tokens,
+            Role::Assistant => by_role.assistant += tokens,
+            Role::FunctionResult => by_role.function_result += tokens,
+            Role::Custom => by_role.custom += tokens,
+        }
+    }
+    by_role
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    fn msg(v: serde_json::Value) -> AgentMessage {
+        serde_json::from_value(v).unwrap()
+    }
+
+    #[test]
+    fn heuristic_is_serialized_chars_over_four() {
+        let m = msg(json!({
+            "role": "user", "content": [{ "type": "text", "text": "hello" }], "timestamp": 1
+        }));
+        let chars = serde_json::to_string(&m).unwrap().len();
+        let est = HeuristicEstimator;
+        assert_eq!(est.message(&m), (chars / 4) as u64);
+        assert_eq!(est.kind(), EstimatorKind::Heuristic);
+    }
+
+    #[test]
+    fn text_estimate_is_chars_over_four() {
+        assert_eq!(HeuristicEstimator.text("x".repeat(400).as_str()), 100);
+        assert_eq!(HeuristicEstimator.text(""), 0);
+    }
+
+    #[test]
+    fn by_role_partitions_every_role() {
+        let messages = vec![
+            msg(
+                json!({ "role": "user", "content": [{ "type": "text", "text": "q" }], "timestamp": 1 }),
+            ),
+            msg(
+                json!({ "role": "assistant", "content": [], "stop_reason": "end",
+                        "model": "m", "provider": "p", "timestamp": 2 }),
+            ),
+            msg(
+                json!({ "role": "function_result", "function_call_id": "c", "function_id": "f",
+                        "content": [], "timestamp": 3 }),
+            ),
+            msg(json!({ "role": "custom", "custom_type": "t", "content": [], "timestamp": 4 })),
+        ];
+        let est = HeuristicEstimator;
+        let by_role = estimate_by_role(&est, &messages);
+        assert!(by_role.user > 0 && by_role.assistant > 0);
+        assert!(by_role.function_result > 0 && by_role.custom > 0);
+        assert_eq!(
+            by_role.user + by_role.assistant + by_role.function_result + by_role.custom,
+            estimate_messages(&est, &messages)
+        );
+    }
+}
diff --git a/context-manager/src/core/lease.rs b/context-manager/src/core/lease.rs
new file mode 100644
index 00000000..05e3bf05
--- /dev/null
+++ b/context-manager/src/core/lease.rs
@@ -0,0 +1,146 @@
+//! Compaction mutual exclusion (context-manager.md § State): a
+//! `{nonce, ts}` claim under scope `context_lease`, keyed by
+//! `options.lease_key` (e.g. a session id) or a hash of the message
+//! set. TTL is enforced by readers — a claim older than the TTL reads
+//! as free, folding crash recovery into acquisition. Ported from
+//! harness `runtime/lease.ts`.
+
+use sha2::{Digest, Sha256};
+
+use crate::ports::{Clock, LeaseRecord, LeaseStore};
+use crate::types::AgentMessage;
+
+/// The single lease scope this worker writes (a subdirectory under the
+/// configured `lease_dir`).
+pub const LEASE_SCOPE: &str = "context_lease";
+
+/// Default lease key: sha256 over the serialized message set, so two
+/// callers compacting the same logical history contend even without an
+/// explicit `lease_key`, while different histories never block each
+/// other. sha2 (not a std hasher) keeps the key stable across
+/// processes and Rust versions.
+pub fn default_lease_key(messages: &[AgentMessage]) -> String {
+    let mut hasher = Sha256::new();
+    for message in messages {
+        hasher.update(serde_json::to_string(message).unwrap_or_default());
+        hasher.update([0u8]); // message separator
+    }
+    format!("{:x}", hasher.finalize())
+}
+
+fn is_active(record: Option<&LeaseRecord>, now_ms: i64, ttl_ms: i64) -> bool {
+    record.is_some_and(|r| now_ms - r.ts < ttl_ms)
+}
+
+/// Try to acquire the lease. Returns the nonce to release with, or
+/// `None` when another holder owns a live claim. Store failures read
+/// as "not acquired" (never a win) — an outage must not let every
+/// contender through, and a transient `busy` is retryable.
+pub async fn acquire(
+    store: &dyn LeaseStore,
+    clock: &dyn Clock,
+    key: &str,
+    ttl_ms: i64,
+) -> Option<String> {
+    let now = clock.now_ms();
+    match store.get(LEASE_SCOPE, key).await {
+        Ok(existing) if is_active(existing.as_ref(), now, ttl_ms) => return None,
+        Ok(_) => {}
+        Err(e) => {
+            tracing::warn!(error = %e, key, "lease get failed; treating as busy");
+            return None;
+        }
+    }
+
+    let nonce = format!("{}-{}", std::process::id(), uuid::Uuid::new_v4().simple());
+    let claim = LeaseRecord {
+        nonce: nonce.clone(),
+        ts: now,
+    };
+    let prior = match store.swap(LEASE_SCOPE, key, claim).await {
+        Ok(prior) => prior,
+        Err(e) => {
+            tracing::warn!(error = %e, key, "lease swap failed; never treating as a win");
+            return None;
+        }
+    };
+
+    // We clobbered a still-valid claim (always someone else's — our
+    // nonce is fresh): restore it and bow out.
+    if is_active(prior.as_ref(), now, ttl_ms) {
+        if let Err(e) = store.set(LEASE_SCOPE, key, prior).await {
+            tracing::warn!(error = %e, key, "lease restore failed");
+        }
+        return None;
+    }
+    Some(nonce)
+}
+
+/// Release only our own claim: a holder that lost its lease to TTL
+/// takeover must not clear the new holder's claim.
+pub async fn release(store: &dyn LeaseStore, key: &str, nonce: &str) {
+    match store.get(LEASE_SCOPE, key).await {
+        Ok(Some(stored)) if stored.nonce == nonce => {
+            if let Err(e) = store.set(LEASE_SCOPE, key, None).await {
+                tracing::warn!(error = %e, key, "lease clear failed");
+            }
+        }
+        Ok(_) => {}
+        Err(e) => tracing::warn!(error = %e, key, "lease release read failed"),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn default_key_is_stable_and_input_sensitive() {
+        let a: AgentMessage = serde_json::from_value(json!({
+            "role": "user", "content": [{ "type": "text", "text": "hi" }], "timestamp": 1
+        }))
+        .unwrap();
+        let b: AgentMessage = serde_json::from_value(json!({
+            "role": "user", "content": [{ "type": "text", "text": "bye" }], "timestamp": 1
+        }))
+        .unwrap();
+
+        let key_a1 = default_lease_key(std::slice::from_ref(&a));
+        let key_a2 = default_lease_key(std::slice::from_ref(&a));
+        let key_b = default_lease_key(std::slice::from_ref(&b));
+        assert_eq!(key_a1, key_a2, "same messages must contend");
+        assert_ne!(
+            key_a1, key_b,
+            "different messages must not block each other"
+        );
+        assert_eq!(key_a1.len(), 64, "sha256 hex");
+    }
+
+    #[test]
+    fn concatenation_is_not_ambiguous() {
+        // ["ab"] vs ["a", "b"]-style boundary games must not collide;
+        // the separator byte keeps message boundaries in the hash.
+        let one: Vec<AgentMessage> = serde_json::from_value(json!([
+            { "role": "user", "content": [{ "type": "text", "text": "ab" }], "timestamp": 1 }
+        ]))
+        .unwrap();
+        let two: Vec<AgentMessage> = serde_json::from_value(json!([
+            { "role": "user", "content": [{ "type": "text", "text": "a" }], "timestamp": 1 },
+            { "role": "user", "content": [{ "type": "text", "text": "b" }], "timestamp": 1 }
+        ]))
+        .unwrap();
+        assert_ne!(default_lease_key(&one), default_lease_key(&two));
+    }
+
+    #[test]
+    fn activity_is_ttl_relative() {
+        let rec = LeaseRecord {
+            nonce: "n".into(),
+            ts: 1_000,
+        };
+        assert!(is_active(Some(&rec), 1_500, 1_000)); // 500ms old, 1s TTL
+        assert!(!is_active(Some(&rec), 2_500, 1_000)); // expired
+        assert!(!is_active(None, 0, 1_000));
+    }
+}
diff --git a/context-manager/src/core/mod.rs b/context-manager/src/core/mod.rs
new file mode 100644
index 00000000..0f38b275
--- /dev/null
+++ b/context-manager/src/core/mod.rs
@@ -0,0 +1,8 @@
+//! Pure context logic: no I/O, no engine, fully unit/BDD-testable.
+
+pub mod budget;
+pub mod estimate;
+pub mod lease;
+pub mod prune;
+pub mod selection;
+pub mod summary;
diff --git a/context-manager/src/core/prune.rs b/context-manager/src/core/prune.rs
new file mode 100644
index 00000000..4e71cb34
--- /dev/null
+++ b/context-manager/src/core/prune.rs
@@ -0,0 +1,282 @@
+//! Function-output pruning (context-manager.md § context::prune):
+//! replace verbose `function_result` outputs with placeholders, newest
+//! to oldest, freeing tokens outside a protected recent window.
+//!
+//! Structural invariant (§ Structural invariants): **prune replaces,
+//! never removes** — the message, its `function_call_id` linkage, and
+//! the message order all survive; only the content is rewritten to a
+//! single text placeholder carrying the freed size.
+//!
+//! Eligibility (ported from harness `context-compaction/prune.ts`):
+//! - the most recent two user turns are never touched;
+//! - outputs inside the newest `protect_recent_tokens` window are kept;
+//! - `protected_functions` are never pruned;
+//! - outputs whose text is at or under `max_output_chars` are not
+//!   "verbose" and stay (pruning them frees almost nothing);
+//! - when everything prunable frees under `min_free_tokens`, nothing is
+//!   touched at all (a no-op beats a destroyed-but-still-over context).
+
+use crate::core::estimate::Estimator;
+use crate::types::{AgentMessage, ContentBlock, Role};
+
+/// Recent user turns that are always exempt, independent of the token
+/// window (prior-art constant, not operator-tunable).
+const PROTECTED_USER_TURNS: usize = 2;
+
+#[derive(Debug, Clone)]
+pub struct PruneParams {
+    /// Newest function-output tokens kept verbatim.
+    pub protect_recent_tokens: u64,
+    /// Skip the whole pass when it would free less than this.
+    pub min_free_tokens: u64,
+    /// Outputs at or under this many chars are not considered verbose.
+    pub max_output_chars: usize,
+    /// `function_id`s whose outputs are never pruned.
+    pub protected_functions: Vec<String>,
+}
+
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct PruneStats {
+    /// Estimated tokens freed.
+    pub pruned_tokens: u64,
+    /// Number of outputs replaced with placeholders.
+    pub pruned_parts: u64,
+    /// Number of prunable outputs examined (excludes protected
+    /// functions and the always-exempt recent turns).
+    pub scanned_parts: u64,
+}
+
+/// The spec's placeholder shape: `[output pruned: was ~N tokens]`.
+pub fn placeholder(tokens: u64) -> String {
+    format!("[output pruned: was ~{tokens} tokens]")
+}
+
+fn text_of(blocks: &[ContentBlock]) -> String {
+    let mut out = String::new();
+    for block in blocks {
+        if let ContentBlock::Text { text } = block {
+            out.push_str(text);
+        }
+    }
+    out
+}
+
+/// Rewrite verbose outputs in place. Returns the stats; `messages` is
+/// mutated only when the pass actually runs (the `min_free_tokens`
+/// guard fires before any rewrite).
+pub fn prune(
+    messages: &mut [AgentMessage],
+    params: &PruneParams,
+    estimator: &dyn Estimator,
+) -> PruneStats {
+    let mut scanned: u64 = 0;
+    let mut window_tokens: u64 = 0;
+    let mut user_turns = 0usize;
+    // (message index, estimated tokens of its text content)
+    let mut queue: Vec<(usize, u64)> = Vec::new();
+
+    for idx in (0..messages.len()).rev() {
+        let message = &messages[idx];
+        if message.role() == Role::User {
+            user_turns += 1;
+            continue;
+        }
+        if user_turns < PROTECTED_USER_TURNS {
+            continue;
+        }
+        let AgentMessage::FunctionResult {
+            function_id,
+            content,
+            ..
+        } = message
+        else {
+            continue;
+        };
+        if params.protected_functions.iter().any(|f| f == function_id) {
+            continue;
+        }
+
+        let text = text_of(content);
+        let tokens = estimator.text(&text);
+        scanned += 1;
+        window_tokens += tokens;
+        if window_tokens <= params.protect_recent_tokens {
+            continue;
+        }
+        if text.len() <= params.max_output_chars {
+            continue;
+        }
+        queue.push((idx, tokens));
+    }
+
+    // Net tokens freed: each pruned output is replaced by a placeholder
+    // that itself costs a few tokens, so the real saving is the original
+    // size minus that placeholder. Gauge `min_free_tokens` on the net.
+    let pruned_tokens: u64 = queue
+        .iter()
+        .map(|(_, tokens)| tokens.saturating_sub(estimator.text(&placeholder(*tokens))))
+        .sum();
+    if pruned_tokens < params.min_free_tokens {
+        return PruneStats {
+            pruned_tokens: 0,
+            pruned_parts: 0,
+            scanned_parts: scanned,
+        };
+    }
+
+    for (idx, tokens) in &queue {
+        messages[*idx].set_content(vec![ContentBlock::Text {
+            text: placeholder(*tokens),
+        }]);
+    }
+
+    PruneStats {
+        pruned_tokens,
+        pruned_parts: queue.len() as u64,
+        scanned_parts: scanned,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::core::estimate::HeuristicEstimator;
+    use serde_json::json;
+
+    fn user(text: &str, ts: i64) -> AgentMessage {
+        serde_json::from_value(json!({
+            "role": "user", "content": [{ "type": "text", "text": text }], "timestamp": ts
+        }))
+        .unwrap()
+    }
+
+    fn result(function_id: &str, chars: usize, ts: i64) -> AgentMessage {
+        serde_json::from_value(json!({
+            "role": "function_result", "function_call_id": format!("c{ts}"),
+            "function_id": function_id,
+            "content": [{ "type": "text", "text": "x".repeat(chars) }],
+            "timestamp": ts
+        }))
+        .unwrap()
+    }
+
+    fn params() -> PruneParams {
+        PruneParams {
+            protect_recent_tokens: 100,
+            min_free_tokens: 1,
+            max_output_chars: 100,
+            protected_functions: vec![],
+        }
+    }
+
+    /// History: old verbose output, then two user turns (exempt zone).
+    fn history() -> Vec<AgentMessage> {
+        vec![
+            user("first", 1),
+            result("shell::run", 8_000, 2), // ~2000 tokens, old
+            user("second", 3),
+            result("shell::run", 8_000, 4), // inside the 2-user-turn exemption
+            user("third", 5),
+        ]
+    }
+
+    #[test]
+    fn prunes_old_verbose_output_and_keeps_recent_turns() {
+        let mut messages = history();
+        let stats = prune(&mut messages, &params(), &HeuristicEstimator);
+        assert_eq!(stats.pruned_parts, 1);
+        // Net of the placeholder written back: 2000 minus the tokens of
+        // "[output pruned: was ~2000 tokens]" (33 chars / 4 = 8).
+        assert_eq!(stats.pruned_tokens, 1_992);
+        // Oldest output replaced...
+        assert_eq!(
+            messages[1].content(),
+            &[ContentBlock::Text {
+                text: placeholder(2_000)
+            }]
+        );
+        // ...the one inside the last two user turns untouched.
+        assert_eq!(text_of(messages[3].content()).len(), 8_000);
+    }
+
+    #[test]
+    fn replaces_but_never_removes() {
+        let mut messages = history();
+        let before = messages.len();
+        prune(&mut messages, &params(), &HeuristicEstimator);
+        assert_eq!(messages.len(), before);
+        let AgentMessage::FunctionResult {
+            function_call_id, ..
+        } = &messages[1]
+        else {
+            panic!("message kind changed");
+        };
+        assert_eq!(function_call_id, "c2");
+    }
+
+    #[test]
+    fn min_free_guard_skips_everything() {
+        let mut messages = history();
+        let mut p = params();
+        p.min_free_tokens = 1_000_000;
+        let stats = prune(&mut messages, &p, &HeuristicEstimator);
+        assert_eq!(stats.pruned_parts, 0);
+        assert_eq!(stats.pruned_tokens, 0);
+        assert_eq!(stats.scanned_parts, 1);
+        assert_eq!(text_of(messages[1].content()).len(), 8_000);
+    }
+
+    #[test]
+    fn protected_functions_are_exempt() {
+        let mut messages = history();
+        let mut p = params();
+        p.protected_functions = vec!["shell::run".into()];
+        let stats = prune(&mut messages, &p, &HeuristicEstimator);
+        assert_eq!(stats.pruned_parts, 0);
+        assert_eq!(stats.scanned_parts, 0);
+    }
+
+    #[test]
+    fn small_outputs_are_not_verbose() {
+        let mut messages = vec![
+            user("first", 1),
+            result("shell::run", 50, 2), // tiny, under max_output_chars
+            user("second", 3),
+            user("third", 4),
+        ];
+        let mut p = params();
+        p.protect_recent_tokens = 0;
+        let stats = prune(&mut messages, &p, &HeuristicEstimator);
+        assert_eq!(stats.pruned_parts, 0);
+        assert_eq!(stats.scanned_parts, 1);
+        assert_eq!(text_of(messages[1].content()).len(), 50);
+    }
+
+    #[test]
+    fn protect_window_counts_newest_first() {
+        // Two old outputs; the window covers the newer one only.
+        let mut messages = vec![
+            user("a", 1),
+            result("f", 4_000, 2), // older: outside window once newer fills it
+            result("f", 4_000, 3), // newer: inside the 1000-token window
+            user("b", 4),
+            user("c", 5),
+        ];
+        let mut p = params();
+        p.protect_recent_tokens = 1_000;
+        let stats = prune(&mut messages, &p, &HeuristicEstimator);
+        assert_eq!(stats.pruned_parts, 1);
+        assert_eq!(text_of(messages[2].content()).len(), 4_000); // newer kept
+        assert!(text_of(messages[1].content()).starts_with("[output pruned"));
+    }
+
+    #[test]
+    fn idempotent_re_prune_is_a_no_op() {
+        let mut messages = history();
+        let first = prune(&mut messages, &params(), &HeuristicEstimator);
+        assert_eq!(first.pruned_parts, 1);
+        let second = prune(&mut messages, &params(), &HeuristicEstimator);
+        // The placeholder is tiny, so nothing is verbose any more.
+        assert_eq!(second.pruned_parts, 0);
+    }
+}
diff --git a/context-manager/src/core/selection.rs b/context-manager/src/core/selection.rs
new file mode 100644
index 00000000..ab3de21f
--- /dev/null
+++ b/context-manager/src/core/selection.rs
@@ -0,0 +1,310 @@
+//! Token-aware verbatim-tail selection for compaction. Pure logic.
+//!
+//! Ported from harness `context-compaction/selection.ts`, hardened for
+//! the spec's structural invariant (context-manager.md § Structural
+//! invariants): a `function_call` and its `function_result` always land
+//! on the same side of the head/tail boundary, so the tail may only
+//! start at a *safe cut* — a user or assistant message (never a
+//! `function_result` message, and never a user message carrying inline
+//! `function_result` blocks).
+
+use crate::core::estimate::Estimator;
+use crate::types::{AgentMessage, Role};
+
+/// One conversation turn: starts at a user message, ends right before
+/// the next one (exclusive).
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct Turn {
+    pub start: usize,
+    pub end: usize,
+}
+
+/// Head/tail split. `tail_start_index` is `None` when nothing could be
+/// safely kept verbatim — the whole history is the head (and the spec's
+/// `tail_start_index` is `null`).
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Selection {
+    pub head_len: usize,
+    pub tail_start_index: Option<usize>,
+}
+
+/// Partition into turns; messages before the first user message belong
+/// to no turn (they are always head material).
+pub fn turns(messages: &[AgentMessage]) -> Vec<Turn> {
+    let mut result: Vec<Turn> = Vec::new();
+    for (i, m) in messages.iter().enumerate() {
+        if m.role() == Role::User {
+            result.push(Turn {
+                start: i,
+                end: messages.len(),
+            });
+        }
+    }
+    for i in 0..result.len().saturating_sub(1) {
+        result[i].end = result[i + 1].start;
+    }
+    result
+}
+
+/// A boundary the tail may start at without orphaning a result:
+/// user/assistant messages only, and the user message must not carry
+/// inline `function_result` blocks (providers reject results whose
+/// call landed in the summarised head).
+fn is_safe_cut(message: &AgentMessage) -> bool {
+    match message.role() {
+        Role::Assistant => true,
+        Role::User => !message.has_function_result_block(),
+        Role::FunctionResult | Role::Custom => false,
+    }
+}
+
+/// Find a partial tail inside `turn` that fits `budget`, scanning from
+/// the oldest in-turn position forward; only safe cuts qualify.
+fn split_turn(
+    messages: &[AgentMessage],
+    turn: Turn,
+    budget: u64,
+    estimator: &dyn Estimator,
+) -> Option<usize> {
+    if budget == 0 || turn.end.saturating_sub(turn.start) <= 1 {
+        return None;
+    }
+    for start in (turn.start + 1)..turn.end {
+        if !is_safe_cut(&messages[start]) {
+            continue;
+        }
+        let size: u64 = messages[start..turn.end]
+            .iter()
+            .map(|m| estimator.message(m))
+            .sum();
+        if size > budget {
+            continue;
+        }
+        return Some(start);
+    }
+    None
+}
+
+/// Select the verbatim tail: keep up to the last `tail_turns` whole
+/// turns that fit `budget` (newest first); when a whole turn does not
+/// fit, fall back to a safe partial cut inside it. Everything before
+/// the kept tail is the head to summarise.
+pub fn select(
+    messages: &[AgentMessage],
+    budget: u64,
+    tail_turns: usize,
+    estimator: &dyn Estimator,
+) -> Selection {
+    let whole_head = Selection {
+        head_len: messages.len(),
+        tail_start_index: None,
+    };
+    if tail_turns == 0 {
+        return whole_head;
+    }
+    let all = turns(messages);
+    if all.is_empty() {
+        return whole_head;
+    }
+    let recent = &all[all.len().saturating_sub(tail_turns)..];
+
+    let mut total: u64 = 0;
+    let mut keep: Option<usize> = None;
+    for turn in recent.iter().rev() {
+        let size: u64 = messages[turn.start..turn.end]
+            .iter()
+            .map(|m| estimator.message(m))
+            .sum();
+        if total + size <= budget {
+            total += size;
+            // A turn whose user message carries inline function_result
+            // blocks is not a safe boundary; accumulate it and let an
+            // older turn (or a split) cover it — the tail is a suffix,
+            // so the unsafe turn still lands tail-side with its call.
+            if is_safe_cut(&messages[turn.start]) {
+                keep = Some(turn.start);
+            }
+            continue;
+        }
+        let remaining = budget.saturating_sub(total);
+        if let Some(split) = split_turn(messages, *turn, remaining, estimator) {
+            keep = Some(split);
+        }
+        break;
+    }
+
+    match keep {
+        None | Some(0) => whole_head,
+        Some(start) => Selection {
+            head_len: start,
+            tail_start_index: Some(start),
+        },
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::core::estimate::HeuristicEstimator;
+    use serde_json::json;
+
+    fn user(text: &str, ts: i64) -> AgentMessage {
+        serde_json::from_value(json!({
+            "role": "user", "content": [{ "type": "text", "text": text }], "timestamp": ts
+        }))
+        .unwrap()
+    }
+
+    fn assistant(text: &str, ts: i64) -> AgentMessage {
+        serde_json::from_value(json!({
+            "role": "assistant", "content": [{ "type": "text", "text": text }],
+            "stop_reason": "end", "model": "m", "provider": "p", "timestamp": ts
+        }))
+        .unwrap()
+    }
+
+    fn assistant_call(call_id: &str, ts: i64) -> AgentMessage {
+        serde_json::from_value(json!({
+            "role": "assistant",
+            "content": [{ "type": "function_call", "id": call_id, "function_id": "f::g",
+                          "arguments": {} }],
+            "stop_reason": "function_call", "model": "m", "provider": "p", "timestamp": ts
+        }))
+        .unwrap()
+    }
+
+    fn result(call_id: &str, chars: usize, ts: i64) -> AgentMessage {
+        serde_json::from_value(json!({
+            "role": "function_result", "function_call_id": call_id, "function_id": "f::g",
+            "content": [{ "type": "text", "text": "x".repeat(chars) }], "timestamp": ts
+        }))
+        .unwrap()
+    }
+
+    #[test]
+    fn turns_start_at_user_messages() {
+        let messages = vec![
+            user("a", 1),
+            assistant("r", 2),
+            user("b", 3),
+            assistant("r", 4),
+        ];
+        let t = turns(&messages);
+        assert_eq!(
+            t,
+            vec![Turn { start: 0, end: 2 }, Turn { start: 2, end: 4 }]
+        );
+    }
+
+    #[test]
+    fn keeps_whole_recent_turns_within_budget() {
+        let messages = vec![
+            user("old question", 1),
+            assistant("old answer", 2),
+            user("recent question", 3),
+            assistant("recent answer", 4),
+        ];
+        let sel = select(&messages, 1_000, 1, &HeuristicEstimator);
+        assert_eq!(sel.tail_start_index, Some(2));
+        assert_eq!(sel.head_len, 2);
+    }
+
+    #[test]
+    fn zero_tail_turns_summarises_everything() {
+        let messages = vec![user("a", 1), assistant("b", 2)];
+        let sel = select(&messages, 1_000, 0, &HeuristicEstimator);
+        assert_eq!(sel.tail_start_index, None);
+        assert_eq!(sel.head_len, 2);
+    }
+
+    #[test]
+    fn tail_covering_everything_means_no_compaction_head() {
+        // Both turns fit: keep.start == 0 → whole history is "head",
+        // i.e. the caller sees nothing to summarise... actually the
+        // selection reports the whole history as head (no tail), which
+        // compact turns into tokens_before over everything.
+        let messages = vec![user("a", 1), assistant("b", 2)];
+        let sel = select(&messages, u64::MAX, 2, &HeuristicEstimator);
+        assert_eq!(sel.tail_start_index, None);
+        assert_eq!(sel.head_len, 2);
+    }
+
+    #[test]
+    fn split_inside_oversized_turn_never_cuts_at_a_function_result() {
+        // One huge turn: user, assistant(call), result(big), assistant.
+        // The only cut that fits the budget straddles the result — the
+        // split must land on the closing assistant message instead.
+        let messages = vec![
+            user("q", 1),
+            assistant_call("c1", 2),
+            result("c1", 4_000, 3),
+            assistant("done", 4),
+        ];
+        let sel = select(&messages, 100, 1, &HeuristicEstimator);
+        // Tail = just the final assistant message (index 3): the cut at
+        // index 2 (the function_result) is unsafe even though it fits.
+        assert_eq!(sel.tail_start_index, Some(3));
+    }
+
+    #[test]
+    fn user_message_with_inline_result_block_is_not_a_cut() {
+        let inline_result_user: AgentMessage = serde_json::from_value(json!({
+            "role": "user",
+            "content": [{ "type": "function_result", "function_call_id": "c9", "content": [] }],
+            "timestamp": 5
+        }))
+        .unwrap();
+        assert!(!is_safe_cut(&inline_result_user));
+        assert!(is_safe_cut(&user("plain", 1)));
+        assert!(is_safe_cut(&assistant("a", 2)));
+        assert!(!is_safe_cut(&result("c", 10, 3)));
+    }
+
+    #[test]
+    fn unsafe_turn_start_defers_to_an_older_safe_turn() {
+        // The newest turn starts at a user message carrying an inline
+        // function_result whose call sits in the previous turn. Cutting
+        // there would orphan the result, so the kept tail must start at
+        // the older (safe) turn and carry both turns together.
+        let inline_result_user: AgentMessage = serde_json::from_value(json!({
+            "role": "user",
+            "content": [{ "type": "function_result", "function_call_id": "c7", "content": [] }],
+            "timestamp": 5
+        }))
+        .unwrap();
+        let messages = vec![
+            user("old", 1),          // 0: turn 1
+            assistant("r0", 2),      // 1
+            user("follow", 3),       // 2: turn 2 (safe)
+            assistant_call("c7", 4), // 3
+            inline_result_user,      // 4: turn 3 (UNSAFE start)
+            assistant("done", 5),    // 5
+        ];
+        let sel = select(&messages, 1_000, 2, &HeuristicEstimator);
+        assert_eq!(
+            sel.tail_start_index,
+            Some(2),
+            "must cut at the safe turn 2 start"
+        );
+    }
+
+    #[test]
+    fn nothing_fits_summarises_everything() {
+        let messages = vec![
+            user("q1", 1),
+            assistant("a1", 2),
+            user("q2", 3),
+            assistant("a2", 4),
+        ];
+        let sel = select(&messages, 0, 2, &HeuristicEstimator);
+        assert_eq!(sel.tail_start_index, None);
+        assert_eq!(sel.head_len, 4);
+    }
+
+    #[test]
+    fn history_without_user_messages_is_all_head() {
+        let messages = vec![assistant("a", 1), assistant("b", 2)];
+        let sel = select(&messages, 1_000, 2, &HeuristicEstimator);
+        assert_eq!(sel.tail_start_index, None);
+    }
+}
diff --git a/context-manager/src/core/summary.rs b/context-manager/src/core/summary.rs
new file mode 100644
index 00000000..8e975fb1
--- /dev/null
+++ b/context-manager/src/core/summary.rs
@@ -0,0 +1,292 @@
+//! Summariser prompt construction (context-manager.md §
+//! context::compact): the fixed Markdown template, previous-summary
+//! anchoring (update-in-place so summaries converge instead of
+//! growing), media stripping, and the `# Conversation summary` system
+//! prompt section `assemble` renders.
+
+use crate::types::{AgentMessage, ContentBlock};
+
+/// The fixed summary structure (spec: Goal / Constraints / Progress /
+/// Key Decisions / Actions Taken / Next Steps / Critical Context /
+/// Relevant Files).
+pub const SUMMARY_TEMPLATE: &str = r#"Output exactly the Markdown structure shown inside <template> and keep the section order unchanged. Do not include the <template> tags in your response.
+<template>
+## Goal
+- [single-sentence task summary]
+
+## Constraints & Preferences
+- [user constraints, preferences, specs, or "(none)"]
+
+## Progress
+### Done
+- [completed work or "(none)"]
+
+### In Progress
+- [current work or "(none)"]
+
+### Blocked
+- [blockers or "(none)"]
+
+## Key Decisions
+- [decision and why, or "(none)"]
+
+## Actions Taken
+- [notable function calls with their arguments (terse), or "(none)"]
+
+## Next Steps
+- [ordered next actions or "(none)"]
+
+## Critical Context
+- [important technical facts, errors, open questions, or "(none)"]
+
+## Relevant Files
+- [file or directory path: why it matters, or "(none)"]
+</template>
+
+Rules:
+- Keep every section, even when empty.
+- Use terse bullets, not prose paragraphs.
+- Preserve exact file paths, commands, error strings, and identifiers when known.
+- Do not mention the summary process or that context was compacted."#;
+
+/// System prompt for the summariser turn. With a `previous_summary`
+/// the anchor instructs an update-in-place merge instead of a restart.
+pub fn build_system_prompt(previous_summary: Option<&str>) -> String {
+    let anchor = match previous_summary {
+        Some(prior) => format!(
+            "Update the anchored summary below using the conversation history above.\n\
+             Preserve still-true details, remove stale details, and merge in the new facts.\n\
+             <previous-summary>\n{prior}\n</previous-summary>"
+        ),
+        None => "Create a new anchored summary from the conversation history above.".to_string(),
+    };
+    format!("{anchor}\n\n{SUMMARY_TEMPLATE}")
+}
+
+/// User prompt: the head messages rendered as a `<conversation>` block
+/// (text blocks verbatim, function calls as terse one-liners).
+pub fn render_user_prompt(older: &[AgentMessage]) -> String {
+    let mut out = String::from(
+        "Summarise the conversation below following the system-prompt structure exactly. \
+         Keep identifiers verbatim.\n\n<conversation>\n",
+    );
+    for message in older {
+        out.push_str(&format!("\n[{}]\n", role_label(message)));
+        for block in message.content() {
+            match block {
+                ContentBlock::Text { text } => {
+                    out.push_str(text);
+                    out.push('\n');
+                }
+                ContentBlock::FunctionCall {
+                    function_id,
+                    arguments,
+                    ..
+                } => {
+                    out.push_str(&format!(
+                        "[tool_call] {function_id} {}\n",
+                        serde_json::to_string(arguments).unwrap_or_default()
+                    ));
+                }
+                _ => {}
+            }
+        }
+    }
+    out.push_str("</conversation>\n");
+    out
+}
+
+fn role_label(message: &AgentMessage) -> &'static str {
+    match message {
+        AgentMessage::User { .. } => "user",
+        AgentMessage::Assistant { .. } => "assistant",
+        AgentMessage::FunctionResult { .. } => "function_result",
+        AgentMessage::Custom { .. } => "custom",
+    }
+}
+
+/// Copy of `messages` fit for the summariser: images become
+/// `[image stripped]` placeholders everywhere; `function_result` text
+/// blocks are truncated to `max_output_chars`.
+pub fn strip_media(messages: &[AgentMessage], max_output_chars: usize) -> Vec<AgentMessage> {
+    messages
+        .iter()
+        .map(|message| {
+            let truncate = matches!(message, AgentMessage::FunctionResult { .. });
+            let mut stripped = message.clone();
+            let content = stripped
+                .content()
+                .iter()
+                .map(|block| {
+                    let block = strip_image(block);
+                    if truncate {
+                        truncate_text(block, max_output_chars)
+                    } else {
+                        block
+                    }
+                })
+                .collect();
+            stripped.set_content(content);
+            stripped
+        })
+        .collect()
+}
+
+fn strip_image(block: &ContentBlock) -> ContentBlock {
+    match block {
+        ContentBlock::Image { .. } => ContentBlock::Text {
+            text: "[image stripped]".to_string(),
+        },
+        other => other.clone(),
+    }
+}
+
+fn truncate_text(block: ContentBlock, max: usize) -> ContentBlock {
+    match block {
+        ContentBlock::Text { text } if text.len() > max => {
+            let mut cut = max;
+            while !text.is_char_boundary(cut) {
+                cut -= 1;
+            }
+            ContentBlock::Text {
+                text: format!("{}... [truncated]", &text[..cut]),
+            }
+        }
+        other => other,
+    }
+}
+
+/// Heading under which `assemble` renders the conversation summary into
+/// the returned system prompt (context-manager.md § The compaction
+/// round trip).
+pub const SUMMARY_HEADING: &str = "# Conversation summary";
+
+/// Base system prompt + the summary section. An empty base yields just
+/// the section; no summary yields just the base.
+pub fn render_system_prompt(base: Option<&str>, summary: Option<&str>) -> String {
+    let base = base.unwrap_or("").trim_end();
+    match summary {
+        None => base.to_string(),
+        Some(summary) => {
+            if base.is_empty() {
+                format!("{SUMMARY_HEADING}\n\n{summary}")
+            } else {
+                format!("{base}\n\n{SUMMARY_HEADING}\n\n{summary}")
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    fn message(v: serde_json::Value) -> AgentMessage {
+        serde_json::from_value(v).unwrap()
+    }
+
+    #[test]
+    fn fresh_prompt_instructs_creation_and_carries_every_section() {
+        let prompt = build_system_prompt(None);
+        assert!(prompt.starts_with("Create a new anchored summary"));
+        for section in [
+            "## Goal",
+            "## Constraints & Preferences",
+            "## Progress",
+            "## Key Decisions",
+            "## Actions Taken",
+            "## Next Steps",
+            "## Critical Context",
+            "## Relevant Files",
+        ] {
+            assert!(prompt.contains(section), "missing {section}");
+        }
+    }
+
+    #[test]
+    fn previous_summary_switches_to_update_mode() {
+        let prompt = build_system_prompt(Some("## Goal\n- ship it"));
+        assert!(prompt.starts_with("Update the anchored summary"));
+        assert!(prompt.contains("<previous-summary>\n## Goal\n- ship it\n</previous-summary>"));
+    }
+
+    #[test]
+    fn user_prompt_renders_text_and_tool_calls() {
+        let messages = vec![
+            message(json!({
+                "role": "user", "content": [{ "type": "text", "text": "find the bug" }],
+                "timestamp": 1
+            })),
+            message(json!({
+                "role": "assistant",
+                "content": [{ "type": "function_call", "id": "c1", "function_id": "coder::search",
+                              "arguments": { "q": "panic" } }],
+                "stop_reason": "function_call", "model": "m", "provider": "p", "timestamp": 2
+            })),
+        ];
+        let prompt = render_user_prompt(&messages);
+        assert!(prompt.contains("[user]\nfind the bug"));
+        assert!(prompt.contains("[tool_call] coder::search {\"q\":\"panic\"}"));
+        assert!(prompt.contains("<conversation>") && prompt.contains("</conversation>"));
+    }
+
+    #[test]
+    fn strip_media_replaces_images_and_truncates_outputs() {
+        let messages = vec![
+            message(json!({
+                "role": "user",
+                "content": [{ "type": "image", "mime": "image/png", "data": "AAAA" }],
+                "timestamp": 1
+            })),
+            message(json!({
+                "role": "function_result", "function_call_id": "c1", "function_id": "f",
+                "content": [{ "type": "text", "text": "x".repeat(50) }], "timestamp": 2
+            })),
+        ];
+        let stripped = strip_media(&messages, 10);
+        assert_eq!(
+            stripped[0].content(),
+            &[ContentBlock::Text {
+                text: "[image stripped]".into()
+            }]
+        );
+        let ContentBlock::Text { text } = &stripped[1].content()[0] else {
+            panic!("expected text");
+        };
+        assert_eq!(text, &format!("{}... [truncated]", "x".repeat(10)));
+        // Originals untouched (the head is summarised; the transcript
+        // the caller holds must never be mutated).
+        assert!(matches!(
+            messages[0].content()[0],
+            ContentBlock::Image { .. }
+        ));
+    }
+
+    #[test]
+    fn user_text_is_never_truncated_by_strip_media() {
+        let messages = vec![message(json!({
+            "role": "user", "content": [{ "type": "text", "text": "y".repeat(100) }],
+            "timestamp": 1
+        }))];
+        let stripped = strip_media(&messages, 10);
+        let ContentBlock::Text { text } = &stripped[0].content()[0] else {
+            panic!("expected text");
+        };
+        assert_eq!(text.len(), 100);
+    }
+
+    #[test]
+    fn render_system_prompt_combinations() {
+        assert_eq!(render_system_prompt(Some("Base."), None), "Base.");
+        assert_eq!(render_system_prompt(None, None), "");
+        assert_eq!(
+            render_system_prompt(Some("Base."), Some("S")),
+            "Base.\n\n# Conversation summary\n\nS"
+        );
+        assert_eq!(
+            render_system_prompt(None, Some("S")),
+            "# Conversation summary\n\nS"
+        );
+    }
+}
diff --git a/context-manager/src/error.rs b/context-manager/src/error.rs
new file mode 100644
index 00000000..728f2028
--- /dev/null
+++ b/context-manager/src/error.rs
@@ -0,0 +1,80 @@
+//! Error conventions: every error crossing the bus carries a stable
+//! snake_case, worker-prefixed code in a `code: message` shape (see
+//! tech-specs/2026-06-agentic/README.md § Error conventions).
+//!
+//! The two messages the spec names verbatim — `messages is required` and
+//! `could not resolve model limits` — are kept word-for-word in the
+//! message part so callers can match on the documented strings.
+
+use iii_sdk::IIIError;
+
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum ContextError {
+    /// The request shape is invalid beyond what serde can reject
+    /// (e.g. `messages` absent: the spec's `messages is required`).
+    #[error("context/invalid_request: {0}")]
+    InvalidRequest(String),
+
+    /// Neither inline limits nor `llm-router` could provide model limits
+    /// and the conservative fallback is disabled.
+    #[error("context/model_unresolved: could not resolve model limits ({0})")]
+    ModelUnresolved(String),
+
+    /// A filesystem operation backing the compaction lease failed.
+    #[error("context/state: {0}")]
+    State(String),
+}
+
+impl ContextError {
+    /// The stable error code (the part before `:` on the wire).
+    pub fn code(&self) -> &'static str {
+        match self {
+            ContextError::InvalidRequest(_) => "context/invalid_request",
+            ContextError::ModelUnresolved(_) => "context/model_unresolved",
+            ContextError::State(_) => "context/state",
+        }
+    }
+}
+
+impl From<ContextError> for IIIError {
+    fn from(e: ContextError) -> Self {
+        IIIError::Handler(e.to_string())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn display_is_code_colon_message() {
+        let e = ContextError::InvalidRequest("messages is required".into());
+        assert_eq!(
+            e.to_string(),
+            "context/invalid_request: messages is required"
+        );
+        assert_eq!(e.code(), "context/invalid_request");
+    }
+
+    #[test]
+    fn model_unresolved_keeps_the_spec_string() {
+        let e = ContextError::ModelUnresolved("no inline limits, router absent".into());
+        assert!(e.to_string().contains("could not resolve model limits"));
+    }
+
+    #[test]
+    fn every_variant_display_starts_with_its_code() {
+        let variants = vec![
+            ContextError::InvalidRequest("m".into()),
+            ContextError::ModelUnresolved("m".into()),
+            ContextError::State("m".into()),
+        ];
+        for v in variants {
+            assert!(
+                v.to_string().starts_with(&format!("{}: ", v.code())),
+                "display `{v}` must start with code `{}`",
+                v.code()
+            );
+        }
+    }
+}
diff --git a/context-manager/src/functions/assemble.rs b/context-manager/src/functions/assemble.rs
new file mode 100644
index 00000000..72c07466
--- /dev/null
+++ b/context-manager/src/functions/assemble.rs
@@ -0,0 +1,283 @@
+//! `context::assemble` — build the model-ready context from a history
+//! (context-manager.md § context::assemble). The pipeline, in order:
+//! count -> (if over) prune function outputs -> (if still over) compact
+//! the head -> assemble the final list.
+//!
+//! Structural guarantees: `role: "custom"` messages never reach the
+//! model-facing list (nor the count); `applied.tail_start_index`
+//! indexes the *request* messages array so callers can map it onto
+//! their own storage; a busy lease or failed summariser degrades to
+//! best effort instead of erroring the turn.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::core::budget::{default_reserved, preserve_recent_budget, usable};
+use crate::core::estimate::{estimator_for_model, Estimator};
+use crate::core::lease;
+use crate::core::prune::{prune as run_prune, PruneParams};
+use crate::core::selection::select;
+use crate::core::summary::{
+    build_system_prompt, render_system_prompt, render_user_prompt, strip_media,
+};
+use crate::error::ContextError;
+use crate::functions::resolve_model;
+use crate::ports::{Deps, SummarizeRequest};
+use crate::types::{AgentMessage, ModelInput, Role, ThinkingLevel};
+
+#[derive(Debug, Default, Deserialize, JsonSchema)]
+pub struct AssembleOptions {
+    /// Override the default reserve (`min(20000, 10% of context_window)`).
+    #[serde(default)]
+    pub reserved_tokens: Option<u64>,
+    /// user+assistant pairs always kept verbatim (default 2).
+    #[serde(default)]
+    pub tail_turns: Option<usize>,
+    /// Default true.
+    #[serde(default)]
+    pub allow_compaction: Option<bool>,
+    /// Default true.
+    #[serde(default)]
+    pub allow_prune: Option<bool>,
+    /// `function_id`s whose outputs are never pruned.
+    #[serde(default)]
+    pub protected_functions: Option<Vec<String>>,
+    /// Reserve the model's thinking budget for this tier.
+    #[serde(default)]
+    pub thinking_level: Option<ThinkingLevel>,
+    /// Compaction mutual-exclusion key (e.g. a session id); default:
+    /// hash of the message set.
+    #[serde(default)]
+    pub lease_key: Option<String>,
+    /// Persisted summary from a prior compaction (see the spec's "The
+    /// compaction round trip"); rendered into the system prompt and
+    /// used as the anchor if compaction triggers again.
+    #[serde(default)]
+    pub previous_summary: Option<String>,
+}
+
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct AssembleRequest {
+    /// Full candidate history, oldest first.
+    pub messages: Option<Vec<AgentMessage>>,
+    pub model: ModelInput,
+    /// Base system prompt to prepend/merge.
+    #[serde(default)]
+    pub system_prompt: Option<String>,
+    #[serde(default)]
+    pub options: Option<AssembleOptions>,
+}
+
+/// How the model limits were resolved.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ModelResolvedWire {
+    Inline,
+    Router,
+    Fallback,
+}
+
+/// What the pipeline actually did this call.
+#[derive(Debug, Serialize, JsonSchema)]
+pub struct Applied {
+    pub pruned: bool,
+    pub pruned_tokens: u64,
+    pub compacted: bool,
+    /// Present when compacted; the caller should persist it and pass
+    /// it back as `options.previous_summary` (compaction round trip).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub summary: Option<String>,
+    /// Present when compacted: index into the request messages where
+    /// the verbatim tail begins (`null` = everything was summarised).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tail_start_index: Option<Option<usize>>,
+    /// Present when compacted: estimated tokens of the summarised head.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tokens_before: Option<u64>,
+}
+
+#[derive(Debug, Serialize, JsonSchema)]
+pub struct AssembleResponse {
+    pub system_prompt: String,
+    /// Budgeted, ready to send to llm-router.
+    pub messages: Vec<AgentMessage>,
+    /// Estimated tokens of the returned context (messages + system
+    /// prompt). Can exceed `usable` when prune/compaction were
+    /// disabled, unavailable, or insufficient — best effort, visible.
+    pub token_count: u64,
+    /// The budget the context was fit into.
+    pub usable: u64,
+    pub model_resolved: ModelResolvedWire,
+    pub applied: Applied,
+}
+
+pub async fn handle(deps: &Deps, req: AssembleRequest) -> Result<AssembleResponse, ContextError> {
+    let request_messages = req
+        .messages
+        .ok_or_else(|| ContextError::InvalidRequest("messages is required".into()))?;
+    let options = req.options.unwrap_or_default();
+    let config = deps.config().await;
+
+    let resolved = resolve_model(deps, &req.model).await?;
+    let reserved = options
+        .reserved_tokens
+        .unwrap_or_else(|| default_reserved(&config, resolved.limits.context_window));
+    let thinking_budget = resolved.thinking_budget(options.thinking_level);
+    let usable_budget = usable(&resolved.limits, reserved, thinking_budget);
+
+    let estimator = estimator_for_model(&req.model.id);
+
+    // Model-facing view: custom messages have no provider wire mapping
+    // — they are excluded from the list and the count. `view_to_orig`
+    // maps view indices back to request indices for tail_start_index.
+    let mut working: Vec<AgentMessage> = Vec::new();
+    let mut view_to_orig: Vec<usize> = Vec::new();
+    for (idx, message) in request_messages.iter().enumerate() {
+        if message.role() != Role::Custom {
+            working.push(message.clone());
+            view_to_orig.push(idx);
+        }
+    }
+
+    let base_prompt = req.system_prompt.as_deref();
+    let mut system_prompt = render_system_prompt(base_prompt, options.previous_summary.as_deref());
+
+    let count = |messages: &[AgentMessage], prompt: &str| -> u64 {
+        messages.iter().map(|m| estimator.message(m)).sum::<u64>() + estimator.text(prompt)
+    };
+
+    let mut applied = Applied {
+        pruned: false,
+        pruned_tokens: 0,
+        compacted: false,
+        summary: None,
+        tail_start_index: None,
+        tokens_before: None,
+    };
+
+    let mut token_count = count(&working, &system_prompt);
+
+    // Step 1: prune function outputs.
+    if token_count > usable_budget && options.allow_prune.unwrap_or(true) {
+        let params = PruneParams {
+            protect_recent_tokens: config.protect_recent_tokens,
+            min_free_tokens: config.min_free_tokens,
+            max_output_chars: config.max_output_chars,
+            protected_functions: options.protected_functions.clone().unwrap_or_default(),
+        };
+        let stats = run_prune(&mut working, &params, estimator);
+        applied.pruned = stats.pruned_parts > 0;
+        applied.pruned_tokens = stats.pruned_tokens;
+        token_count = count(&working, &system_prompt);
+    }
+
+    // Step 2: compact the head.
+    if token_count > usable_budget && options.allow_compaction.unwrap_or(true) {
+        // The default lease key hashes the *request* message set —
+        // the same derivation context::compact uses — so callers
+        // hitting both functions with the same history contend on the
+        // same claim.
+        let lease_key = options
+            .lease_key
+            .clone()
+            .unwrap_or_else(|| lease::default_lease_key(&request_messages));
+        if let Some(compaction) = try_compact(
+            deps,
+            &req.model,
+            &working,
+            usable_budget,
+            options.tail_turns.unwrap_or(config.tail_turns),
+            &lease_key,
+            options.previous_summary.as_deref(),
+            estimator,
+        )
+        .await
+        {
+            system_prompt = render_system_prompt(base_prompt, Some(&compaction.summary));
+            working = working.split_off(compaction.head_len);
+            applied.compacted = true;
+            applied.tail_start_index = Some(compaction.tail_start_view.map(|v| view_to_orig[v]));
+            applied.tokens_before = Some(compaction.tokens_before);
+            applied.summary = Some(compaction.summary);
+            token_count = count(&working, &system_prompt);
+        }
+    }
+
+    Ok(AssembleResponse {
+        system_prompt,
+        messages: working,
+        token_count,
+        usable: usable_budget,
+        model_resolved: match resolved.resolved {
+            crate::core::budget::ModelResolved::Inline => ModelResolvedWire::Inline,
+            crate::core::budget::ModelResolved::Router => ModelResolvedWire::Router,
+            crate::core::budget::ModelResolved::Fallback => ModelResolvedWire::Fallback,
+        },
+        applied,
+    })
+}
+
+struct CompactionOutcome {
+    summary: String,
+    head_len: usize,
+    tail_start_view: Option<usize>,
+    tokens_before: u64,
+}
+
+/// Inline compaction under the lease. `None` means "no compaction this
+/// call" — lease busy, nothing to summarise, or summariser failure —
+/// and assemble degrades to best effort.
+#[allow(clippy::too_many_arguments)]
+async fn try_compact(
+    deps: &Deps,
+    model: &ModelInput,
+    working: &[AgentMessage],
+    usable_budget: u64,
+    tail_turns: usize,
+    lease_key: &str,
+    previous_summary: Option<&str>,
+    estimator: &dyn Estimator,
+) -> Option<CompactionOutcome> {
+    let config = deps.config().await;
+    let ttl_ms = (config.lease_ttl_secs * 1_000) as i64;
+    let nonce =
+        lease::acquire(deps.leases.as_ref(), deps.clock.as_ref(), lease_key, ttl_ms).await?;
+
+    let outcome = async {
+        let budget = preserve_recent_budget(usable_budget, None);
+        let selection = select(working, budget, tail_turns, estimator);
+        let head = &working[..selection.head_len];
+        if head.is_empty() {
+            return None;
+        }
+
+        let tokens_before: u64 = head.iter().map(|m| estimator.message(m)).sum();
+        let stripped = strip_media(head, config.max_output_chars);
+        let request = SummarizeRequest {
+            system_prompt: build_system_prompt(previous_summary),
+            user_prompt: render_user_prompt(&stripped),
+            model: model.id.clone(),
+            provider: model.provider.clone(),
+        };
+
+        match deps.summarizer.summarize(request).await {
+            Ok(summary) => Some(CompactionOutcome {
+                summary,
+                head_len: selection.head_len,
+                tail_start_view: selection.tail_start_index,
+                tokens_before,
+            }),
+            Err(err) => {
+                // Assemble never fails the turn on a summariser error;
+                // the caller still gets a (possibly over-budget)
+                // context and can see token_count > usable.
+                tracing::warn!(error = %err, "assemble: compaction skipped");
+                None
+            }
+        }
+    }
+    .await;
+
+    lease::release(deps.leases.as_ref(), lease_key, &nonce).await;
+    outcome
+}
diff --git a/context-manager/src/functions/compact.rs b/context-manager/src/functions/compact.rs
new file mode 100644
index 00000000..eb865722
--- /dev/null
+++ b/context-manager/src/functions/compact.rs
@@ -0,0 +1,183 @@
+//! `context::compact` — summarise the head of a history into a single
+//! compaction summary, keeping a recent tail verbatim
+//! (context-manager.md § context::compact).
+//!
+//! Transient and storage-agnostic: the summary is returned for the
+//! caller to persist; no session is touched. A short-lived lease
+//! (scope `context_lease`) keeps two callers from summarising the same
+//! logical history concurrently. Without `llm-router` the summariser
+//! is unavailable and the response is `{ status: "overflow" }` —
+//! callers treat it as "compaction unavailable".
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::core::budget::{default_reserved, preserve_recent_budget, usable};
+use crate::core::estimate::{estimator_for_model, Estimator};
+use crate::core::lease;
+use crate::core::selection::select;
+use crate::core::summary::{build_system_prompt, render_user_prompt, strip_media};
+use crate::error::ContextError;
+use crate::functions::resolve_model;
+use crate::ports::{Deps, SummarizeError, SummarizeRequest};
+use crate::types::{AgentMessage, ModelInput};
+
+#[derive(Debug, Default, Deserialize, JsonSchema)]
+pub struct CompactOptions {
+    /// user+assistant pairs kept verbatim (default 2).
+    #[serde(default)]
+    pub tail_turns: Option<usize>,
+    /// Anchor from a prior compaction so summaries converge instead of
+    /// growing; the summariser updates it rather than starting over.
+    #[serde(default)]
+    pub previous_summary: Option<String>,
+    /// Override the adaptive verbatim-tail token budget.
+    #[serde(default)]
+    pub preserve_recent_tokens: Option<u64>,
+    /// Mutual-exclusion key (e.g. a session id); default: hash of the
+    /// message set.
+    #[serde(default)]
+    pub lease_key: Option<String>,
+}
+
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct CompactRequest {
+    /// Full candidate history, oldest first.
+    pub messages: Option<Vec<AgentMessage>>,
+    pub model: ModelInput,
+    #[serde(default)]
+    pub options: Option<CompactOptions>,
+}
+
+/// Discriminated on `status`.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(tag = "status", rename_all = "snake_case")]
+pub enum CompactResponse {
+    /// Compaction ran; the caller should persist `summary` and map
+    /// `tail_start_index` onto its own storage ids.
+    Ok {
+        summary: String,
+        /// Index into the request `messages` where the verbatim tail
+        /// begins; `null` when everything was summarised.
+        tail_start_index: Option<usize>,
+        /// Estimated tokens of the summarised head.
+        tokens_before: u64,
+        /// Estimated tokens of summary + verbatim tail.
+        tokens_after: u64,
+        used_prior_summary: bool,
+    },
+    /// A compaction lease is held; the caller may retry.
+    Busy,
+    /// Nothing to compact.
+    Empty,
+    /// The summariser is unavailable or itself overflowed.
+    Overflow,
+}
+
+pub async fn handle(deps: &Deps, req: CompactRequest) -> Result<CompactResponse, ContextError> {
+    let messages = req
+        .messages
+        .ok_or_else(|| ContextError::InvalidRequest("messages is required".into()))?;
+    if messages.is_empty() {
+        return Ok(CompactResponse::Empty);
+    }
+
+    let config = deps.config().await;
+    let options = req.options.unwrap_or_default();
+    let resolved = resolve_model(deps, &req.model).await?;
+    let estimator = estimator_for_model(&req.model.id);
+
+    let reserved = default_reserved(&config, resolved.limits.context_window);
+    let budget = preserve_recent_budget(
+        usable(&resolved.limits, reserved, 0),
+        options.preserve_recent_tokens,
+    );
+    let tail_turns = options.tail_turns.unwrap_or(config.tail_turns);
+
+    let lease_key = options
+        .lease_key
+        .clone()
+        .unwrap_or_else(|| lease::default_lease_key(&messages));
+    let ttl_ms = (config.lease_ttl_secs * 1_000) as i64;
+    let Some(nonce) = lease::acquire(
+        deps.leases.as_ref(),
+        deps.clock.as_ref(),
+        &lease_key,
+        ttl_ms,
+    )
+    .await
+    else {
+        return Ok(CompactResponse::Busy);
+    };
+
+    let outcome = summarise(
+        deps,
+        &req.model,
+        &messages,
+        budget,
+        tail_turns,
+        options.previous_summary.as_deref(),
+        estimator,
+    )
+    .await;
+
+    lease::release(deps.leases.as_ref(), &lease_key, &nonce).await;
+    Ok(outcome)
+}
+
+/// The summarisation pipeline between lease acquire and release.
+async fn summarise(
+    deps: &Deps,
+    model: &ModelInput,
+    messages: &[AgentMessage],
+    budget: u64,
+    tail_turns: usize,
+    previous_summary: Option<&str>,
+    estimator: &dyn Estimator,
+) -> CompactResponse {
+    let selection = select(messages, budget, tail_turns, estimator);
+    let head = &messages[..selection.head_len];
+    if head.is_empty() {
+        return CompactResponse::Empty;
+    }
+    let tail = &messages[selection.head_len..];
+
+    let tokens_before: u64 = head.iter().map(|m| estimator.message(m)).sum();
+    let stripped = strip_media(head, deps.config().await.max_output_chars);
+
+    let request = SummarizeRequest {
+        system_prompt: build_system_prompt(previous_summary),
+        user_prompt: render_user_prompt(&stripped),
+        model: model.id.clone(),
+        provider: model.provider.clone(),
+    };
+
+    let summary = match deps.summarizer.summarize(request).await {
+        Ok(summary) => summary,
+        Err(SummarizeError::Empty) => {
+            tracing::warn!("summariser produced an empty summary; nothing to compact");
+            return CompactResponse::Empty;
+        }
+        Err(err @ SummarizeError::Unavailable(_)) => {
+            // Spec: without llm-router, compact returns overflow with a
+            // permanent error_kind in the worker log.
+            tracing::error!(error = %err, error_kind = "permanent", "compaction unavailable");
+            return CompactResponse::Overflow;
+        }
+        Err(err) => {
+            tracing::error!(error = %err, "summariser failed");
+            return CompactResponse::Overflow;
+        }
+    };
+
+    let tokens_after =
+        estimator.text(&summary) + tail.iter().map(|m| estimator.message(m)).sum::<u64>();
+
+    CompactResponse::Ok {
+        summary,
+        tail_start_index: selection.tail_start_index,
+        tokens_before,
+        tokens_after,
+        used_prior_summary: previous_summary.is_some(),
+    }
+}
diff --git a/context-manager/src/functions/count_tokens.rs b/context-manager/src/functions/count_tokens.rs
new file mode 100644
index 00000000..7da3b9fa
--- /dev/null
+++ b/context-manager/src/functions/count_tokens.rs
@@ -0,0 +1,92 @@
+//! `context::count-tokens` — estimate token usage for a set of
+//! messages, optionally including invocation schemas and a system
+//! prompt, vs a model (context-manager.md § context::count-tokens).
+//!
+//! Pure and router-free: the model selects a tokenizer (v1 always
+//! falls back to the generic heuristic, reported in `estimator`), so
+//! cost-sensitive callers can run this with no `llm-router` installed.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::core::estimate::{estimate_by_role, estimate_messages, estimator_for_model};
+use crate::error::ContextError;
+use crate::ports::Deps;
+use crate::types::{AgentFunction, AgentMessage, ModelInput};
+
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct CountTokensRequest {
+    /// Messages to estimate, oldest first.
+    pub messages: Option<Vec<AgentMessage>>,
+    /// Counted on top of the messages when present.
+    #[serde(default)]
+    pub system_prompt: Option<String>,
+    /// Invocation schema(s) to include in the estimate (typically the
+    /// single `agent_trigger` entry).
+    #[serde(default)]
+    pub tools: Option<Vec<AgentFunction>>,
+    /// Tokenizer selection; falls back to a generic estimator.
+    pub model: ModelInput,
+}
+
+/// Per-role token breakdown of the `messages` array.
+#[derive(Debug, Serialize, JsonSchema)]
+pub struct ByRoleTokens {
+    pub user: u64,
+    pub assistant: u64,
+    pub function_result: u64,
+    pub custom: u64,
+}
+
+/// Which estimator produced the count.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum EstimatorName {
+    Tokenizer,
+    Heuristic,
+}
+
+#[derive(Debug, Serialize, JsonSchema)]
+pub struct CountTokensResponse {
+    /// Total estimate: messages + system prompt + tools.
+    pub tokens: u64,
+    /// Breakdown of the message tokens by role (system prompt and
+    /// tools are not part of any role bucket).
+    pub by_role: Option<ByRoleTokens>,
+    pub estimator: EstimatorName,
+}
+
+pub async fn handle(
+    _deps: &Deps,
+    req: CountTokensRequest,
+) -> Result<CountTokensResponse, ContextError> {
+    let messages = req
+        .messages
+        .ok_or_else(|| ContextError::InvalidRequest("messages is required".into()))?;
+
+    let estimator = estimator_for_model(&req.model.id);
+
+    let mut tokens = estimate_messages(estimator, &messages);
+    if let Some(system_prompt) = &req.system_prompt {
+        tokens += estimator.text(system_prompt);
+    }
+    for tool in req.tools.iter().flatten() {
+        tokens += estimator.function(tool);
+    }
+
+    let by_role = estimate_by_role(estimator, &messages);
+
+    Ok(CountTokensResponse {
+        tokens,
+        by_role: Some(ByRoleTokens {
+            user: by_role.user,
+            assistant: by_role.assistant,
+            function_result: by_role.function_result,
+            custom: by_role.custom,
+        }),
+        estimator: match estimator.kind() {
+            crate::core::estimate::EstimatorKind::Tokenizer => EstimatorName::Tokenizer,
+            crate::core::estimate::EstimatorKind::Heuristic => EstimatorName::Heuristic,
+        },
+    })
+}
diff --git a/context-manager/src/functions/mod.rs b/context-manager/src/functions/mod.rs
new file mode 100644
index 00000000..b07a594a
--- /dev/null
+++ b/context-manager/src/functions/mod.rs
@@ -0,0 +1,165 @@
+//! The four `context::*` functions.
+//!
+//! Each `<verb>.rs` holds the request/response types (serde +
+//! `schemars::JsonSchema`, so the SDK emits request/response schemas)
+//! and a `pub async fn handle(deps, req)` that the registration closure
+//! wraps. BDD scenarios call the same `handle` functions directly, so
+//! engine-free tests exercise the exact production code path.
+
+pub mod assemble;
+pub mod compact;
+pub mod count_tokens;
+pub mod prune;
+
+use std::future::Future;
+use std::sync::Arc;
+
+use iii_sdk::{IIIError, RegisterFunction, III};
+use schemars::JsonSchema;
+use serde::de::DeserializeOwned;
+use serde::Serialize;
+
+use crate::core::budget::{fallback_model, ResolvedModel};
+use crate::error::ContextError;
+use crate::ports::Deps;
+use crate::types::ModelInput;
+
+pub const ASSEMBLE_ID: &str = "context::assemble";
+pub const ASSEMBLE_DESC: &str =
+    "Build the model-ready context (system prompt + budgeted messages) from a history; \
+     applies prune and/or compaction as needed to fit the model's usable token budget.";
+
+pub const COMPACT_ID: &str = "context::compact";
+pub const COMPACT_DESC: &str =
+    "Summarise the head of a history into a single compaction summary, keeping a recent \
+     tail verbatim. Transient: the caller persists the result.";
+
+pub const PRUNE_ID: &str = "context::prune";
+pub const PRUNE_DESC: &str =
+    "Replace verbose function outputs with placeholders without summarising. \
+     A cheaper first pass before compaction.";
+
+pub const COUNT_TOKENS_ID: &str = "context::count-tokens";
+pub const COUNT_TOKENS_DESC: &str =
+    "Estimate token usage for a set of messages (+ optional invocation schemas / system \
+     prompt) vs a model.";
+
+/// Resolve model limits per the spec order: inline `limits` ->
+/// `router::models::get` -> conservative fallback (when allowed).
+pub(crate) async fn resolve_model(
+    deps: &Deps,
+    input: &ModelInput,
+) -> Result<ResolvedModel, ContextError> {
+    if let Some(limits) = input.limits {
+        return Ok(ResolvedModel::from_inline(limits));
+    }
+    let reason = match deps
+        .resolver
+        .get_model(input.provider.as_deref(), &input.id)
+        .await
+    {
+        Ok(Some(model)) => return Ok(ResolvedModel::from_router(&model)),
+        Ok(None) => format!("model {} is not in the router catalog", input.id),
+        Err(e) => format!("router unavailable: {e}"),
+    };
+    if deps.config().await.allow_fallback_limits {
+        tracing::warn!(model = %input.id, %reason, "using conservative fallback limits");
+        return Ok(fallback_model());
+    }
+    Err(ContextError::ModelUnresolved(reason))
+}
+
+/// Register one typed handler under `id`, mapping `ContextError` into
+/// the bus error shape (`code: message`).
+fn register<Req, Resp, F, Fut>(
+    iii: &Arc<III>,
+    deps: &Arc<Deps>,
+    id: &str,
+    description: &str,
+    handler: F,
+) where
+    Req: DeserializeOwned + JsonSchema + Send + 'static,
+    Resp: Serialize + JsonSchema + Send + 'static,
+    F: Fn(Arc<Deps>, Req) -> Fut + Send + Sync + Clone + 'static,
+    Fut: Future<Output = Result<Resp, ContextError>> + Send + 'static,
+{
+    let deps = deps.clone();
+    iii.register_function(
+        id,
+        RegisterFunction::new_async(move |req: Req| {
+            let deps = deps.clone();
+            let handler = handler.clone();
+            async move { handler(deps, req).await.map_err(IIIError::from) }
+        })
+        .description(description),
+    );
+}
+
+pub fn register_all(iii: &Arc<III>, deps: &Arc<Deps>) {
+    register(iii, deps, ASSEMBLE_ID, ASSEMBLE_DESC, |d, r| async move {
+        assemble::handle(&d, r).await
+    });
+    register(iii, deps, COMPACT_ID, COMPACT_DESC, |d, r| async move {
+        compact::handle(&d, r).await
+    });
+    register(iii, deps, PRUNE_ID, PRUNE_DESC, |d, r| async move {
+        prune::handle(&d, r).await
+    });
+    register(
+        iii,
+        deps,
+        COUNT_TOKENS_ID,
+        COUNT_TOKENS_DESC,
+        |d, r| async move { count_tokens::handle(&d, r).await },
+    );
+
+    tracing::info!("all context::* functions registered");
+}
+
+/// One function's complete agent-facing wire surface: id, registration
+/// description, and the schemars-derived request/response schemas.
+pub struct FunctionSpec {
+    pub function_id: &'static str,
+    pub description: &'static str,
+    pub request_schema: schemars::schema::RootSchema,
+    pub response_schema: schemars::schema::RootSchema,
+}
+
+/// Schema generation MUST mirror iii-sdk's internal `json_schema_for`
+/// (`SchemaSettings::draft07()` on the handler's request/response
+/// types): `RegisterFunction::new_async` auto-extracts schemas from the
+/// SAME structs referenced here, with the same schemars 0.8 generator
+/// settings, so a catalog snapshot pins exactly what registration
+/// emits.
+fn schema_of<T: schemars::JsonSchema>() -> schemars::schema::RootSchema {
+    schemars::r#gen::SchemaSettings::draft07()
+        .into_generator()
+        .into_root_schema_for::<T>()
+}
+
+fn spec<Req, Resp>(function_id: &'static str, description: &'static str) -> FunctionSpec
+where
+    Req: schemars::JsonSchema,
+    Resp: schemars::JsonSchema,
+{
+    FunctionSpec {
+        function_id,
+        description,
+        request_schema: schema_of::<Req>(),
+        response_schema: schema_of::<Resp>(),
+    }
+}
+
+/// The full wire-surface catalog, in registration order. Golden-tested
+/// in `tests/schemas.rs`; keep in lockstep with `register_all`.
+pub fn catalog() -> Vec<FunctionSpec> {
+    vec![
+        spec::<assemble::AssembleRequest, assemble::AssembleResponse>(ASSEMBLE_ID, ASSEMBLE_DESC),
+        spec::<compact::CompactRequest, compact::CompactResponse>(COMPACT_ID, COMPACT_DESC),
+        spec::<prune::PruneRequest, prune::PruneResponse>(PRUNE_ID, PRUNE_DESC),
+        spec::<count_tokens::CountTokensRequest, count_tokens::CountTokensResponse>(
+            COUNT_TOKENS_ID,
+            COUNT_TOKENS_DESC,
+        ),
+    ]
+}
diff --git a/context-manager/src/functions/prune.rs b/context-manager/src/functions/prune.rs
new file mode 100644
index 00000000..627f08e1
--- /dev/null
+++ b/context-manager/src/functions/prune.rs
@@ -0,0 +1,84 @@
+//! `context::prune` — replace verbose function outputs with
+//! placeholders without summarising (context-manager.md §
+//! context::prune). The cheaper first pass before compaction; safe to
+//! expose to agents (no LLM call, no state).
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::core::estimate::estimator_for_model;
+use crate::core::prune::{prune as run_prune, PruneParams};
+use crate::error::ContextError;
+use crate::ports::Deps;
+use crate::types::{AgentMessage, ModelInput};
+
+#[derive(Debug, Default, Deserialize, JsonSchema)]
+pub struct PruneOptions {
+    /// Newest function-output tokens never pruned (default 40000).
+    #[serde(default)]
+    pub protect_recent_tokens: Option<u64>,
+    /// Skip the pass entirely when it would free less (default 20000).
+    #[serde(default)]
+    pub min_free_tokens: Option<u64>,
+    /// Per-output verbosity threshold in chars (default 2000): outputs
+    /// at or under it are never pruned.
+    #[serde(default)]
+    pub max_output_chars: Option<usize>,
+    /// `function_id`s whose outputs are never pruned.
+    #[serde(default)]
+    pub protected_functions: Option<Vec<String>>,
+}
+
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct PruneRequest {
+    /// Messages to prune, oldest first.
+    pub messages: Option<Vec<AgentMessage>>,
+    /// Only used for token math; the heuristic estimator needs no
+    /// model, so this is optional.
+    #[serde(default)]
+    pub model: Option<ModelInput>,
+    #[serde(default)]
+    pub options: Option<PruneOptions>,
+}
+
+#[derive(Debug, Serialize, JsonSchema)]
+pub struct PruneResponse {
+    /// Same array with pruned outputs replaced by placeholders — no
+    /// message or block is ever removed.
+    pub messages: Vec<AgentMessage>,
+    /// Estimated tokens freed.
+    pub pruned_tokens: u64,
+    /// Outputs replaced with placeholders.
+    pub pruned_parts: u64,
+    /// Prunable outputs examined.
+    pub scanned_parts: u64,
+}
+
+pub async fn handle(deps: &Deps, req: PruneRequest) -> Result<PruneResponse, ContextError> {
+    let mut messages = req
+        .messages
+        .ok_or_else(|| ContextError::InvalidRequest("messages is required".into()))?;
+
+    let config = deps.config().await;
+    let options = req.options.unwrap_or_default();
+    let params = PruneParams {
+        protect_recent_tokens: options
+            .protect_recent_tokens
+            .unwrap_or(config.protect_recent_tokens),
+        min_free_tokens: options.min_free_tokens.unwrap_or(config.min_free_tokens),
+        max_output_chars: options.max_output_chars.unwrap_or(config.max_output_chars),
+        protected_functions: options.protected_functions.unwrap_or_default(),
+    };
+
+    let model_id = req.model.as_ref().map(|m| m.id.as_str()).unwrap_or("");
+    let estimator = estimator_for_model(model_id);
+
+    let stats = run_prune(&mut messages, &params, estimator);
+
+    Ok(PruneResponse {
+        messages,
+        pruned_tokens: stats.pruned_tokens,
+        pruned_parts: stats.pruned_parts,
+        scanned_parts: stats.scanned_parts,
+    })
+}
diff --git a/context-manager/src/lib.rs b/context-manager/src/lib.rs
new file mode 100644
index 00000000..3ed4e106
--- /dev/null
+++ b/context-manager/src/lib.rs
@@ -0,0 +1,17 @@
+//! `context-manager` — model-ready context assembly for the agentic
+//! worker family (tech-specs/2026-06-agentic/context-manager.md).
+//!
+//! Pure transforms over caller-supplied `AgentMessage[]`: token
+//! counting, function-result pruning, and history compaction, behind
+//! four `context::*` functions. Stateless with respect to conversation
+//! storage; the only state is short-lived compaction leases.
+
+pub mod adapters;
+pub mod config;
+pub mod configuration;
+pub mod core;
+pub mod error;
+pub mod functions;
+pub mod manifest;
+pub mod ports;
+pub mod types;
diff --git a/context-manager/src/main.rs b/context-manager/src/main.rs
new file mode 100644
index 00000000..d39fd4bd
--- /dev/null
+++ b/context-manager/src/main.rs
@@ -0,0 +1,161 @@
+//! `context-manager` binary entry.
+//!
+//! Boot sequence:
+//!   1. Parse CLI. An optional `--config` YAML file is only a SEED for the
+//!      first registration; the authoritative config lives in the
+//!      `configuration` worker.
+//!   2. Connect to the local iii engine over WebSocket.
+//!   3. Register the config schema (+ seed) with the `configuration`
+//!      worker and fetch the authoritative, env-expanded value.
+//!      `configuration` is a required boot dependency: a failed
+//!      register/fetch aborts startup.
+//!   4. Wire production adapters behind the ports: model limits via
+//!      `router::models::get` and the summariser via `router::chat`
+//!      (both soft — only compaction degrades without `llm-router`), plus
+//!      compaction leases on the local filesystem (`lease_dir`).
+//!      `lease_dir` and `summarizer_timeout_ms` are read once here and are
+//!      restart-required.
+//!   5. Register the four `context::*` functions, then bind a
+//!      `configuration` trigger that hot-reloads the per-call tuning knobs.
+//!   6. Sleep on Ctrl+C, then `shutdown_async` cleanly.
+
+use std::sync::Arc;
+
+use anyhow::{Context, Result};
+use clap::Parser;
+use iii_sdk::{register_worker, InitOptions, WorkerMetadata};
+use tokio::sync::RwLock;
+
+use context_manager::adapters::fs_lease::FsLeaseStore;
+use context_manager::adapters::router::{RouterModelResolver, RouterSummarizer};
+use context_manager::configuration::{self, ConfigCell};
+use context_manager::ports::{Deps, SystemClock};
+use context_manager::{config, functions, manifest};
+
+#[derive(Parser, Debug)]
+#[command(
+    name = "context-manager",
+    about = "Model-ready context assembly: token counting, pruning, and history compaction."
+)]
+struct Cli {
+    /// Optional seed config.yaml used to populate `initial_value` on the
+    /// first registration. The AUTHORITATIVE config is always fetched from
+    /// the `configuration` worker afterward; this file only seeds it.
+    #[arg(long)]
+    config: Option<String>,
+
+    #[arg(long, default_value = "ws://127.0.0.1:49134")]
+    url: String,
+
+    #[arg(long)]
+    manifest: bool,
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
+        )
+        .init();
+
+    let cli = Cli::parse();
+
+    if cli.manifest {
+        let m = manifest::build_manifest();
+        println!("{}", serde_json::to_string_pretty(&m).unwrap());
+        return Ok(());
+    }
+
+    // Connect first so the configuration round-trip below runs over a live
+    // connection.
+    let iii = Arc::new(register_worker(
+        &cli.url,
+        InitOptions {
+            metadata: Some(WorkerMetadata {
+                runtime: "rust".to_string(),
+                version: env!("CARGO_PKG_VERSION").to_string(),
+                name: "context-manager".to_string(),
+                os: std::env::consts::OS.to_string(),
+                pid: Some(std::process::id()),
+                telemetry: None,
+                ..WorkerMetadata::default()
+            }),
+            ..InitOptions::default()
+        },
+    ));
+
+    // A `--config` file only SEEDS the first registration; a failed parse
+    // WARNS and falls through to None (the authoritative value comes from the
+    // configuration worker). The seed IS env-expanded (`${VAR}`).
+    let seed = match cli.config.as_deref() {
+        Some(path) => match config::WorkerConfig::from_file(path) {
+            Ok(c) => {
+                tracing::info!(path = %path, "loaded seed config for initial registration");
+                Some(c)
+            }
+            Err(e) => {
+                tracing::warn!(
+                    path = %path,
+                    error = %e,
+                    "failed to load seed config; relying on the stored configuration entry"
+                );
+                None
+            }
+        },
+        None => None,
+    };
+
+    // Register the schema (+ optional seed) and fetch the authoritative,
+    // env-expanded config. `configuration` is a required boot dependency.
+    configuration::register_config(&iii, seed.as_ref())
+        .await
+        .map_err(anyhow::Error::msg)
+        .context("registering context-manager configuration schema")?;
+    let cfg = configuration::fetch_config(&iii)
+        .await
+        .map_err(anyhow::Error::msg)
+        .context("loading context-manager configuration")?;
+
+    // Capture the restart-required signature before moving cfg into the cell.
+    let boot_sig = cfg.boot_signature();
+
+    // Boot-time adapters read the restart-required fields ONCE here:
+    // summarizer_timeout_ms -> RouterSummarizer, lease_dir -> FsLeaseStore. A
+    // later change to either is refused on hot-reload. A lease dir we can't
+    // create is boot-fatal — leasing must never silently no-op.
+    let summarizer = Arc::new(RouterSummarizer::new(
+        iii.clone(),
+        cfg.summarizer_timeout_ms,
+    ));
+    let leases = Arc::new(
+        FsLeaseStore::new(cfg.resolved_lease_dir())
+            .map_err(|e| anyhow::anyhow!("opening the compaction lease directory: {e}"))?,
+    );
+
+    // The hot-swappable snapshot shared with every handler.
+    let cell: ConfigCell = Arc::new(RwLock::new(Arc::new(cfg)));
+
+    let deps = Arc::new(Deps {
+        config: cell.clone(),
+        resolver: Arc::new(RouterModelResolver::new(iii.clone())),
+        summarizer,
+        leases,
+        clock: Arc::new(SystemClock),
+    });
+
+    functions::register_all(&iii, &deps);
+
+    // LAST: bind the configuration-change trigger so its handler closes over
+    // the fully-built snapshot cell + boot signature.
+    configuration::register_config_trigger(&iii, cell, boot_sig)
+        .context("registering the configuration change trigger")?;
+
+    tracing::info!("context-manager ready: 4 context::* functions");
+
+    tokio::signal::ctrl_c().await?;
+    tracing::info!("context-manager shutting down");
+    iii.shutdown_async().await;
+    Ok(())
+}
diff --git a/context-manager/src/manifest.rs b/context-manager/src/manifest.rs
new file mode 100644
index 00000000..8371e9cc
--- /dev/null
+++ b/context-manager/src/manifest.rs
@@ -0,0 +1,65 @@
+//! `--manifest` subcommand output for the registry publish pipeline.
+
+use serde::Serialize;
+
+#[derive(Serialize)]
+pub struct ModuleManifest {
+    pub name: String,
+    pub version: String,
+    pub description: String,
+    pub default_config: serde_json::Value,
+    pub supported_targets: Vec<String>,
+}
+
+pub fn build_manifest() -> ModuleManifest {
+    ModuleManifest {
+        name: env!("CARGO_PKG_NAME").to_string(),
+        version: env!("CARGO_PKG_VERSION").to_string(),
+        description:
+            "Model-ready context assembly: token counting, function-result pruning, and history compaction over caller-supplied messages."
+                .to_string(),
+        // Mirrors config::WorkerConfig::default() field-for-field.
+        default_config: serde_json::json!({
+            "reserved_tokens_cap": 20_000,
+            "reserved_pct": 10,
+            "tail_turns": 2,
+            "protect_recent_tokens": 40_000,
+            "min_free_tokens": 20_000,
+            "max_output_chars": 2_000,
+            "lease_ttl_secs": 300,
+            "allow_fallback_limits": true,
+            "summarizer_timeout_ms": 320_000,
+            "lease_dir": "~/.iii/data/context-manager",
+        }),
+        supported_targets: vec![env!("TARGET").to_string()],
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::config::WorkerConfig;
+
+    #[test]
+    fn json_roundtrip_has_required_fields() {
+        let m = build_manifest();
+        let json = serde_json::to_string_pretty(&m).unwrap();
+        let parsed: serde_json::Value = serde_json::from_str(&json).unwrap();
+        assert_eq!(parsed["name"], env!("CARGO_PKG_NAME"));
+        assert_eq!(parsed["version"], env!("CARGO_PKG_VERSION"));
+        assert!(parsed["description"]
+            .as_str()
+            .is_some_and(|d| !d.is_empty()));
+        assert!(parsed["default_config"].is_object());
+        assert!(!parsed["supported_targets"].as_array().unwrap().is_empty());
+    }
+
+    #[test]
+    fn default_config_mirrors_worker_config_default() {
+        let m = build_manifest();
+        let cfg = WorkerConfig::default();
+        let from_manifest: WorkerConfig =
+            serde_json::from_value(m.default_config.clone()).expect("default_config parses");
+        assert_eq!(from_manifest, cfg);
+    }
+}
diff --git a/context-manager/src/ports.rs b/context-manager/src/ports.rs
new file mode 100644
index 00000000..9f24fbed
--- /dev/null
+++ b/context-manager/src/ports.rs
@@ -0,0 +1,128 @@
+//! Ports: the only seams between the pure context logic and the
+//! outside world (llm-router, iii state, wall clock).
+//!
+//! The binary wires production adapters (`crate::adapters`); BDD wires
+//! deterministic fakes. Handlers receive everything through [`Deps`],
+//! so the full `context::*` surface is testable without an engine.
+
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use serde::{Deserialize, Serialize};
+
+use crate::config::WorkerConfig;
+use crate::configuration::ConfigCell;
+use crate::types::Model;
+
+/// Resolves a model id to its router catalog record.
+#[async_trait]
+pub trait ModelResolver: Send + Sync {
+    /// `Ok(None)` when the router is installed but does not know the
+    /// model; `Err` when the router is unreachable or absent. Both
+    /// degrade to the conservative fallback (when allowed).
+    async fn get_model(&self, provider: Option<&str>, id: &str) -> Result<Option<Model>, String>;
+}
+
+/// One summariser call. The core renders the prompts (template,
+/// previous-summary anchoring, media stripping) so this port only
+/// transports them.
+#[async_trait]
+pub trait Summarizer: Send + Sync {
+    async fn summarize(&self, req: SummarizeRequest) -> Result<String, SummarizeError>;
+}
+
+#[derive(Debug, Clone, PartialEq)]
+pub struct SummarizeRequest {
+    pub system_prompt: String,
+    pub user_prompt: String,
+    /// Model id the summary runs on (same model as the request).
+    pub model: String,
+    pub provider: Option<String>,
+}
+
+/// Why a summariser call produced no summary.
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum SummarizeError {
+    /// `llm-router` is not installed / `router::chat` is not routable.
+    /// Compaction is unavailable; `context::compact` maps this to
+    /// `status: "overflow"` with a permanent error_kind in the log.
+    #[error("summariser unavailable: {0}")]
+    Unavailable(String),
+
+    /// The summariser ran and failed (provider error, stream error,
+    /// or the summariser itself overflowed).
+    #[error("summariser failed: {0}")]
+    Failed(String),
+
+    /// The summariser returned an empty summary.
+    #[error("summariser returned an empty summary")]
+    Empty,
+}
+
+/// A `{nonce, ts}` lease claim (harness `runtime/lease.ts` semantics:
+/// TTL is enforced by readers — a claim older than the TTL reads as
+/// free, folding crash recovery into acquisition).
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct LeaseRecord {
+    pub nonce: String,
+    /// Claim time, milliseconds since epoch.
+    pub ts: i64,
+}
+
+/// Key-value storage backing compaction leases (scope `context_lease`).
+#[async_trait]
+pub trait LeaseStore: Send + Sync {
+    async fn get(&self, scope: &str, key: &str) -> Result<Option<LeaseRecord>, String>;
+
+    /// `Some` writes the record; `None` clears the key.
+    async fn set(&self, scope: &str, key: &str, value: Option<LeaseRecord>) -> Result<(), String>;
+
+    /// Atomic read-modify-write: store `value` and return the previous
+    /// record in one operation, so exactly one concurrent acquirer sees
+    /// a free/expired prior and wins.
+    async fn swap(
+        &self,
+        scope: &str,
+        key: &str,
+        value: LeaseRecord,
+    ) -> Result<Option<LeaseRecord>, String>;
+}
+
+/// Wall clock, faked in tests to drive lease TTL expiry.
+pub trait Clock: Send + Sync {
+    /// Milliseconds since epoch.
+    fn now_ms(&self) -> i64;
+}
+
+/// Production clock.
+pub struct SystemClock;
+
+impl Clock for SystemClock {
+    fn now_ms(&self) -> i64 {
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .map(|d| d.as_millis() as i64)
+            .unwrap_or(0)
+    }
+}
+
+/// Everything a `context::*` handler needs.
+pub struct Deps {
+    /// Hot-swappable config snapshot, sourced from the `configuration`
+    /// worker and reloaded live (see [`crate::configuration`]).
+    pub config: ConfigCell,
+    pub resolver: Arc<dyn ModelResolver>,
+    pub summarizer: Arc<dyn Summarizer>,
+    pub leases: Arc<dyn LeaseStore>,
+    pub clock: Arc<dyn Clock>,
+}
+
+impl Deps {
+    /// Take a cheap snapshot of the current config. Handlers call this
+    /// once at the top of a request and read fields off the returned
+    /// `Arc`; a concurrent hot-reload swaps the cell without disturbing
+    /// snapshots already taken.
+    pub async fn config(&self) -> Arc<WorkerConfig> {
+        self.config.read().await.clone()
+    }
+}
diff --git a/context-manager/src/types.rs b/context-manager/src/types.rs
new file mode 100644
index 00000000..335c2022
--- /dev/null
+++ b/context-manager/src/types.rs
@@ -0,0 +1,396 @@
+//! Cross-cutting wire contracts shared by the agentic worker family.
+//!
+//! These shapes mirror the spec's "Cross-cutting contracts"
+//! (tech-specs/2026-06-agentic/README.md): `ContentBlock`,
+//! `AgentMessage`, `AgentFunction`, `Model`, `ThinkingLevel` — plus this
+//! worker's `ModelInput` (context-manager.md § Model input). Serde
+//! renames keep the wire format byte-compatible with the TypeScript
+//! definitions (and with session-manager's Rust copy).
+
+use std::collections::BTreeMap;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+
+/// Message role discriminator.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum Role {
+    User,
+    Assistant,
+    FunctionResult,
+    Custom,
+}
+
+/// Why an assistant message stopped generating.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum StopReason {
+    End,
+    Length,
+    FunctionCall,
+    Aborted,
+    Error,
+}
+
+/// Coarse error classification carried by failed assistant messages.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ErrorKind {
+    AuthExpired,
+    RateLimited,
+    ContextOverflow,
+    Transient,
+    Permanent,
+}
+
+/// Token / cost accounting reported by providers.
+#[derive(Debug, Clone, PartialEq, Default, Serialize, Deserialize, JsonSchema)]
+pub struct Usage {
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub input: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_read: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cache_write: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub reasoning: Option<u64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub cost_usd: Option<f64>,
+}
+
+/// The atomic unit of message content. A message's `content` is an
+/// ordered array of these.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum ContentBlock {
+    Text {
+        text: String,
+    },
+    Image {
+        /// MIME type, e.g. `image/png`.
+        mime: String,
+        /// Base64-encoded image bytes.
+        data: String,
+    },
+    Thinking {
+        text: String,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        signature: Option<String>,
+    },
+    FunctionCall {
+        /// Unique per call, echoed by the result.
+        id: String,
+        /// The iii function id to invoke.
+        function_id: String,
+        /// Model-produced arguments (JSON).
+        #[serde(default)]
+        arguments: Value,
+    },
+    FunctionResult {
+        function_call_id: String,
+        content: Vec<ContentBlock>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        is_error: Option<bool>,
+    },
+}
+
+/// The canonical transcript message union, discriminated by `role`.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "role", rename_all = "snake_case")]
+pub enum AgentMessage {
+    User {
+        content: Vec<ContentBlock>,
+        /// Milliseconds since epoch.
+        timestamp: i64,
+    },
+    Assistant {
+        content: Vec<ContentBlock>,
+        stop_reason: StopReason,
+        /// Provider's raw finish reason, passed through untouched.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        native_stop_reason: Option<String>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        error_message: Option<String>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        error_kind: Option<ErrorKind>,
+        /// Report-and-continue notices (e.g. dropped unsupported param).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        warnings: Option<Vec<String>>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        usage: Option<Usage>,
+        model: String,
+        provider: String,
+        /// Milliseconds since epoch.
+        timestamp: i64,
+    },
+    FunctionResult {
+        /// Echoes the `function_call` block id this result answers.
+        function_call_id: String,
+        function_id: String,
+        content: Vec<ContentBlock>,
+        /// Opaque structured payload kept alongside the rendered content.
+        #[serde(default)]
+        details: Value,
+        #[serde(default)]
+        is_error: bool,
+        /// Milliseconds since epoch.
+        timestamp: i64,
+    },
+    /// Escape hatch for app-specific transcript items (system notices,
+    /// UI markers, attachments, ...). No provider wire mapping —
+    /// `context::assemble` excludes these from the model-facing list.
+    Custom {
+        /// App-defined discriminator.
+        custom_type: String,
+        content: Vec<ContentBlock>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        display: Option<String>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        details: Option<Value>,
+        /// Milliseconds since epoch.
+        timestamp: i64,
+    },
+}
+
+impl AgentMessage {
+    pub fn role(&self) -> Role {
+        match self {
+            AgentMessage::User { .. } => Role::User,
+            AgentMessage::Assistant { .. } => Role::Assistant,
+            AgentMessage::FunctionResult { .. } => Role::FunctionResult,
+            AgentMessage::Custom { .. } => Role::Custom,
+        }
+    }
+
+    pub fn content(&self) -> &[ContentBlock] {
+        match self {
+            AgentMessage::User { content, .. }
+            | AgentMessage::Assistant { content, .. }
+            | AgentMessage::FunctionResult { content, .. }
+            | AgentMessage::Custom { content, .. } => content,
+        }
+    }
+
+    pub fn set_content(&mut self, new_content: Vec<ContentBlock>) {
+        match self {
+            AgentMessage::User { content, .. }
+            | AgentMessage::Assistant { content, .. }
+            | AgentMessage::FunctionResult { content, .. }
+            | AgentMessage::Custom { content, .. } => *content = new_content,
+        }
+    }
+
+    /// True when any top-level content block is a `function_result` —
+    /// used by tail selection to avoid cutting a result away from its
+    /// call. Function-result content holds rendered output (text/image),
+    /// not nested `function_result` blocks, so a top-level scan suffices.
+    pub fn has_function_result_block(&self) -> bool {
+        fn scan(blocks: &[ContentBlock]) -> bool {
+            blocks
+                .iter()
+                .any(|b| matches!(b, ContentBlock::FunctionResult { .. }))
+        }
+        scan(self.content())
+    }
+}
+
+/// Reasoning-effort tier; levels map to provider-native knobs via
+/// `Model.thinking_budgets`.
+#[derive(
+    Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord, Serialize, Deserialize, JsonSchema,
+)]
+#[serde(rename_all = "snake_case")]
+pub enum ThinkingLevel {
+    Minimal,
+    Low,
+    Medium,
+    High,
+    Xhigh,
+}
+
+/// Token limits a caller may supply inline to stay fully standalone.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+pub struct ModelLimits {
+    /// Total context window in tokens.
+    pub context_window: u64,
+    /// Maximum output tokens the model can produce per call.
+    pub max_output_tokens: u64,
+    /// Usable input budget when distinct from `context_window`.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub input_limit: Option<u64>,
+}
+
+/// How callers identify the target model (context-manager.md § Model
+/// input). Inline `limits` keep the call standalone; `id`/`provider`
+/// alone defer limit resolution to `router::models::get`.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+pub struct ModelInput {
+    /// Model id, e.g. `claude-sonnet-4`.
+    pub id: String,
+    /// Provider id, e.g. `anthropic`; disambiguates when a model id
+    /// exists on multiple providers.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub provider: Option<String>,
+    /// Inline token limits; when present no router lookup happens.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub limits: Option<ModelLimits>,
+}
+
+/// The capability record `llm-router` serves (README.md § Model
+/// descriptor). Only the fields this worker reads are modelled;
+/// unknown fields are ignored on deserialisation.
+#[derive(Debug, Clone, PartialEq, Default, Serialize, Deserialize, JsonSchema)]
+pub struct Model {
+    /// Model id, e.g. `claude-sonnet-4`.
+    pub id: String,
+    /// Provider id, e.g. `anthropic`.
+    #[serde(default)]
+    pub provider: String,
+    /// Total context window in tokens.
+    pub context_window: u64,
+    /// Maximum output tokens per call.
+    pub max_output_tokens: u64,
+    /// Usable input budget when distinct from `context_window`.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub input_limit: Option<u64>,
+    /// Reasoning-token budgets per thinking tier.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub thinking_budgets: Option<BTreeMap<ThinkingLevel, u64>>,
+}
+
+/// An invocation-surface schema entry (README.md § AgentFunction) —
+/// accepted by `context::count-tokens` so callers can include the
+/// `agent_trigger` schema in their estimate.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+pub struct AgentFunction {
+    /// Invocation surface name (`agent_trigger`, `submit_result`, or a
+    /// function id in native exposure mode).
+    pub name: String,
+    pub description: String,
+    /// JSON Schema for the function parameters.
+    #[serde(default)]
+    pub parameters: Value,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub label: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub execution_mode: Option<ExecutionMode>,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ExecutionMode {
+    Parallel,
+    Sequential,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn content_block_wire_shapes_match_spec() {
+        let blocks: Vec<ContentBlock> = serde_json::from_value(json!([
+            { "type": "text", "text": "hi" },
+            { "type": "image", "mime": "image/png", "data": "aGk=" },
+            { "type": "thinking", "text": "hmm", "signature": "sig" },
+            { "type": "function_call", "id": "c1", "function_id": "f::g", "arguments": { "a": 1 } },
+            { "type": "function_result", "function_call_id": "c1",
+              "content": [{ "type": "text", "text": "out" }], "is_error": false }
+        ]))
+        .unwrap();
+        assert_eq!(blocks.len(), 5);
+        let round = serde_json::to_value(&blocks).unwrap();
+        assert_eq!(round[0], json!({ "type": "text", "text": "hi" }));
+        assert_eq!(round[3]["function_id"], "f::g");
+    }
+
+    #[test]
+    fn agent_message_roundtrip_all_roles() {
+        let msgs: Vec<AgentMessage> = serde_json::from_value(json!([
+            { "role": "user", "content": [{ "type": "text", "text": "q" }], "timestamp": 1 },
+            { "role": "assistant", "content": [], "stop_reason": "end",
+              "model": "m1", "provider": "p1", "timestamp": 2 },
+            { "role": "function_result", "function_call_id": "c1", "function_id": "f::g",
+              "content": [], "details": { "x": 1 }, "is_error": false, "timestamp": 3 },
+            { "role": "custom", "custom_type": "notice", "content": [], "timestamp": 4 }
+        ]))
+        .unwrap();
+        assert_eq!(msgs[0].role(), Role::User);
+        assert_eq!(msgs[1].role(), Role::Assistant);
+        assert_eq!(msgs[2].role(), Role::FunctionResult);
+        assert_eq!(msgs[3].role(), Role::Custom);
+
+        let round = serde_json::to_value(&msgs).unwrap();
+        assert_eq!(round[1]["stop_reason"], "end");
+        assert!(round[1].get("error_message").is_none());
+        assert_eq!(round[2]["details"], json!({ "x": 1 }));
+    }
+
+    #[test]
+    fn model_input_minimal_and_inline_forms() {
+        let bare: ModelInput = serde_json::from_value(json!({ "id": "gpt-x" })).unwrap();
+        assert_eq!(bare.id, "gpt-x");
+        assert!(bare.provider.is_none() && bare.limits.is_none());
+
+        let inline: ModelInput = serde_json::from_value(json!({
+            "id": "m", "provider": "p",
+            "limits": { "context_window": 1000, "max_output_tokens": 100, "input_limit": 900 }
+        }))
+        .unwrap();
+        let limits = inline.limits.unwrap();
+        assert_eq!(limits.context_window, 1000);
+        assert_eq!(limits.input_limit, Some(900));
+    }
+
+    #[test]
+    fn model_descriptor_ignores_unknown_router_fields() {
+        let m: Model = serde_json::from_value(json!({
+            "id": "claude-sonnet-4", "provider": "anthropic",
+            "context_window": 200_000, "max_output_tokens": 16_000,
+            "thinking_budgets": { "low": 1024, "high": 8192 },
+            "display_name": "Sonnet", "supports_tools": true,
+            "pricing": { "input": 3.0 }
+        }))
+        .unwrap();
+        assert_eq!(m.context_window, 200_000);
+        assert_eq!(
+            m.thinking_budgets.unwrap().get(&ThinkingLevel::High),
+            Some(&8192)
+        );
+    }
+
+    #[test]
+    fn thinking_level_wire_names_are_snake_case() {
+        for (level, wire) in [
+            (ThinkingLevel::Minimal, "minimal"),
+            (ThinkingLevel::Low, "low"),
+            (ThinkingLevel::Medium, "medium"),
+            (ThinkingLevel::High, "high"),
+            (ThinkingLevel::Xhigh, "xhigh"),
+        ] {
+            assert_eq!(serde_json::to_value(level).unwrap(), json!(wire));
+        }
+    }
+
+    #[test]
+    fn has_function_result_block_detects_inline_results() {
+        let user_with_result: AgentMessage = serde_json::from_value(json!({
+            "role": "user",
+            "content": [{ "type": "function_result", "function_call_id": "c1", "content": [] }],
+            "timestamp": 1
+        }))
+        .unwrap();
+        assert!(user_with_result.has_function_result_block());
+
+        let plain: AgentMessage = serde_json::from_value(json!({
+            "role": "user", "content": [{ "type": "text", "text": "hi" }], "timestamp": 1
+        }))
+        .unwrap();
+        assert!(!plain.has_function_result_block());
+    }
+}
diff --git a/context-manager/tests/bdd.rs b/context-manager/tests/bdd.rs
new file mode 100644
index 00000000..92dd1ada
--- /dev/null
+++ b/context-manager/tests/bdd.rs
@@ -0,0 +1,28 @@
+//! Cucumber BDD entry point. All `.feature` files under
+//! `tests/features/**` are discovered at runtime.
+
+mod common;
+mod steps;
+
+use cucumber::World as _;
+
+use crate::common::world::ContextWorld;
+
+#[tokio::main]
+async fn main() {
+    // Bring up the shared engine connection + in-process registration
+    // exactly once. On a host without an engine this returns None and
+    // every `@engine` scenario short-circuits into a soft skip.
+    let _ = common::engine::get_or_init().await;
+
+    ContextWorld::cucumber()
+        .max_concurrent_scenarios(1)
+        .before(|_feature, _rule, _scenario, world| {
+            Box::pin(async move {
+                world.reset_pure();
+                world.engine = common::engine::get_or_init().await;
+            })
+        })
+        .run_and_exit("tests/features")
+        .await;
+}
diff --git a/context-manager/tests/common/engine.rs b/context-manager/tests/common/engine.rs
new file mode 100644
index 00000000..29b45b85
--- /dev/null
+++ b/context-manager/tests/common/engine.rs
@@ -0,0 +1,63 @@
+//! Connect-or-skip wrapper around the iii SDK.
+//!
+//! One engine connection per test binary process via OnceCell. On
+//! success the production surface (the four `context::*` functions
+//! with production adapters) is registered in-process so @engine
+//! scenarios drive the exact code the binary would run.
+
+use std::sync::Arc;
+use std::time::Duration;
+
+use iii_sdk::{register_worker, IIIError, InitOptions, TriggerRequest, III};
+use serde_json::json;
+use tokio::sync::OnceCell;
+
+const DEFAULT_WS_URL: &str = "ws://127.0.0.1:49134";
+
+static ENGINE: OnceCell<Option<Arc<III>>> = OnceCell::const_new();
+
+pub fn ws_url() -> String {
+    std::env::var("III_ENGINE_WS_URL").unwrap_or_else(|_| DEFAULT_WS_URL.to_string())
+}
+
+async fn try_connect_raw() -> Option<Arc<III>> {
+    let url = ws_url();
+    let iii = Arc::new(register_worker(&url, InitOptions::default()));
+
+    for _ in 0..20 {
+        tokio::time::sleep(Duration::from_millis(250)).await;
+        let probe = iii
+            .trigger(TriggerRequest {
+                function_id: "engine::workers::list".to_string(),
+                payload: json!({}),
+                action: None,
+                timeout_ms: Some(800),
+            })
+            .await;
+        match probe {
+            Ok(_) => return Some(iii),
+            Err(IIIError::NotConnected) => continue,
+            Err(_) => continue,
+        }
+    }
+
+    eprintln!(
+        "[skip] iii engine not reachable at {url}; \
+         set III_ENGINE_WS_URL or start `iii` to enable engine-bound BDD scenarios"
+    );
+    iii.shutdown_async().await;
+    None
+}
+
+/// Get-or-init the shared engine handle, registering the production
+/// context-manager surface in-process on first success.
+pub async fn get_or_init() -> Option<Arc<III>> {
+    ENGINE
+        .get_or_init(|| async {
+            let iii = try_connect_raw().await?;
+            super::workers::register_all(&iii).await;
+            Some(iii)
+        })
+        .await
+        .clone()
+}
diff --git a/context-manager/tests/common/fakes.rs b/context-manager/tests/common/fakes.rs
new file mode 100644
index 00000000..39a29fd6
--- /dev/null
+++ b/context-manager/tests/common/fakes.rs
@@ -0,0 +1,259 @@
+//! Deterministic fakes behind the ports for @pure scenarios.
+//!
+//! Each fake records what the production code asked of it (summariser
+//! prompts, resolver lookups) so scenarios can assert not just the
+//! response but the *interaction* — e.g. that a second compaction was
+//! anchored on the previous summary instead of starting over.
+
+use std::collections::HashMap;
+use std::sync::atomic::{AtomicBool, AtomicI64, Ordering};
+use std::sync::Mutex;
+
+use async_trait::async_trait;
+
+use context_manager::ports::{
+    Clock, LeaseRecord, LeaseStore, ModelResolver, SummarizeError, SummarizeRequest, Summarizer,
+};
+use context_manager::types::Model;
+
+// ---------------------------------------------------------------------------
+// Clock
+// ---------------------------------------------------------------------------
+
+/// Starts at a fixed epoch so lease timestamps are reproducible;
+/// scenarios advance it explicitly to drive TTL expiry.
+pub struct FakeClock {
+    now_ms: AtomicI64,
+}
+
+pub const CLOCK_START_MS: i64 = 1_000_000;
+
+impl FakeClock {
+    pub fn new() -> Self {
+        Self {
+            now_ms: AtomicI64::new(CLOCK_START_MS),
+        }
+    }
+
+    pub fn advance(&self, ms: i64) {
+        self.now_ms.fetch_add(ms, Ordering::SeqCst);
+    }
+}
+
+impl Clock for FakeClock {
+    fn now_ms(&self) -> i64 {
+        self.now_ms.load(Ordering::SeqCst)
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Model resolver
+// ---------------------------------------------------------------------------
+
+/// Scripted router catalog. `unavailable` simulates `llm-router` not
+/// being installed (every lookup errors); an id missing from `models`
+/// simulates a router that does not know the model.
+pub struct FakeModelResolver {
+    pub models: Mutex<HashMap<String, Model>>,
+    pub unavailable: AtomicBool,
+    /// Recorded `(provider, id)` lookups.
+    pub calls: Mutex<Vec<(Option<String>, String)>>,
+}
+
+impl FakeModelResolver {
+    pub fn new() -> Self {
+        Self {
+            models: Mutex::new(HashMap::new()),
+            unavailable: AtomicBool::new(false),
+            calls: Mutex::new(Vec::new()),
+        }
+    }
+
+    pub fn insert(&self, model: Model) {
+        self.models
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .insert(model.id.clone(), model);
+    }
+}
+
+#[async_trait]
+impl ModelResolver for FakeModelResolver {
+    async fn get_model(&self, provider: Option<&str>, id: &str) -> Result<Option<Model>, String> {
+        self.calls
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .push((provider.map(str::to_string), id.to_string()));
+        if self.unavailable.load(Ordering::SeqCst) {
+            return Err("router::models::get is not routable".to_string());
+        }
+        Ok(self
+            .models
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .get(id)
+            .cloned())
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Summariser
+// ---------------------------------------------------------------------------
+
+#[derive(Debug, Clone)]
+pub enum SummarizerScript {
+    Return(String),
+    Unavailable,
+    Fail(String),
+    Empty,
+}
+
+pub struct FakeSummarizer {
+    pub script: Mutex<SummarizerScript>,
+    /// Every request the production pipeline sent, in order.
+    pub calls: Mutex<Vec<SummarizeRequest>>,
+}
+
+impl FakeSummarizer {
+    pub fn new() -> Self {
+        Self {
+            // A recognisable default so scenarios that don't care about
+            // the text still see a plausible summary flow through.
+            script: Mutex::new(SummarizerScript::Return(
+                "## Goal\n- (scripted default summary)".to_string(),
+            )),
+            calls: Mutex::new(Vec::new()),
+        }
+    }
+
+    pub fn set_script(&self, script: SummarizerScript) {
+        *self
+            .script
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner()) = script;
+    }
+
+    pub fn calls(&self) -> Vec<SummarizeRequest> {
+        self.calls
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .clone()
+    }
+}
+
+#[async_trait]
+impl Summarizer for FakeSummarizer {
+    async fn summarize(&self, req: SummarizeRequest) -> Result<String, SummarizeError> {
+        self.calls
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .push(req);
+        let script = self
+            .script
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .clone();
+        match script {
+            SummarizerScript::Return(summary) => Ok(summary),
+            SummarizerScript::Unavailable => Err(SummarizeError::Unavailable(
+                "router::chat is not routable".to_string(),
+            )),
+            SummarizerScript::Fail(reason) => Err(SummarizeError::Failed(reason)),
+            SummarizerScript::Empty => Err(SummarizeError::Empty),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Lease store
+// ---------------------------------------------------------------------------
+
+/// In-memory `(scope, key) -> LeaseRecord` map. `swap` is genuinely
+/// atomic under the mutex, mirroring `FsLeaseStore`'s under-lock swap.
+/// `fail_all` simulates a lease-store outage (every call errors).
+pub struct InMemoryLeaseStore {
+    records: Mutex<HashMap<(String, String), LeaseRecord>>,
+    pub fail_all: AtomicBool,
+}
+
+impl InMemoryLeaseStore {
+    pub fn new() -> Self {
+        Self {
+            records: Mutex::new(HashMap::new()),
+            fail_all: AtomicBool::new(false),
+        }
+    }
+
+    fn check(&self) -> Result<(), String> {
+        if self.fail_all.load(Ordering::SeqCst) {
+            return Err("state is unavailable".to_string());
+        }
+        Ok(())
+    }
+
+    /// Direct write for Given steps (a foreign holder's claim).
+    pub fn put(&self, scope: &str, key: &str, record: LeaseRecord) {
+        self.records
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .insert((scope.to_string(), key.to_string()), record);
+    }
+
+    /// Direct read for Then steps.
+    pub fn peek(&self, scope: &str, key: &str) -> Option<LeaseRecord> {
+        self.records
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .get(&(scope.to_string(), key.to_string()))
+            .cloned()
+    }
+
+    /// All keys currently holding a claim (any scope).
+    pub fn keys(&self) -> Vec<String> {
+        self.records
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner())
+            .keys()
+            .map(|(_, key)| key.clone())
+            .collect()
+    }
+}
+
+#[async_trait]
+impl LeaseStore for InMemoryLeaseStore {
+    async fn get(&self, scope: &str, key: &str) -> Result<Option<LeaseRecord>, String> {
+        self.check()?;
+        Ok(self.peek(scope, key))
+    }
+
+    async fn set(&self, scope: &str, key: &str, value: Option<LeaseRecord>) -> Result<(), String> {
+        self.check()?;
+        let mut records = self
+            .records
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner());
+        match value {
+            Some(record) => {
+                records.insert((scope.to_string(), key.to_string()), record);
+            }
+            None => {
+                records.remove(&(scope.to_string(), key.to_string()));
+            }
+        }
+        Ok(())
+    }
+
+    async fn swap(
+        &self,
+        scope: &str,
+        key: &str,
+        value: LeaseRecord,
+    ) -> Result<Option<LeaseRecord>, String> {
+        self.check()?;
+        let mut records = self
+            .records
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner());
+        Ok(records.insert((scope.to_string(), key.to_string()), value))
+    }
+}
diff --git a/context-manager/tests/common/mod.rs b/context-manager/tests/common/mod.rs
new file mode 100644
index 00000000..14c3a19c
--- /dev/null
+++ b/context-manager/tests/common/mod.rs
@@ -0,0 +1,4 @@
+pub mod engine;
+pub mod fakes;
+pub mod workers;
+pub mod world;
diff --git a/context-manager/tests/common/workers.rs b/context-manager/tests/common/workers.rs
new file mode 100644
index 00000000..1aa1e741
--- /dev/null
+++ b/context-manager/tests/common/workers.rs
@@ -0,0 +1,86 @@
+//! In-process registration of the production `context-manager` surface
+//! against the shared SDK handle.
+//!
+//! Re-uses the production entry points in the binary's boot order —
+//! production adapters (`RouterModelResolver`, `RouterSummarizer`,
+//! `FsLeaseStore`, `SystemClock`) and `functions::register_all` — so
+//! @engine scenarios exercise identical code paths end to end,
+//! including real filesystem lease writes (under a per-process temp
+//! dir). `llm-router` is absent on the test engine, which is exactly
+//! the degraded mode the spec defines (fallback limits; compact ->
+//! overflow).
+
+use std::sync::Arc;
+use std::time::Duration;
+
+use iii_sdk::{TriggerRequest, III};
+use tokio::sync::{OnceCell, RwLock};
+
+use context_manager::adapters::fs_lease::FsLeaseStore;
+use context_manager::adapters::router::{RouterModelResolver, RouterSummarizer};
+use context_manager::config::WorkerConfig;
+use context_manager::functions;
+use context_manager::ports::{Deps, SystemClock};
+
+static REGISTERED: OnceCell<()> = OnceCell::const_new();
+
+/// Idempotent: the first caller registers; subsequent callers reuse.
+pub async fn register_all(iii: &Arc<III>) {
+    REGISTERED
+        .get_or_init(|| async {
+            let cfg = WorkerConfig::default();
+            let deps = Arc::new(Deps {
+                resolver: Arc::new(RouterModelResolver::new(iii.clone())),
+                summarizer: Arc::new(RouterSummarizer::new(
+                    iii.clone(),
+                    // Keep the @engine compact-overflow scenario fast:
+                    // router::chat is absent, so only the trigger error
+                    // path runs and a long stream budget buys nothing.
+                    5_000,
+                )),
+                leases: Arc::new(
+                    FsLeaseStore::new(
+                        std::env::temp_dir()
+                            .join(format!("context-manager-bdd-leases-{}", std::process::id())),
+                    )
+                    .expect("create bdd lease dir"),
+                ),
+                clock: Arc::new(SystemClock),
+                config: Arc::new(RwLock::new(Arc::new(cfg))),
+            });
+            functions::register_all(iii, &deps);
+
+            // Registrations flow over one connection in order, so the
+            // last one being routable means the batch landed.
+            wait_until_routable(iii, "context::count-tokens").await;
+        })
+        .await;
+}
+
+/// Poll the engine until `function_id` is routable, panicking after a
+/// deadline. An `Err` carrying a `context/` code also counts as
+/// routable — it proves the call reached the production handler.
+async fn wait_until_routable(iii: &Arc<III>, function_id: &str) {
+    let deadline = std::time::Instant::now() + Duration::from_secs(10);
+    loop {
+        let res = iii
+            .trigger(TriggerRequest {
+                function_id: function_id.to_string(),
+                payload: serde_json::json!({ "model": { "id": "bdd-readiness-probe" } }),
+                action: None,
+                timeout_ms: Some(1_000),
+            })
+            .await;
+        match res {
+            Ok(_) => return,
+            Err(e) if e.to_string().contains("context/") => return,
+            Err(e) => {
+                assert!(
+                    std::time::Instant::now() < deadline,
+                    "{function_id} did not become routable within 10s: {e}"
+                );
+                tokio::time::sleep(Duration::from_millis(50)).await;
+            }
+        }
+    }
+}
diff --git a/context-manager/tests/common/world.rs b/context-manager/tests/common/world.rs
new file mode 100644
index 00000000..b833e7c4
--- /dev/null
+++ b/context-manager/tests/common/world.rs
@@ -0,0 +1,216 @@
+//! Per-scenario World.
+//!
+//! Carries two stacks side by side:
+//!
+//! - **@pure**: a fresh production stack per scenario — the real
+//!   `handle` functions over real core logic, with deterministic fakes
+//!   behind the ports (scripted router catalog, scripted summariser,
+//!   in-memory atomic lease store, settable clock). No engine; counts,
+//!   placeholders, and lease timestamps are exact.
+//! - **@engine**: the shared engine connection (`None` when
+//!   unreachable, so `@engine` steps soft-skip) driving the same
+//!   handlers registered in-process with the *production* adapters.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Duration;
+
+use cucumber::World;
+use iii_sdk::{TriggerRequest, III};
+use serde_json::Value;
+use tokio::sync::RwLock;
+
+use context_manager::config::WorkerConfig;
+use context_manager::error::ContextError;
+use context_manager::functions::{assemble, compact, count_tokens, prune};
+use context_manager::ports::Deps;
+
+use super::fakes::{FakeClock, FakeModelResolver, FakeSummarizer, InMemoryLeaseStore};
+
+#[derive(World)]
+#[world(init = Self::new)]
+pub struct ContextWorld {
+    /// @pure stack, rebuilt before every scenario.
+    pub config: WorkerConfig,
+    pub resolver: Arc<FakeModelResolver>,
+    pub summarizer: Arc<FakeSummarizer>,
+    pub leases: Arc<InMemoryLeaseStore>,
+    pub clock: Arc<FakeClock>,
+
+    /// The history the Given steps build, as raw JSON messages.
+    pub messages: Vec<Value>,
+    /// Named model inputs (`ModelInput` JSON) the call steps reference.
+    pub model_inputs: HashMap<String, Value>,
+    /// Monotonic timestamp for built messages.
+    pub next_ts: i64,
+
+    /// Outcome of the most recent call (pure or engine).
+    pub last_response: Option<Value>,
+    pub last_error: Option<String>,
+    /// Set when an engine-dependent step soft-skipped (no engine
+    /// reachable); generic assertion steps then pass vacuously.
+    pub soft_skipped: bool,
+
+    /// Alias -> value used by `${X}` substitution in docstrings.
+    pub aliases: HashMap<String, String>,
+
+    /// Shared engine connection; `None` soft-skips @engine steps.
+    pub engine: Option<Arc<III>>,
+}
+
+impl std::fmt::Debug for ContextWorld {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("ContextWorld")
+            .field("engine_connected", &self.engine.is_some())
+            .field("messages", &self.messages.len())
+            .field("last_error", &self.last_error)
+            .finish()
+    }
+}
+
+impl ContextWorld {
+    pub fn new() -> Self {
+        Self {
+            config: WorkerConfig::default(),
+            resolver: Arc::new(FakeModelResolver::new()),
+            summarizer: Arc::new(FakeSummarizer::new()),
+            leases: Arc::new(InMemoryLeaseStore::new()),
+            clock: Arc::new(FakeClock::new()),
+            messages: Vec::new(),
+            model_inputs: HashMap::new(),
+            next_ts: 1_000,
+            last_response: None,
+            last_error: None,
+            soft_skipped: false,
+            aliases: HashMap::new(),
+            engine: None,
+        }
+    }
+
+    /// Fresh pure stack + cleared bookkeeping. Runs before every
+    /// scenario.
+    pub fn reset_pure(&mut self) {
+        *self = Self::new();
+    }
+
+    /// Build the production `Deps` over the current fakes + config. The
+    /// per-scenario `config` is wrapped in a fresh `ConfigCell` so handlers
+    /// take the snapshot the Given steps set up.
+    pub fn deps(&self) -> Deps {
+        Deps {
+            config: Arc::new(RwLock::new(Arc::new(self.config.clone()))),
+            resolver: self.resolver.clone(),
+            summarizer: self.summarizer.clone(),
+            leases: self.leases.clone(),
+            clock: self.clock.clone(),
+        }
+    }
+
+    /// Mark the scenario as engine-dependent-but-skipped. Returns true
+    /// (and records the skip) when no engine is reachable.
+    pub fn skip_without_engine(&mut self) -> bool {
+        if self.engine.is_none() {
+            self.soft_skipped = true;
+            return true;
+        }
+        false
+    }
+
+    /// Next message timestamp (monotonic).
+    pub fn tick(&mut self) -> i64 {
+        self.next_ts += 1;
+        self.next_ts
+    }
+
+    /// The `ModelInput` JSON registered under `name`, defaulting to a
+    /// bare `{ "id": name }`.
+    pub fn model_input(&self, name: &str) -> Value {
+        self.model_inputs
+            .get(name)
+            .cloned()
+            .unwrap_or_else(|| serde_json::json!({ "id": name }))
+    }
+
+    /// Replace `${alias}` placeholders; `${messages}` expands to the
+    /// built history as a JSON array.
+    pub fn substitute(&self, text: &str) -> String {
+        let mut out = text.to_string();
+        if out.contains("${messages}") {
+            let rendered =
+                serde_json::to_string(&self.messages).expect("history serializes to JSON");
+            out = out.replace("${messages}", &rendered);
+        }
+        for (alias, value) in &self.aliases {
+            out = out.replace(&format!("${{{alias}}}"), value);
+        }
+        out
+    }
+
+    /// Drive a `context::*` function through the production handler
+    /// (@pure). Mirrors the bus contract: serde rejects malformed
+    /// requests, handler errors carry `code: message`.
+    pub async fn call_pure(&mut self, function: &str, payload: Value) {
+        let deps = self.deps();
+        let result = dispatch(&deps, function, payload).await;
+        self.record_outcome(result);
+    }
+
+    /// Drive a function over the live engine (@engine).
+    pub async fn call_engine(&mut self, function: &str, payload: Value) {
+        let Some(engine) = self.engine.clone() else {
+            self.soft_skipped = true;
+            return;
+        };
+        let result = tokio::time::timeout(
+            Duration::from_secs(15),
+            engine.trigger(TriggerRequest {
+                function_id: function.to_string(),
+                payload,
+                action: None,
+                timeout_ms: Some(12_000),
+            }),
+        )
+        .await;
+        let outcome = match result {
+            Ok(Ok(v)) => Ok(v),
+            Ok(Err(e)) => Err(e.to_string()),
+            Err(_) => Err("engine trigger timed out".to_string()),
+        };
+        self.record_outcome(outcome);
+    }
+
+    fn record_outcome(&mut self, result: Result<Value, String>) {
+        match result {
+            Ok(v) => {
+                self.last_response = Some(v);
+                self.last_error = None;
+            }
+            Err(e) => {
+                self.last_response = None;
+                self.last_error = Some(e);
+            }
+        }
+    }
+}
+
+fn parse<T: serde::de::DeserializeOwned>(payload: Value) -> Result<T, String> {
+    serde_json::from_value(payload).map_err(|e| format!("invalid request: {e}"))
+}
+
+fn out<T: serde::Serialize>(result: Result<T, ContextError>) -> Result<Value, String> {
+    result
+        .map(|v| serde_json::to_value(v).expect("responses always serialize"))
+        .map_err(|e| e.to_string())
+}
+
+/// Dispatch by function id to the same `handle` functions the binary
+/// registers — the full production path minus the WebSocket.
+pub async fn dispatch(deps: &Deps, function: &str, payload: Value) -> Result<Value, String> {
+    match function {
+        "context::assemble" => out(assemble::handle(deps, parse(payload)?).await),
+        "context::compact" => out(compact::handle(deps, parse(payload)?).await),
+        "context::prune" => out(prune::handle(deps, parse(payload)?).await),
+        "context::count-tokens" => out(count_tokens::handle(deps, parse(payload)?).await),
+        other => Err(format!("unknown function {other}")),
+    }
+}
diff --git a/context-manager/tests/features/assemble.feature b/context-manager/tests/features/assemble.feature
new file mode 100644
index 00000000..af9770f5
--- /dev/null
+++ b/context-manager/tests/features/assemble.feature
@@ -0,0 +1,213 @@
+@pure
+Feature: context::assemble — the model-ready context pipeline
+
+  Contract (context-manager.md § context::assemble): count -> (if over)
+  prune function outputs -> (if still over) compact the head ->
+  assemble the final list. The response reports what actually happened
+  (`applied`), the budget it fit into (`usable`), and how the model was
+  resolved. Degradations (busy lease, failed summariser, disabled
+  steps) are best effort and visible — never an error on the hot path.
+
+  # Prevents: the happy path being mangled — a context under budget
+  # must pass through byte-identical, with nothing applied and no
+  # summariser cost.
+  Scenario: under budget passes through untouched
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message "hello"
+    And an assistant message "hi there"
+    When I assemble the history with model "big" and system prompt "You are helpful."
+    Then the call succeeds
+    And the response field "system_prompt" is "You are helpful."
+    And the response field "token_count" does not exceed 100
+    And the response field "applied.pruned" is false
+    And the response field "applied.compacted" is false
+    And the response field "applied.pruned_tokens" is 0
+    And the response has no field "applied.summary"
+    And the response messages equal the request history
+    And the summariser was never invoked
+
+  # Prevents: prune kicking in while the context still fits — pruning
+  # under budget destroys context for nothing.
+  Scenario: prune never runs under budget
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And config "protect_recent_tokens" is 0
+    And config "min_free_tokens" is 1
+    And a user message "task"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~5000 tokens
+    And a user message "next"
+    And a user message "done"
+    When I assemble the history with model "big"
+    Then the response field "applied.pruned" is false
+    And response message 2 text has 20000 chars
+
+  # Prevents: an over-budget context reaching the model when freeing
+  # old tool outputs would have been enough — the cheap pass must run
+  # first and suffice alone.
+  Scenario: over budget, prune alone brings the context home
+    Given inline model "small" with context window 5000 and max output 500
+    And config "protect_recent_tokens" is 0
+    And config "min_free_tokens" is 1
+    And the summariser returns "should never be needed"
+    And a user message "task"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~5000 tokens
+    And an assistant message "ok"
+    And a user message "next"
+    And a user message "done"
+    When I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "applied.pruned" is true
+    And the response field "applied.pruned_tokens" is 4992
+    And the response field "applied.compacted" is false
+    And the response field "token_count" does not exceed 4000
+    And response message 2 text is "[output pruned: was ~5000 tokens]"
+    And the response messages have as many messages as the request
+    And call/result pairing is intact in the response messages
+    And the summariser was never invoked
+
+  # Prevents: prune running against the caller's explicit wishes — and
+  # proves the over-budget result is still returned, visibly over.
+  Scenario: disabling prune and compaction returns a visible best effort
+    Given inline model "small" with context window 5000 and max output 500
+    And config "protect_recent_tokens" is 0
+    And config "min_free_tokens" is 1
+    And a user message "task"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~5000 tokens
+    And an assistant message "ok"
+    And a user message "next"
+    And a user message "done"
+    When I assemble the history with model "small" and options:
+      """
+      { "allow_prune": false, "allow_compaction": false }
+      """
+    Then the call succeeds
+    And the response field "applied.pruned" is false
+    And the response field "applied.compacted" is false
+    And the response field "token_count" exceeds 4000
+    And response message 2 text has 20000 chars
+
+  # Prevents: protected functions being pruned on the assemble path —
+  # the per-call protection list must reach the prune pass.
+  Scenario: protected functions survive the assemble prune pass
+    Given inline model "small" with context window 5000 and max output 500
+    And config "protect_recent_tokens" is 0
+    And config "min_free_tokens" is 1
+    And a user message "task"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~5000 tokens
+    And an assistant message "ok"
+    And a user message "next"
+    And a user message "done"
+    When I assemble the history with model "small" and options:
+      """
+      { "protected_functions": ["shell::run"], "allow_compaction": false }
+      """
+    Then the response field "applied.pruned" is false
+    And response message 2 text has 20000 chars
+    And the response field "token_count" exceeds 4000
+
+  # Prevents: histories with nothing prunable never compacting — when
+  # prune cannot help, compaction must take over and the returned tail
+  # must map onto the REQUEST message indices.
+  Scenario: still over after prune, the head is compacted away
+    Given inline model "small" with context window 5000 and max output 500
+    And the summariser returns "## Goal\n- ship the feature"
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small" and system prompt "Base prompt."
+    Then the call succeeds
+    And the response field "applied.compacted" is true
+    And the response field "applied.summary" is "## Goal\n- ship the feature"
+    And the response field "applied.tail_start_index" is 3
+    And the response field "applied.tokens_before" exceeds 6000
+    And the response field "token_count" does not exceed 4000
+    And the response field "system_prompt" contains "Base prompt."
+    And the response field "system_prompt" contains "# Conversation summary"
+    And the response field "system_prompt" contains "ship the feature"
+    And the response messages start at request message 3
+    And the summariser was invoked 1 time
+    And no lease claim remains
+
+  # Prevents: both passes fighting instead of stacking — prune frees
+  # the tool output, compaction then folds the rest of the head.
+  Scenario: prune and compaction stack when one is not enough
+    Given inline model "small" with context window 5000 and max output 500
+    And config "protect_recent_tokens" is 0
+    And config "min_free_tokens" is 1
+    And the summariser returns "## Goal\n- both passes ran"
+    And a user message "task"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~3000 tokens
+    And an assistant message "ok"
+    And a user message of ~4000 tokens
+    And an assistant message "r1"
+    And a user message "recent"
+    And an assistant message "fine"
+    When I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "applied.pruned" is true
+    And the response field "applied.pruned_tokens" is 2992
+    And the response field "applied.compacted" is true
+    And the response field "applied.tail_start_index" is 5
+    And the response field "token_count" does not exceed 4000
+    And the response messages start at request message 5
+    And call/result pairing is intact in the response messages
+
+  # Prevents: compaction running against the caller's explicit wishes.
+  Scenario: allow_compaction false leaves an over-budget context as is
+    Given inline model "small" with context window 5000 and max output 500
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small" and options:
+      """
+      { "allow_compaction": false }
+      """
+    Then the call succeeds
+    And the response field "applied.compacted" is false
+    And the response field "token_count" exceeds 4000
+    And the summariser was never invoked
+
+  # Prevents: a busy lease failing the turn — assemble must degrade to
+  # best effort, spend no summariser call, and leave the holder alone.
+  Scenario: a busy lease degrades assemble to best effort
+    Given inline model "small" with context window 5000 and max output 500
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    And a foreign lease on the history's default key taken 0 ms ago
+    When I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "applied.compacted" is false
+    And the response field "token_count" exceeds 4000
+    And the summariser was never invoked
+    And the lease on the history's default key still belongs to the foreign holder
+
+  # Prevents: a summariser failure erroring the hot path or leaving the
+  # lease stuck — the turn proceeds over budget, the lease is freed.
+  Scenario: a summariser failure degrades assemble to best effort
+    Given inline model "small" with context window 5000 and max output 500
+    And the summariser fails with "provider exploded"
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "applied.compacted" is false
+    And the response field "token_count" exceeds 4000
+    And no lease claim remains
diff --git a/context-manager/tests/features/assemble_round_trip.feature b/context-manager/tests/features/assemble_round_trip.feature
new file mode 100644
index 00000000..17154e0b
--- /dev/null
+++ b/context-manager/tests/features/assemble_round_trip.feature
@@ -0,0 +1,96 @@
+@pure
+Feature: the compaction round trip — summaries persist, converge, never grow
+
+  Contract (context-manager.md § The compaction round trip): the worker
+  never persists a summary; the caller does. previous_summary is
+  rendered into the system prompt under "# Conversation summary"; when
+  compaction triggers again the summariser UPDATES that summary instead
+  of starting over, so summaries converge. Callers that skip
+  persistence still get correct output — at one summariser call per
+  over-budget request.
+
+  # Prevents: a persisted summary silently vanishing from the model's
+  # view — the round trip's whole point is that the summary comes back
+  # every turn.
+  Scenario: a previous summary is rendered into the system prompt
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message "hello again"
+    When I assemble the history with model "big", system prompt "Base." and options:
+      """
+      { "previous_summary": "## Goal\n- carried forward from last time" }
+      """
+    Then the call succeeds
+    And the response field "applied.compacted" is false
+    And the response field "system_prompt" contains "Base."
+    And the response field "system_prompt" contains "# Conversation summary"
+    And the response field "system_prompt" contains "carried forward from last time"
+    And the summariser was never invoked
+
+  # Prevents: stale and fresh summaries stacking up in the prompt — a
+  # new compaction must REPLACE the rendered summary, not append to it.
+  Scenario: a fresh compaction replaces the rendered previous summary
+    Given inline model "small" with context window 5000 and max output 500
+    And the summariser returns "NEW-SUMMARY-TEXT"
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small" and options:
+      """
+      { "previous_summary": "OLD-ANCHOR-TEXT" }
+      """
+    Then the response field "applied.compacted" is true
+    And the response field "applied.summary" is "NEW-SUMMARY-TEXT"
+    And the response field "system_prompt" contains "NEW-SUMMARY-TEXT"
+    And the response field "system_prompt" does not contain "OLD-ANCHOR-TEXT"
+
+  # Prevents: summaries growing forever — the second compaction must
+  # run in update mode, anchored on the previous summary.
+  Scenario: re-compaction is anchored on the previous summary
+    Given inline model "small" with context window 5000 and max output 500
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small" and options:
+      """
+      { "previous_summary": "OLD-ANCHOR-TEXT" }
+      """
+    Then the response field "applied.compacted" is true
+    And the summariser system prompt contains "Update the anchored summary"
+    And the summariser system prompt contains "OLD-ANCHOR-TEXT"
+
+  # Prevents: a first compaction hallucinating an anchor it never had.
+  Scenario: a first compaction starts fresh
+    Given inline model "small" with context window 5000 and max output 500
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small"
+    Then the response field "applied.compacted" is true
+    And the summariser system prompt contains "Create a new anchored summary"
+    And the summariser system prompt does not contain "<previous-summary>"
+
+  # Prevents: misunderstanding the cost model — a caller that never
+  # persists the summary stays correct but pays one summariser call per
+  # over-budget request. This scenario IS the documented trade-off.
+  Scenario: skipping persistence costs one summariser call per request
+    Given inline model "small" with context window 5000 and max output 500
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small"
+    And I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "applied.compacted" is true
+    And the summariser was invoked 2 times
diff --git a/context-manager/tests/features/compact.feature b/context-manager/tests/features/compact.feature
new file mode 100644
index 00000000..a531fccf
--- /dev/null
+++ b/context-manager/tests/features/compact.feature
@@ -0,0 +1,176 @@
+@pure
+Feature: context::compact — summarise the head, keep a verbatim tail
+
+  Contract (context-manager.md § context::compact): summarise the head
+  of a history into a single compaction summary, keeping a recent tail
+  verbatim. Transient and storage-agnostic — the caller persists the
+  result; no session is touched. The response is a status union:
+  ok | busy | empty | overflow. `tail_start_index` indexes the request
+  messages array (this worker never sees storage entry ids).
+
+  # Prevents: summarising nothing and burning an LLM call on an empty
+  # transcript.
+  Scenario: an empty history is empty, not an error
+    Given an empty history
+    And the router knows model "big" with context window 200000 and max output 8000
+    When I compact the history with model "big"
+    Then the call succeeds
+    And the response field "status" is "empty"
+    And the summariser was never invoked
+    And no lease claim remains
+
+  # Prevents: the core contract breaking — the head must be summarised,
+  # the tail index must point into the REQUEST array, and the lease
+  # must be released for the next caller.
+  Scenario: a long head is summarised and the recent tail survives verbatim
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And the summariser returns "## Goal\n- finish the migration"
+    And a user message of ~9000 tokens
+    And an assistant message "r1"
+    And a user message "q2"
+    And an assistant message "r2"
+    And a user message "q3"
+    And an assistant message "r3"
+    When I compact the history with model "big"
+    Then the call succeeds
+    And the response field "status" is "ok"
+    And the response field "summary" is "## Goal\n- finish the migration"
+    And the response field "tail_start_index" is 2
+    And the response field "tokens_before" exceeds 9000
+    And the response field "tokens_after" does not exceed 1000
+    And the response field "used_prior_summary" is false
+    And the summariser was invoked 1 time
+    And no lease claim remains
+
+  # Prevents: tail_turns being ignored — operators tune how much recent
+  # conversation survives compaction verbatim.
+  Scenario: tail_turns controls how many recent turns survive
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message of ~9000 tokens
+    And an assistant message "r1"
+    And a user message "q2"
+    And an assistant message "r2"
+    And a user message "q3"
+    And an assistant message "r3"
+    When I compact the history with model "big" and options:
+      """
+      { "tail_turns": 1 }
+      """
+    Then the response field "status" is "ok"
+    And the response field "tail_start_index" is 4
+
+  # Prevents: callers misreading "no tail" — when everything was
+  # summarised the index must be an explicit null, not 0 or missing.
+  Scenario: tail_turns zero summarises everything and reports a null tail
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message "only message"
+    And an assistant message "only reply"
+    When I compact the history with model "big" and options:
+      """
+      { "tail_turns": 0 }
+      """
+    Then the response field "status" is "ok"
+    And the response field "tail_start_index" is null
+
+  # Prevents: convergence breaking — with a prior summary the
+  # summariser must be anchored on it (update-in-place), not start
+  # over, and the response must say the anchor was used.
+  Scenario: a previous summary anchors the summariser
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big" and options:
+      """
+      { "previous_summary": "## Goal\n- the prior anchor" }
+      """
+    Then the response field "status" is "ok"
+    And the response field "used_prior_summary" is true
+    And the summariser system prompt contains "Update the anchored summary"
+    And the summariser system prompt contains "the prior anchor"
+
+  # Prevents: a first compaction accidentally running in update mode
+  # against a phantom summary.
+  Scenario: a first compaction creates a fresh summary
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big"
+    Then the response field "status" is "ok"
+    And the summariser system prompt contains "Create a new anchored summary"
+    And the summariser system prompt does not contain "<previous-summary>"
+
+  # Prevents: "no router" looking like success — the spec maps a
+  # missing summariser to overflow so callers treat compaction as
+  # unavailable, and the lease must not stay stuck.
+  Scenario: an unavailable summariser reports overflow and frees the lease
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And the summariser is unavailable
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big"
+    Then the call succeeds
+    And the response field "status" is "overflow"
+    And no lease claim remains
+
+  # Prevents: provider failures mid-summarisation leaking as errors or
+  # stuck leases.
+  Scenario: a failing summariser reports overflow and frees the lease
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And the summariser fails with "429 rate limited"
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big"
+    Then the response field "status" is "overflow"
+    And no lease claim remains
+
+  # Prevents: persisting an empty string as the conversation's memory.
+  Scenario: an empty summary is reported as empty, never persisted as ok
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And the summariser returns an empty summary
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big"
+    Then the response field "status" is "empty"
+    And no lease claim remains
+
+  # Prevents: base64 image payloads being shipped to the summariser —
+  # they cost real tokens and carry nothing summarisable.
+  Scenario: images are stripped before the summariser sees the head
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message containing an image of 400 chars
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big"
+    Then the response field "status" is "ok"
+    And the summariser user prompt contains "[image stripped]"
+    And the summariser user prompt does not contain "AAAA"
+
+  # Prevents: a single huge tool output dominating the summariser's
+  # input — function outputs are truncated to max_output_chars first.
+  Scenario: verbose function outputs are truncated before summarisation
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message "start"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~5000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big"
+    Then the response field "status" is "ok"
+    And the summariser user prompt contains "... [truncated]"
+
+  # Prevents: the summariser running on a surprise model — it must run
+  # on exactly the model the caller passed.
+  Scenario: the summariser runs on the request model
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I compact the history with model "big"
+    Then the response field "status" is "ok"
+    And the summariser ran on model "big"
diff --git a/context-manager/tests/features/compact_lease.feature b/context-manager/tests/features/compact_lease.feature
new file mode 100644
index 00000000..ada902eb
--- /dev/null
+++ b/context-manager/tests/features/compact_lease.feature
@@ -0,0 +1,77 @@
+@pure
+Feature: compaction leases — one summariser per logical session
+
+  Contract (context-manager.md § State): scope `context_lease`, key =
+  options.lease_key (e.g. a session id) or a hash of the message set;
+  value {nonce, ts}; TTL ~300s. Two callers must not summarise the same
+  logical session concurrently; crashed holders expire via TTL; a
+  holder must never clear someone else's claim.
+
+  Background:
+    Given the router knows model "big" with context window 200000 and max output 8000
+    And a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+
+  # Prevents: two harness instances double-summarising one session —
+  # the second caller must see busy and must NOT consume an LLM call.
+  Scenario: a held lease turns the second caller away as busy
+    Given a foreign lease on the history's default key taken 0 ms ago
+    When I compact the history with model "big"
+    Then the call succeeds
+    And the response field "status" is "busy"
+    And the summariser was never invoked
+    And the lease on the history's default key still belongs to the foreign holder
+
+  # Prevents: a crashed holder locking a session forever — claims older
+  # than the TTL must read as free.
+  Scenario: an expired lease is taken over
+    Given a foreign lease on the history's default key taken 0 ms ago
+    And the clock advances by 300001 ms
+    When I compact the history with model "big"
+    Then the response field "status" is "ok"
+    And no lease claim remains
+
+  # Prevents: a lease one millisecond short of the TTL being treated as
+  # expired — early takeover is the double-summarisation bug again.
+  Scenario: a lease just inside the TTL still wins
+    Given a foreign lease on the history's default key taken 0 ms ago
+    And the clock advances by 299999 ms
+    When I compact the history with model "big"
+    Then the response field "status" is "busy"
+
+  # Prevents: callers with an explicit session key colliding with (or
+  # missing) each other — the same lease_key must contend.
+  Scenario: an explicit lease_key contends with its holder
+    Given a foreign lease on key "session-42" taken 0 ms ago
+    When I compact the history with model "big" and options:
+      """
+      { "lease_key": "session-42" }
+      """
+    Then the response field "status" is "busy"
+    And the lease on key "session-42" still belongs to "foreign-holder"
+
+  # Prevents: unrelated histories blocking each other — the default
+  # key is a hash of THIS message set, so a foreign claim on some other
+  # session must not contend.
+  Scenario: different histories never contend by default
+    Given a foreign lease on key "some-other-session" taken 0 ms ago
+    When I compact the history with model "big"
+    Then the response field "status" is "ok"
+    And the lease on key "some-other-session" still belongs to "foreign-holder"
+
+  # Prevents: a successful compaction leaving its claim behind and
+  # blocking every later compaction until TTL.
+  Scenario: success releases the lease
+    When I compact the history with model "big"
+    Then the response field "status" is "ok"
+    And no lease claim remains
+
+  # Prevents: a state outage letting every contender through at once —
+  # an unverifiable claim must read as busy, never as a win.
+  Scenario: a state outage reads as busy, never as a win
+    Given the lease store is unavailable
+    When I compact the history with model "big"
+    Then the call succeeds
+    And the response field "status" is "busy"
+    And the summariser was never invoked
diff --git a/context-manager/tests/features/count_tokens.feature b/context-manager/tests/features/count_tokens.feature
new file mode 100644
index 00000000..49121c21
--- /dev/null
+++ b/context-manager/tests/features/count_tokens.feature
@@ -0,0 +1,78 @@
+@pure
+Feature: context::count-tokens — estimate token usage for a message set
+
+  Contract (context-manager.md § context::count-tokens): estimate token
+  usage for messages, optionally including invocation schemas and a
+  system prompt, vs a model. The model selects a tokenizer; v1 always
+  falls back to the generic chars/4 heuristic and says so in
+  `estimator`. Pure and router-free — safe for cost-sensitive callers
+  with no llm-router installed.
+
+  # Prevents: silent estimator swaps — callers budgeting real money on
+  # these numbers must always be able to see which estimator ran.
+  Scenario: every count declares the estimator that produced it
+    Given a user message "hello"
+    When I count tokens with model "any-model"
+    Then the call succeeds
+    And the response field "estimator" is "heuristic"
+    And the router was never queried
+
+  # Prevents: the total drifting away from the by_role breakdown —
+  # callers subtract buckets (e.g. custom) from the total, so the two
+  # views must reconcile exactly.
+  Scenario: the total equals the sum of the by_role buckets
+    Given a user message "what is the weather"
+    And an assistant message "sunny with a chance of tokens"
+    And a function result for call "c1" from "weather::get" saying "22 degrees"
+    And a custom message of type "ui_marker"
+    When I count tokens with model "any-model"
+    Then the call succeeds
+    And the response field "tokens" equals the sum of the "by_role" fields
+    And the response field "by_role.user" exceeds 0
+    And the response field "by_role.assistant" exceeds 0
+    And the response field "by_role.function_result" exceeds 0
+    And the response field "by_role.custom" exceeds 0
+
+  # Prevents: count_tokens silently dropping custom messages the way
+  # assemble does — this function counts what it is GIVEN; the by_role
+  # split is how callers see (and can subtract) the custom share.
+  Scenario: custom messages are counted here, unlike in assemble
+    Given a custom message of type "compaction" of ~500 tokens
+    When I count tokens with model "any-model"
+    Then the call succeeds
+    And the response field "by_role.custom" exceeds 499
+    And the response field "tokens" exceeds 499
+
+  # Prevents: the system prompt being forgotten in the estimate — it
+  # occupies real context window and a 0-message count must reflect it.
+  Scenario: the system prompt is part of the total but no role bucket
+    Given an empty history
+    When I count tokens with model "any-model" and system prompt "xxxxxxxxxxxxxxxx"
+    Then the call succeeds
+    And the response field "tokens" is 4
+    And the response field "by_role.user" is 0
+    And the response field "by_role.assistant" is 0
+
+  # Prevents: invocation schemas (the agent_trigger entry) being
+  # invisible to the budget — tool schemas can be thousands of tokens.
+  Scenario: tools add to the total without polluting role buckets
+    Given an empty history
+    When I count tokens with model "any-model" and tools:
+      """
+      [{ "name": "agent_trigger", "description": "Invoke any allowed iii function.",
+         "parameters": { "type": "object", "properties": { "function": { "type": "string" } } } }]
+      """
+    Then the call succeeds
+    And the response field "tokens" exceeds 0
+    And the response field "by_role.user" is 0
+    And the response field "by_role.assistant" is 0
+    And the response field "by_role.function_result" is 0
+    And the response field "by_role.custom" is 0
+
+  # Prevents: an empty history erroring or returning phantom tokens.
+  Scenario: an empty message set counts to zero
+    Given an empty history
+    When I count tokens with model "any-model"
+    Then the call succeeds
+    And the response field "tokens" is 0
+    And the response field "by_role.user" is 0
diff --git a/context-manager/tests/features/engine_roundtrip.feature b/context-manager/tests/features/engine_roundtrip.feature
new file mode 100644
index 00000000..fc3b2efa
--- /dev/null
+++ b/context-manager/tests/features/engine_roundtrip.feature
@@ -0,0 +1,108 @@
+@engine
+Feature: engine round trip — the production surface over a live bus
+
+  The four functions registered in-process with PRODUCTION adapters
+  (router resolver, router summariser, state-backed leases) against a
+  real engine. llm-router is absent on the test engine — which is
+  exactly the degraded mode the spec defines: limits fall back, prune
+  and count stay fully functional, compact reports overflow after a
+  real state::* lease acquire/release cycle. Scenarios soft-skip when
+  no engine is reachable.
+
+  Background:
+    Given an engine connection
+
+  # Prevents: the registered wire surface drifting from the handlers —
+  # a real trigger must reach the production count path end to end.
+  Scenario: count_tokens answers over the bus
+    Given a user message "hello engine"
+    When I call "context::count-tokens" on the engine with:
+      """
+      { "messages": ${messages}, "model": { "id": "any-model" } }
+      """
+    Then the call succeeds
+    And the response field "estimator" is "heuristic"
+    And the response field "tokens" exceeds 0
+
+  # Prevents: prune needing anything beyond the bus — no router, no
+  # state, just the transform.
+  Scenario: prune rewrites outputs over the bus
+    Given a user message "investigate"
+    And an assistant function call "e1" to "shell::run"
+    And a function result for call "e1" from "shell::run" of ~2000 tokens
+    And a user message "next"
+    And a user message "done"
+    When I call "context::prune" on the engine with:
+      """
+      { "messages": ${messages},
+        "options": { "protect_recent_tokens": 100, "min_free_tokens": 1, "max_output_chars": 100 } }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 1
+    And response message 2 text is "[output pruned: was ~2000 tokens]"
+
+  # Prevents: assemble requiring llm-router for the inline-limits path
+  # — the standalone mode must work on a bare engine.
+  Scenario: assemble passes through under budget with inline limits
+    Given a user message "hello"
+    When I call "context::assemble" on the engine with:
+      """
+      { "messages": ${messages},
+        "model": { "id": "standalone",
+                   "limits": { "context_window": 10000, "max_output_tokens": 1000 } },
+        "system_prompt": "You are helpful." }
+      """
+    Then the call succeeds
+    And the response field "model_resolved" is "inline"
+    And the response field "usable" is 8000
+    And the response field "applied.compacted" is false
+    And the response field "system_prompt" is "You are helpful."
+
+  # Prevents: a missing router erroring the hot path — an over-budget
+  # assemble must still answer (best effort), exercising a REAL
+  # state::* lease acquire/release around the failed summarisation.
+  Scenario: over budget without a router degrades to best effort
+    Given a user message of ~3000 tokens
+    And a user message of ~3000 tokens
+    And a user message "recent"
+    When I call "context::assemble" on the engine with:
+      """
+      { "messages": ${messages},
+        "model": { "id": "small",
+                   "limits": { "context_window": 5000, "max_output_tokens": 500 } } }
+      """
+    Then the call succeeds
+    And the response field "applied.compacted" is false
+    And the response field "token_count" exceeds 4000
+
+  # Prevents: "router absent" surfacing as anything but the documented
+  # overflow status — and proves the real lease cycle completes (a
+  # second compact is NOT busy).
+  Scenario: compact without a router reports overflow and frees its lease
+    Given a user message of ~9000 tokens
+    And a user message "q2"
+    And a user message "q3"
+    When I call "context::compact" on the engine with:
+      """
+      { "messages": ${messages},
+        "model": { "id": "big",
+                   "limits": { "context_window": 200000, "max_output_tokens": 8000 } } }
+      """
+    Then the call succeeds
+    And the response field "status" is "overflow"
+    When I call "context::compact" on the engine with:
+      """
+      { "messages": ${messages},
+        "model": { "id": "big",
+                   "limits": { "context_window": 200000, "max_output_tokens": 8000 } } }
+      """
+    Then the response field "status" is "overflow"
+
+  # Prevents: handler errors losing their stable code across the bus.
+  Scenario: the spec error string survives the wire
+    When I call "context::assemble" on the engine with:
+      """
+      { "model": { "id": "m1" } }
+      """
+    Then the error contains "context/invalid_request"
+    And the error contains "messages is required"
diff --git a/context-manager/tests/features/errors.feature b/context-manager/tests/features/errors.feature
new file mode 100644
index 00000000..e4e97419
--- /dev/null
+++ b/context-manager/tests/features/errors.feature
@@ -0,0 +1,93 @@
+@pure
+Feature: error contract — spec strings, stable codes, tolerant unknowns
+
+  Contract (context-manager.md § context::assemble Errors, plus the
+  repo error convention `code: message`): `messages is required` and
+  `could not resolve model limits` are the spec's literal strings;
+  malformed shapes are rejected by deserialisation; unknown EXTRA
+  fields are tolerated so older callers survive additive evolution.
+
+  # Prevents: the documented error string drifting — callers match on
+  # "messages is required" across all four functions.
+  Scenario: assemble without messages fails with the spec string
+    When I call "context::assemble" with:
+      """
+      { "model": { "id": "m1" } }
+      """
+    Then the call fails with code "context/invalid_request"
+    And the error contains "messages is required"
+
+  Scenario: compact without messages fails with the spec string
+    When I call "context::compact" with:
+      """
+      { "model": { "id": "m1" } }
+      """
+    Then the call fails with code "context/invalid_request"
+    And the error contains "messages is required"
+
+  Scenario: prune without messages fails with the spec string
+    When I call "context::prune" with:
+      """
+      {}
+      """
+    Then the call fails with code "context/invalid_request"
+    And the error contains "messages is required"
+
+  Scenario: count_tokens without messages fails with the spec string
+    When I call "context::count-tokens" with:
+      """
+      { "model": { "id": "m1" } }
+      """
+    Then the call fails with code "context/invalid_request"
+    And the error contains "messages is required"
+
+  # Prevents: `messages: null` sneaking past as an empty history.
+  Scenario: null messages are as missing as absent messages
+    When I call "context::assemble" with:
+      """
+      { "messages": null, "model": { "id": "m1" } }
+      """
+    Then the call fails with code "context/invalid_request"
+    And the error contains "messages is required"
+
+  # Prevents: malformed roles silently coercing — a typo'd role must be
+  # rejected at the boundary, not guessed at.
+  Scenario: an unknown message role is rejected at deserialisation
+    When I call "context::count-tokens" with:
+      """
+      { "messages": [{ "role": "wizard", "content": [], "timestamp": 1 }],
+        "model": { "id": "m1" } }
+      """
+    Then the error contains "invalid request"
+
+  # Prevents: a missing model slipping through and counting against a
+  # phantom tokenizer.
+  Scenario: count_tokens requires a model
+    When I call "context::count-tokens" with:
+      """
+      { "messages": [] }
+      """
+    Then the error contains "invalid request"
+    And the error contains "model"
+
+  # Prevents: additive API evolution breaking old workers — unknown
+  # extra request fields must be ignored, not fatal.
+  Scenario: unknown extra fields are tolerated
+    Given the router knows model "big" with context window 200000 and max output 8000
+    When I call "context::assemble" with:
+      """
+      { "messages": [], "model": { "id": "big" }, "future_option": 42 }
+      """
+    Then the call succeeds
+    And the response field "model_resolved" is "router"
+
+  # Prevents: prune requiring a model it does not need — the heuristic
+  # is model-free and the spec marks `model` optional here.
+  Scenario: prune works without any model
+    Given a user message "hello"
+    When I call "context::prune" with:
+      """
+      { "messages": ${messages} }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 0
diff --git a/context-manager/tests/features/invariants.feature b/context-manager/tests/features/invariants.feature
new file mode 100644
index 00000000..4a18ccac
--- /dev/null
+++ b/context-manager/tests/features/invariants.feature
@@ -0,0 +1,104 @@
+@pure
+Feature: structural invariants — the returned context is always provider-legal
+
+  Contract (context-manager.md § Structural invariants): whatever
+  pruning or compaction does, a function_call and its function_result
+  always land on the same side of any boundary (providers reject
+  orphaned results); prune replaces, never removes; custom messages
+  never reach the model-facing list nor its token count.
+
+  # Prevents: the compaction boundary slicing between an assistant's
+  # call and its result — the only in-budget cut here straddles the
+  # pair, so selection must skip it and cut after the result instead.
+  Scenario: the tail never starts between a call and its result
+    Given inline model "small" with context window 5000 and max output 500
+    And the summariser returns "## Goal\n- boundary test"
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message "follow up"
+    And an assistant function call "c2" to "shell::run"
+    And a function result for call "c2" from "shell::run" of ~2100 tokens
+    And an assistant message "done"
+    When I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "applied.compacted" is true
+    And the response field "applied.tail_start_index" is 5
+    And the response messages start at request message 5
+    And call/result pairing is intact in the response messages
+    And the summariser user prompt contains "[tool_call] shell::run"
+
+  # Prevents: a turn that STARTS with an inline function_result block
+  # (Anthropic-style tool results travel in user messages) being used
+  # as a cut point — its call sits in the previous turn, so the cut
+  # must move back to the older, safe turn and keep the pair together.
+  Scenario: a turn starting with an inline result defers to an older safe cut
+    Given inline model "small" with context window 5000 and max output 500
+    And the summariser returns "## Goal\n- inline result test"
+    And a user message of ~4000 tokens
+    And an assistant message "r0"
+    And a user message "follow up"
+    And an assistant function call "c7" to "coder::search"
+    And a user message carrying an inline function result block for call "c7"
+    And an assistant message "done"
+    When I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "applied.compacted" is true
+    And the response field "applied.tail_start_index" is 2
+    And the response messages start at request message 2
+    And call/result pairing is intact in the response messages
+
+  # Prevents: custom messages leaking to providers (they have no wire
+  # mapping) or their tokens triggering phantom overflows — a huge
+  # custom entry must neither appear nor count.
+  Scenario: custom messages neither reach the model nor count tokens
+    Given inline model "small" with context window 5000 and max output 500
+    And a custom message of type "compaction" of ~10000 tokens
+    And a user message "hi"
+    When I assemble the history with model "small"
+    Then the call succeeds
+    And the response field "token_count" does not exceed 100
+    And the response field "applied.pruned" is false
+    And the response field "applied.compacted" is false
+    And the response messages contain no custom roles
+    And the response field "messages" has 1 items
+    And the summariser was never invoked
+
+  # Prevents: tail_start_index drifting when custom messages sit before
+  # the cut — the index must point into the REQUEST array (customs
+  # included), not the filtered view, or callers persist the wrong
+  # boundary.
+  Scenario: tail_start_index accounts for excluded custom messages
+    Given inline model "small" with context window 5000 and max output 500
+    And the summariser returns "## Goal\n- index mapping"
+    And a custom message of type "ui_marker"
+    And a user message of ~3000 tokens
+    And an assistant message "r0"
+    And a user message of ~3000 tokens
+    And an assistant message "r1"
+    And a user message "recent question"
+    And an assistant message "recent answer"
+    When I assemble the history with model "small"
+    Then the response field "applied.compacted" is true
+    And the response field "applied.tail_start_index" is 4
+    And the response messages contain no custom roles
+
+  # Prevents: prune deleting blocks or messages under pressure — the
+  # message count, the placeholder, and the call linkage must all
+  # survive a maximally aggressive prune.
+  Scenario: an aggressive prune still replaces rather than removes
+    Given a user message "task"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~5000 tokens
+    And an assistant function call "c2" to "coder::read"
+    And a function result for call "c2" from "coder::read" of ~5000 tokens
+    And a user message "next"
+    And a user message "done"
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 0, "min_free_tokens": 1, "max_output_chars": 0 }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 2
+    And the response messages have as many messages as the request
+    And every response message keeps its function_call_id
+    And call/result pairing is intact in the response messages
diff --git a/context-manager/tests/features/model_and_budget.feature b/context-manager/tests/features/model_and_budget.feature
new file mode 100644
index 00000000..8f6de07f
--- /dev/null
+++ b/context-manager/tests/features/model_and_budget.feature
@@ -0,0 +1,116 @@
+@pure
+Feature: model resolution and the usable token budget
+
+  Contract (context-manager.md § Model input, § Token budget model):
+  limits resolve inline -> router::models::get -> conservative fallback
+  (8192/1024), echoed in `model_resolved`. The usable budget is
+    usable = max(0, (input_limit ?? context_window - max_output_tokens)
+                    - reserved - thinking_budget)
+  with reserved defaulting to min(20000, 10% of context_window).
+  assemble is the probe here because it returns both `usable` and
+  `model_resolved`.
+
+  Background:
+    Given an empty history
+
+  # Prevents: the worker phoning the router when the caller supplied
+  # limits — inline limits are the standalone mode; a router lookup
+  # would break deployments that don't install llm-router.
+  Scenario: inline limits resolve without touching the router
+    Given inline model "standalone" with context window 10000 and max output 1000
+    When I assemble the history with model "standalone"
+    Then the call succeeds
+    And the response field "model_resolved" is "inline"
+    And the response field "usable" is 8000
+    And the router was never queried
+
+  # Prevents: ignoring a declared input_limit and overfilling models
+  # whose usable input is smaller than context - output.
+  Scenario: input_limit takes precedence over the derived base
+    Given inline model "capped" with input limit 5000, context window 10000 and max output 1000
+    When I assemble the history with model "capped"
+    Then the response field "usable" is 4000
+
+  # Prevents: the reserve being un-overridable — latency-sensitive
+  # callers tune it per call.
+  Scenario: reserved_tokens overrides the default reserve
+    Given inline model "standalone" with context window 10000 and max output 1000
+    When I assemble the history with model "standalone" and options:
+      """
+      { "reserved_tokens": 0 }
+      """
+    Then the response field "usable" is 9000
+
+  # Prevents: a flat 20k reserve eating a third of a small model's
+  # window — the reserve must scale down with the context window.
+  Scenario: the default reserve is 10% capped at 20000
+    Given the router knows model "small" with context window 32000 and max output 16000
+    When I assemble the history with model "small"
+    Then the response field "model_resolved" is "router"
+    And the response field "usable" is 12800
+    And the router was asked for model "small"
+
+  # Prevents: silent fallback — a caller must be able to detect that
+  # conservative limits were used instead of real catalog data.
+  Scenario: an unknown model falls back conservatively and says so
+    When I assemble the history with model "ghost-model"
+    Then the call succeeds
+    And the response field "model_resolved" is "fallback"
+    And the response field "usable" is 6349
+
+  # Prevents: treating a dead router differently from a missing model —
+  # both degrade to the same detectable fallback.
+  Scenario: a dead router also falls back conservatively
+    Given the router is unavailable
+    When I assemble the history with model "any-model"
+    Then the call succeeds
+    And the response field "model_resolved" is "fallback"
+
+  # Prevents: deployments that disabled the fallback silently running
+  # on 8k limits anyway — they asked for a hard error, give them one.
+  Scenario: fallback disabled turns an unresolvable model into an error
+    Given config allow_fallback_limits is false
+    When I assemble the history with model "ghost-model"
+    Then the call fails with code "context/model_unresolved"
+    And the error contains "could not resolve model limits"
+
+  # Prevents: thinking tokens overflowing the context — when a tier is
+  # requested and the model declares a budget, that budget must be
+  # carved out of usable.
+  Scenario: a requested thinking tier reserves its declared budget
+    Given the router knows model "thinker" with context window 100000 and max output 10000
+    And the router model "thinker" declares a thinking budget of 16000 for "high"
+    When I assemble the history with model "thinker" and options:
+      """
+      { "thinking_level": "high" }
+      """
+    Then the response field "usable" is 64000
+
+  # Prevents: reserving thinking room the model never declared — an
+  # undeclared tier must cost nothing.
+  Scenario: an undeclared thinking tier reserves nothing
+    Given the router knows model "thinker" with context window 100000 and max output 10000
+    And the router model "thinker" declares a thinking budget of 16000 for "high"
+    When I assemble the history with model "thinker" and options:
+      """
+      { "thinking_level": "low" }
+      """
+    Then the response field "usable" is 80000
+
+  # Prevents: inline-limit callers paying a thinking tax — inline
+  # limits carry no budgets, so the tier must be a no-op.
+  Scenario: inline limits ignore the thinking level entirely
+    Given inline model "standalone" with context window 10000 and max output 1000
+    When I assemble the history with model "standalone" and options:
+      """
+      { "thinking_level": "xhigh" }
+      """
+    Then the response field "usable" is 8000
+
+  # Prevents: integer underflow wrapping a tiny budget around to ~2^64
+  # — the budget must clamp to 0, never wrap.
+  Scenario: a model smaller than its own output budget clamps to zero
+    Given inline model "tiny" with context window 1000 and max output 2000
+    When I assemble the history with model "tiny"
+    Then the call succeeds
+    And the response field "usable" is 0
diff --git a/context-manager/tests/features/prune.feature b/context-manager/tests/features/prune.feature
new file mode 100644
index 00000000..019cfe24
--- /dev/null
+++ b/context-manager/tests/features/prune.feature
@@ -0,0 +1,137 @@
+@pure
+Feature: context::prune — placeholder verbose function outputs
+
+  Contract (context-manager.md § context::prune, § Structural
+  invariants): walk function_result content newest to oldest, freeing
+  outputs outside a protected token window. Prune REPLACES, never
+  removes — the block, the message, and the function_call_id linkage
+  all survive; placeholders carry the freed size
+  (`[output pruned: was ~N tokens]`). The most recent two user turns
+  are always exempt (prior-art guard, independent of the window).
+
+  Background:
+    Given a user message "investigate the failure"
+    And an assistant function call "c1" to "shell::run"
+    And a function result for call "c1" from "shell::run" of ~2000 tokens
+    And an assistant message "found it"
+    And a user message "now fix it"
+    And a user message "and add a test"
+
+  # Prevents: deleting messages to save tokens — providers reject a
+  # function_call whose result vanished, so the message count and the
+  # call linkage must survive pruning.
+  Scenario: pruning replaces content but never removes a message
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 100, "min_free_tokens": 1, "max_output_chars": 100 }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 1
+    And the response field "pruned_tokens" is 1992
+    And the response messages have as many messages as the request
+    And every response message keeps its function_call_id
+    And call/result pairing is intact in the response messages
+
+  # Prevents: vague placeholders — the spec's placeholder must tell the
+  # model (and a debugging human) how much was cut.
+  Scenario: the placeholder names the freed size
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 100, "min_free_tokens": 1, "max_output_chars": 100 }
+      """
+    Then response message 2 text is "[output pruned: was ~2000 tokens]"
+
+  # Prevents: destroying context for a marginal win — freeing less than
+  # min_free_tokens must leave the history completely untouched.
+  Scenario: a pass that would free too little is a complete no-op
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 100, "min_free_tokens": 5000, "max_output_chars": 100 }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 0
+    And the response field "pruned_tokens" is 0
+    And the response field "scanned_parts" is 1
+    And response message 2 text has 8000 chars
+
+  # Prevents: pruning the outputs a deployment declared load-bearing
+  # (e.g. a plan document the model must keep verbatim).
+  Scenario: protected functions are never pruned
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 0, "min_free_tokens": 1, "max_output_chars": 100,
+        "protected_functions": ["shell::run"] }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 0
+    And the response field "scanned_parts" is 0
+    And response message 2 text has 8000 chars
+
+  # Prevents: the protected window filling oldest-first — the NEWEST
+  # outputs are the ones the model still needs, so they must consume
+  # the window and push older outputs out of it.
+  Scenario: the newest outputs fill the protected window first
+    Given an assistant function call "c2" to "coder::read"
+    And a function result for call "c2" from "coder::read" of ~90 tokens
+    And a user message "thanks"
+    And a user message "continue"
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 100, "min_free_tokens": 1, "max_output_chars": 100 }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 1
+    And response message 7 text has 360 chars
+    And response message 2 text is "[output pruned: was ~2000 tokens]"
+
+  # Prevents: pruning cheap outputs for cosmetic gains — an output at
+  # or under max_output_chars is not "verbose" and stays, even outside
+  # the protected window.
+  Scenario: small outputs are not verbose and survive
+    Given an empty history
+    And a user message "start"
+    And an assistant function call "c9" to "config::get"
+    And a function result for call "c9" from "config::get" saying "value=42"
+    And a user message "ok"
+    And a user message "done"
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 0, "min_free_tokens": 1, "max_output_chars": 100 }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 0
+    And the response field "scanned_parts" is 1
+    And response message 2 text is "value=42"
+
+  # Prevents: the always-exempt recent turns being pruned under token
+  # pressure — outputs the user is actively working with must survive
+  # even a zero-token protected window.
+  Scenario: outputs inside the last two user turns are untouchable
+    Given an empty history
+    And a user message "look at this"
+    And a user message "do the thing"
+    And an assistant function call "c5" to "shell::run"
+    And a function result for call "c5" from "shell::run" of ~3000 tokens
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 0, "min_free_tokens": 1, "max_output_chars": 100 }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 0
+    And the response field "scanned_parts" is 0
+    And response message 3 text has 12000 chars
+
+  # Prevents: re-running prune from double-pruning placeholders or
+  # re-counting freed tokens — the pass must be idempotent.
+  Scenario: pruning twice is idempotent
+    When I prune the history with options:
+      """
+      { "protect_recent_tokens": 100, "min_free_tokens": 1, "max_output_chars": 100 }
+      """
+    And I prune the response messages again with the same options:
+      """
+      { "protect_recent_tokens": 100, "min_free_tokens": 1, "max_output_chars": 100 }
+      """
+    Then the call succeeds
+    And the response field "pruned_parts" is 0
+    And the response field "pruned_tokens" is 0
diff --git a/context-manager/tests/golden/schemas/context.assemble.json b/context-manager/tests/golden/schemas/context.assemble.json
new file mode 100644
index 00000000..ce66256d
--- /dev/null
+++ b/context-manager/tests/golden/schemas/context.assemble.json
@@ -0,0 +1,1107 @@
+{
+  "description": "Build the model-ready context (system prompt + budgeted messages) from a history; applies prune and/or compaction as needed to fit the model's usable token budget.",
+  "function_id": "context::assemble",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "AgentMessage": {
+        "description": "The canonical transcript message union, discriminated by `role`.",
+        "oneOf": [
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "role": {
+                "enum": [
+                  "user"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "error_kind": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/ErrorKind"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "error_message": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "model": {
+                "type": "string"
+              },
+              "native_stop_reason": {
+                "description": "Provider's raw finish reason, passed through untouched.",
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "provider": {
+                "type": "string"
+              },
+              "role": {
+                "enum": [
+                  "assistant"
+                ],
+                "type": "string"
+              },
+              "stop_reason": {
+                "$ref": "#/definitions/StopReason"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              },
+              "usage": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/Usage"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "warnings": {
+                "description": "Report-and-continue notices (e.g. dropped unsupported param).",
+                "items": {
+                  "type": "string"
+                },
+                "type": [
+                  "array",
+                  "null"
+                ]
+              }
+            },
+            "required": [
+              "content",
+              "model",
+              "provider",
+              "role",
+              "stop_reason",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "details": {
+                "default": null,
+                "description": "Opaque structured payload kept alongside the rendered content."
+              },
+              "function_call_id": {
+                "description": "Echoes the `function_call` block id this result answers.",
+                "type": "string"
+              },
+              "function_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "default": false,
+                "type": "boolean"
+              },
+              "role": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "function_id",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "description": "Escape hatch for app-specific transcript items (system notices, UI markers, attachments, ...). No provider wire mapping — `context::assemble` excludes these from the model-facing list.",
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "custom_type": {
+                "description": "App-defined discriminator.",
+                "type": "string"
+              },
+              "details": true,
+              "display": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "role": {
+                "enum": [
+                  "custom"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "custom_type",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "AssembleOptions": {
+        "properties": {
+          "allow_compaction": {
+            "default": null,
+            "description": "Default true.",
+            "type": [
+              "boolean",
+              "null"
+            ]
+          },
+          "allow_prune": {
+            "default": null,
+            "description": "Default true.",
+            "type": [
+              "boolean",
+              "null"
+            ]
+          },
+          "lease_key": {
+            "default": null,
+            "description": "Compaction mutual-exclusion key (e.g. a session id); default: hash of the message set.",
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "previous_summary": {
+            "default": null,
+            "description": "Persisted summary from a prior compaction (see the spec's \"The compaction round trip\"); rendered into the system prompt and used as the anchor if compaction triggers again.",
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "protected_functions": {
+            "default": null,
+            "description": "`function_id`s whose outputs are never pruned.",
+            "items": {
+              "type": "string"
+            },
+            "type": [
+              "array",
+              "null"
+            ]
+          },
+          "reserved_tokens": {
+            "default": null,
+            "description": "Override the default reserve (`min(20000, 10% of context_window)`).",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "tail_turns": {
+            "default": null,
+            "description": "user+assistant pairs always kept verbatim (default 2).",
+            "format": "uint",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "thinking_level": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ThinkingLevel"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "default": null,
+            "description": "Reserve the model's thinking budget for this tier."
+          }
+        },
+        "type": "object"
+      },
+      "ContentBlock": {
+        "description": "The atomic unit of message content. A message's `content` is an ordered array of these.",
+        "oneOf": [
+          {
+            "properties": {
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "data": {
+                "description": "Base64-encoded image bytes.",
+                "type": "string"
+              },
+              "mime": {
+                "description": "MIME type, e.g. `image/png`.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "image"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "data",
+              "mime",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "signature": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "thinking"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "arguments": {
+                "default": null,
+                "description": "Model-produced arguments (JSON)."
+              },
+              "function_id": {
+                "description": "The iii function id to invoke.",
+                "type": "string"
+              },
+              "id": {
+                "description": "Unique per call, echoed by the result.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "function_call"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "function_id",
+              "id",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "function_call_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "type": [
+                  "boolean",
+                  "null"
+                ]
+              },
+              "type": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ErrorKind": {
+        "description": "Coarse error classification carried by failed assistant messages.",
+        "enum": [
+          "auth_expired",
+          "rate_limited",
+          "context_overflow",
+          "transient",
+          "permanent"
+        ],
+        "type": "string"
+      },
+      "ModelInput": {
+        "description": "How callers identify the target model (context-manager.md § Model input). Inline `limits` keep the call standalone; `id`/`provider` alone defer limit resolution to `router::models::get`.",
+        "properties": {
+          "id": {
+            "description": "Model id, e.g. `claude-sonnet-4`.",
+            "type": "string"
+          },
+          "limits": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ModelLimits"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "description": "Inline token limits; when present no router lookup happens."
+          },
+          "provider": {
+            "description": "Provider id, e.g. `anthropic`; disambiguates when a model id exists on multiple providers.",
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "id"
+        ],
+        "type": "object"
+      },
+      "ModelLimits": {
+        "description": "Token limits a caller may supply inline to stay fully standalone.",
+        "properties": {
+          "context_window": {
+            "description": "Total context window in tokens.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_limit": {
+            "description": "Usable input budget when distinct from `context_window`.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_output_tokens": {
+            "description": "Maximum output tokens the model can produce per call.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "context_window",
+          "max_output_tokens"
+        ],
+        "type": "object"
+      },
+      "StopReason": {
+        "description": "Why an assistant message stopped generating.",
+        "enum": [
+          "end",
+          "length",
+          "function_call",
+          "aborted",
+          "error"
+        ],
+        "type": "string"
+      },
+      "ThinkingLevel": {
+        "description": "Reasoning-effort tier; levels map to provider-native knobs via `Model.thinking_budgets`.",
+        "enum": [
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh"
+        ],
+        "type": "string"
+      },
+      "Usage": {
+        "description": "Token / cost accounting reported by providers.",
+        "properties": {
+          "cache_read": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "input": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      }
+    },
+    "properties": {
+      "messages": {
+        "description": "Full candidate history, oldest first.",
+        "items": {
+          "$ref": "#/definitions/AgentMessage"
+        },
+        "type": [
+          "array",
+          "null"
+        ]
+      },
+      "model": {
+        "$ref": "#/definitions/ModelInput"
+      },
+      "options": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/AssembleOptions"
+          },
+          {
+            "type": "null"
+          }
+        ]
+      },
+      "system_prompt": {
+        "default": null,
+        "description": "Base system prompt to prepend/merge.",
+        "type": [
+          "string",
+          "null"
+        ]
+      }
+    },
+    "required": [
+      "model"
+    ],
+    "title": "AssembleRequest",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "AgentMessage": {
+        "description": "The canonical transcript message union, discriminated by `role`.",
+        "oneOf": [
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "role": {
+                "enum": [
+                  "user"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "error_kind": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/ErrorKind"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "error_message": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "model": {
+                "type": "string"
+              },
+              "native_stop_reason": {
+                "description": "Provider's raw finish reason, passed through untouched.",
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "provider": {
+                "type": "string"
+              },
+              "role": {
+                "enum": [
+                  "assistant"
+                ],
+                "type": "string"
+              },
+              "stop_reason": {
+                "$ref": "#/definitions/StopReason"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              },
+              "usage": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/Usage"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "warnings": {
+                "description": "Report-and-continue notices (e.g. dropped unsupported param).",
+                "items": {
+                  "type": "string"
+                },
+                "type": [
+                  "array",
+                  "null"
+                ]
+              }
+            },
+            "required": [
+              "content",
+              "model",
+              "provider",
+              "role",
+              "stop_reason",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "details": {
+                "default": null,
+                "description": "Opaque structured payload kept alongside the rendered content."
+              },
+              "function_call_id": {
+                "description": "Echoes the `function_call` block id this result answers.",
+                "type": "string"
+              },
+              "function_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "default": false,
+                "type": "boolean"
+              },
+              "role": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "function_id",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "description": "Escape hatch for app-specific transcript items (system notices, UI markers, attachments, ...). No provider wire mapping — `context::assemble` excludes these from the model-facing list.",
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "custom_type": {
+                "description": "App-defined discriminator.",
+                "type": "string"
+              },
+              "details": true,
+              "display": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "role": {
+                "enum": [
+                  "custom"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "custom_type",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "Applied": {
+        "description": "What the pipeline actually did this call.",
+        "properties": {
+          "compacted": {
+            "type": "boolean"
+          },
+          "pruned": {
+            "type": "boolean"
+          },
+          "pruned_tokens": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "summary": {
+            "description": "Present when compacted; the caller should persist it and pass it back as `options.previous_summary` (compaction round trip).",
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "tail_start_index": {
+            "description": "Present when compacted: index into the request messages where the verbatim tail begins (`null` = everything was summarised).",
+            "format": "uint",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "tokens_before": {
+            "description": "Present when compacted: estimated tokens of the summarised head.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "compacted",
+          "pruned",
+          "pruned_tokens"
+        ],
+        "type": "object"
+      },
+      "ContentBlock": {
+        "description": "The atomic unit of message content. A message's `content` is an ordered array of these.",
+        "oneOf": [
+          {
+            "properties": {
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "data": {
+                "description": "Base64-encoded image bytes.",
+                "type": "string"
+              },
+              "mime": {
+                "description": "MIME type, e.g. `image/png`.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "image"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "data",
+              "mime",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "signature": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "thinking"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "arguments": {
+                "default": null,
+                "description": "Model-produced arguments (JSON)."
+              },
+              "function_id": {
+                "description": "The iii function id to invoke.",
+                "type": "string"
+              },
+              "id": {
+                "description": "Unique per call, echoed by the result.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "function_call"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "function_id",
+              "id",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "function_call_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "type": [
+                  "boolean",
+                  "null"
+                ]
+              },
+              "type": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ErrorKind": {
+        "description": "Coarse error classification carried by failed assistant messages.",
+        "enum": [
+          "auth_expired",
+          "rate_limited",
+          "context_overflow",
+          "transient",
+          "permanent"
+        ],
+        "type": "string"
+      },
+      "ModelResolvedWire": {
+        "description": "How the model limits were resolved.",
+        "enum": [
+          "inline",
+          "router",
+          "fallback"
+        ],
+        "type": "string"
+      },
+      "StopReason": {
+        "description": "Why an assistant message stopped generating.",
+        "enum": [
+          "end",
+          "length",
+          "function_call",
+          "aborted",
+          "error"
+        ],
+        "type": "string"
+      },
+      "Usage": {
+        "description": "Token / cost accounting reported by providers.",
+        "properties": {
+          "cache_read": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "input": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      }
+    },
+    "properties": {
+      "applied": {
+        "$ref": "#/definitions/Applied"
+      },
+      "messages": {
+        "description": "Budgeted, ready to send to llm-router.",
+        "items": {
+          "$ref": "#/definitions/AgentMessage"
+        },
+        "type": "array"
+      },
+      "model_resolved": {
+        "$ref": "#/definitions/ModelResolvedWire"
+      },
+      "system_prompt": {
+        "type": "string"
+      },
+      "token_count": {
+        "description": "Estimated tokens of the returned context (messages + system prompt). Can exceed `usable` when prune/compaction were disabled, unavailable, or insufficient — best effort, visible.",
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      },
+      "usable": {
+        "description": "The budget the context was fit into.",
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      }
+    },
+    "required": [
+      "applied",
+      "messages",
+      "model_resolved",
+      "system_prompt",
+      "token_count",
+      "usable"
+    ],
+    "title": "AssembleResponse",
+    "type": "object"
+  }
+}
diff --git a/context-manager/tests/golden/schemas/context.compact.json b/context-manager/tests/golden/schemas/context.compact.json
new file mode 100644
index 00000000..af0ff45b
--- /dev/null
+++ b/context-manager/tests/golden/schemas/context.compact.json
@@ -0,0 +1,643 @@
+{
+  "description": "Summarise the head of a history into a single compaction summary, keeping a recent tail verbatim. Transient: the caller persists the result.",
+  "function_id": "context::compact",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "AgentMessage": {
+        "description": "The canonical transcript message union, discriminated by `role`.",
+        "oneOf": [
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "role": {
+                "enum": [
+                  "user"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "error_kind": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/ErrorKind"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "error_message": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "model": {
+                "type": "string"
+              },
+              "native_stop_reason": {
+                "description": "Provider's raw finish reason, passed through untouched.",
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "provider": {
+                "type": "string"
+              },
+              "role": {
+                "enum": [
+                  "assistant"
+                ],
+                "type": "string"
+              },
+              "stop_reason": {
+                "$ref": "#/definitions/StopReason"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              },
+              "usage": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/Usage"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "warnings": {
+                "description": "Report-and-continue notices (e.g. dropped unsupported param).",
+                "items": {
+                  "type": "string"
+                },
+                "type": [
+                  "array",
+                  "null"
+                ]
+              }
+            },
+            "required": [
+              "content",
+              "model",
+              "provider",
+              "role",
+              "stop_reason",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "details": {
+                "default": null,
+                "description": "Opaque structured payload kept alongside the rendered content."
+              },
+              "function_call_id": {
+                "description": "Echoes the `function_call` block id this result answers.",
+                "type": "string"
+              },
+              "function_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "default": false,
+                "type": "boolean"
+              },
+              "role": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "function_id",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "description": "Escape hatch for app-specific transcript items (system notices, UI markers, attachments, ...). No provider wire mapping — `context::assemble` excludes these from the model-facing list.",
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "custom_type": {
+                "description": "App-defined discriminator.",
+                "type": "string"
+              },
+              "details": true,
+              "display": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "role": {
+                "enum": [
+                  "custom"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "custom_type",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "CompactOptions": {
+        "properties": {
+          "lease_key": {
+            "default": null,
+            "description": "Mutual-exclusion key (e.g. a session id); default: hash of the message set.",
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "preserve_recent_tokens": {
+            "default": null,
+            "description": "Override the adaptive verbatim-tail token budget.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "previous_summary": {
+            "default": null,
+            "description": "Anchor from a prior compaction so summaries converge instead of growing; the summariser updates it rather than starting over.",
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "tail_turns": {
+            "default": null,
+            "description": "user+assistant pairs kept verbatim (default 2).",
+            "format": "uint",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      },
+      "ContentBlock": {
+        "description": "The atomic unit of message content. A message's `content` is an ordered array of these.",
+        "oneOf": [
+          {
+            "properties": {
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "data": {
+                "description": "Base64-encoded image bytes.",
+                "type": "string"
+              },
+              "mime": {
+                "description": "MIME type, e.g. `image/png`.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "image"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "data",
+              "mime",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "signature": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "thinking"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "arguments": {
+                "default": null,
+                "description": "Model-produced arguments (JSON)."
+              },
+              "function_id": {
+                "description": "The iii function id to invoke.",
+                "type": "string"
+              },
+              "id": {
+                "description": "Unique per call, echoed by the result.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "function_call"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "function_id",
+              "id",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "function_call_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "type": [
+                  "boolean",
+                  "null"
+                ]
+              },
+              "type": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ErrorKind": {
+        "description": "Coarse error classification carried by failed assistant messages.",
+        "enum": [
+          "auth_expired",
+          "rate_limited",
+          "context_overflow",
+          "transient",
+          "permanent"
+        ],
+        "type": "string"
+      },
+      "ModelInput": {
+        "description": "How callers identify the target model (context-manager.md § Model input). Inline `limits` keep the call standalone; `id`/`provider` alone defer limit resolution to `router::models::get`.",
+        "properties": {
+          "id": {
+            "description": "Model id, e.g. `claude-sonnet-4`.",
+            "type": "string"
+          },
+          "limits": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ModelLimits"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "description": "Inline token limits; when present no router lookup happens."
+          },
+          "provider": {
+            "description": "Provider id, e.g. `anthropic`; disambiguates when a model id exists on multiple providers.",
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "id"
+        ],
+        "type": "object"
+      },
+      "ModelLimits": {
+        "description": "Token limits a caller may supply inline to stay fully standalone.",
+        "properties": {
+          "context_window": {
+            "description": "Total context window in tokens.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_limit": {
+            "description": "Usable input budget when distinct from `context_window`.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_output_tokens": {
+            "description": "Maximum output tokens the model can produce per call.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "context_window",
+          "max_output_tokens"
+        ],
+        "type": "object"
+      },
+      "StopReason": {
+        "description": "Why an assistant message stopped generating.",
+        "enum": [
+          "end",
+          "length",
+          "function_call",
+          "aborted",
+          "error"
+        ],
+        "type": "string"
+      },
+      "Usage": {
+        "description": "Token / cost accounting reported by providers.",
+        "properties": {
+          "cache_read": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "input": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      }
+    },
+    "properties": {
+      "messages": {
+        "description": "Full candidate history, oldest first.",
+        "items": {
+          "$ref": "#/definitions/AgentMessage"
+        },
+        "type": [
+          "array",
+          "null"
+        ]
+      },
+      "model": {
+        "$ref": "#/definitions/ModelInput"
+      },
+      "options": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/CompactOptions"
+          },
+          {
+            "type": "null"
+          }
+        ]
+      }
+    },
+    "required": [
+      "model"
+    ],
+    "title": "CompactRequest",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "description": "Discriminated on `status`.",
+    "oneOf": [
+      {
+        "description": "Compaction ran; the caller should persist `summary` and map `tail_start_index` onto its own storage ids.",
+        "properties": {
+          "status": {
+            "enum": [
+              "ok"
+            ],
+            "type": "string"
+          },
+          "summary": {
+            "type": "string"
+          },
+          "tail_start_index": {
+            "description": "Index into the request `messages` where the verbatim tail begins; `null` when everything was summarised.",
+            "format": "uint",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "tokens_after": {
+            "description": "Estimated tokens of summary + verbatim tail.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "tokens_before": {
+            "description": "Estimated tokens of the summarised head.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "used_prior_summary": {
+            "type": "boolean"
+          }
+        },
+        "required": [
+          "status",
+          "summary",
+          "tokens_after",
+          "tokens_before",
+          "used_prior_summary"
+        ],
+        "type": "object"
+      },
+      {
+        "description": "A compaction lease is held; the caller may retry.",
+        "properties": {
+          "status": {
+            "enum": [
+              "busy"
+            ],
+            "type": "string"
+          }
+        },
+        "required": [
+          "status"
+        ],
+        "type": "object"
+      },
+      {
+        "description": "Nothing to compact.",
+        "properties": {
+          "status": {
+            "enum": [
+              "empty"
+            ],
+            "type": "string"
+          }
+        },
+        "required": [
+          "status"
+        ],
+        "type": "object"
+      },
+      {
+        "description": "The summariser is unavailable or itself overflowed.",
+        "properties": {
+          "status": {
+            "enum": [
+              "overflow"
+            ],
+            "type": "string"
+          }
+        },
+        "required": [
+          "status"
+        ],
+        "type": "object"
+      }
+    ],
+    "title": "CompactResponse"
+  }
+}
diff --git a/context-manager/tests/golden/schemas/context.count-tokens.json b/context-manager/tests/golden/schemas/context.count-tokens.json
new file mode 100644
index 00000000..bd257343
--- /dev/null
+++ b/context-manager/tests/golden/schemas/context.count-tokens.json
@@ -0,0 +1,635 @@
+{
+  "description": "Estimate token usage for a set of messages (+ optional invocation schemas / system prompt) vs a model.",
+  "function_id": "context::count-tokens",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "AgentFunction": {
+        "description": "An invocation-surface schema entry (README.md § AgentFunction) — accepted by `context::count-tokens` so callers can include the `agent_trigger` schema in their estimate.",
+        "properties": {
+          "description": {
+            "type": "string"
+          },
+          "execution_mode": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ExecutionMode"
+              },
+              {
+                "type": "null"
+              }
+            ]
+          },
+          "label": {
+            "type": [
+              "string",
+              "null"
+            ]
+          },
+          "name": {
+            "description": "Invocation surface name (`agent_trigger`, `submit_result`, or a function id in native exposure mode).",
+            "type": "string"
+          },
+          "parameters": {
+            "default": null,
+            "description": "JSON Schema for the function parameters."
+          }
+        },
+        "required": [
+          "description",
+          "name"
+        ],
+        "type": "object"
+      },
+      "AgentMessage": {
+        "description": "The canonical transcript message union, discriminated by `role`.",
+        "oneOf": [
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "role": {
+                "enum": [
+                  "user"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "error_kind": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/ErrorKind"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "error_message": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "model": {
+                "type": "string"
+              },
+              "native_stop_reason": {
+                "description": "Provider's raw finish reason, passed through untouched.",
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "provider": {
+                "type": "string"
+              },
+              "role": {
+                "enum": [
+                  "assistant"
+                ],
+                "type": "string"
+              },
+              "stop_reason": {
+                "$ref": "#/definitions/StopReason"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              },
+              "usage": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/Usage"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "warnings": {
+                "description": "Report-and-continue notices (e.g. dropped unsupported param).",
+                "items": {
+                  "type": "string"
+                },
+                "type": [
+                  "array",
+                  "null"
+                ]
+              }
+            },
+            "required": [
+              "content",
+              "model",
+              "provider",
+              "role",
+              "stop_reason",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "details": {
+                "default": null,
+                "description": "Opaque structured payload kept alongside the rendered content."
+              },
+              "function_call_id": {
+                "description": "Echoes the `function_call` block id this result answers.",
+                "type": "string"
+              },
+              "function_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "default": false,
+                "type": "boolean"
+              },
+              "role": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "function_id",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "description": "Escape hatch for app-specific transcript items (system notices, UI markers, attachments, ...). No provider wire mapping — `context::assemble` excludes these from the model-facing list.",
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "custom_type": {
+                "description": "App-defined discriminator.",
+                "type": "string"
+              },
+              "details": true,
+              "display": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "role": {
+                "enum": [
+                  "custom"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "custom_type",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ContentBlock": {
+        "description": "The atomic unit of message content. A message's `content` is an ordered array of these.",
+        "oneOf": [
+          {
+            "properties": {
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "data": {
+                "description": "Base64-encoded image bytes.",
+                "type": "string"
+              },
+              "mime": {
+                "description": "MIME type, e.g. `image/png`.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "image"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "data",
+              "mime",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "signature": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "thinking"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "arguments": {
+                "default": null,
+                "description": "Model-produced arguments (JSON)."
+              },
+              "function_id": {
+                "description": "The iii function id to invoke.",
+                "type": "string"
+              },
+              "id": {
+                "description": "Unique per call, echoed by the result.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "function_call"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "function_id",
+              "id",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "function_call_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "type": [
+                  "boolean",
+                  "null"
+                ]
+              },
+              "type": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ErrorKind": {
+        "description": "Coarse error classification carried by failed assistant messages.",
+        "enum": [
+          "auth_expired",
+          "rate_limited",
+          "context_overflow",
+          "transient",
+          "permanent"
+        ],
+        "type": "string"
+      },
+      "ExecutionMode": {
+        "enum": [
+          "parallel",
+          "sequential"
+        ],
+        "type": "string"
+      },
+      "ModelInput": {
+        "description": "How callers identify the target model (context-manager.md § Model input). Inline `limits` keep the call standalone; `id`/`provider` alone defer limit resolution to `router::models::get`.",
+        "properties": {
+          "id": {
+            "description": "Model id, e.g. `claude-sonnet-4`.",
+            "type": "string"
+          },
+          "limits": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ModelLimits"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "description": "Inline token limits; when present no router lookup happens."
+          },
+          "provider": {
+            "description": "Provider id, e.g. `anthropic`; disambiguates when a model id exists on multiple providers.",
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "id"
+        ],
+        "type": "object"
+      },
+      "ModelLimits": {
+        "description": "Token limits a caller may supply inline to stay fully standalone.",
+        "properties": {
+          "context_window": {
+            "description": "Total context window in tokens.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_limit": {
+            "description": "Usable input budget when distinct from `context_window`.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_output_tokens": {
+            "description": "Maximum output tokens the model can produce per call.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "context_window",
+          "max_output_tokens"
+        ],
+        "type": "object"
+      },
+      "StopReason": {
+        "description": "Why an assistant message stopped generating.",
+        "enum": [
+          "end",
+          "length",
+          "function_call",
+          "aborted",
+          "error"
+        ],
+        "type": "string"
+      },
+      "Usage": {
+        "description": "Token / cost accounting reported by providers.",
+        "properties": {
+          "cache_read": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "input": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      }
+    },
+    "properties": {
+      "messages": {
+        "description": "Messages to estimate, oldest first.",
+        "items": {
+          "$ref": "#/definitions/AgentMessage"
+        },
+        "type": [
+          "array",
+          "null"
+        ]
+      },
+      "model": {
+        "allOf": [
+          {
+            "$ref": "#/definitions/ModelInput"
+          }
+        ],
+        "description": "Tokenizer selection; falls back to a generic estimator."
+      },
+      "system_prompt": {
+        "default": null,
+        "description": "Counted on top of the messages when present.",
+        "type": [
+          "string",
+          "null"
+        ]
+      },
+      "tools": {
+        "default": null,
+        "description": "Invocation schema(s) to include in the estimate (typically the single `agent_trigger` entry).",
+        "items": {
+          "$ref": "#/definitions/AgentFunction"
+        },
+        "type": [
+          "array",
+          "null"
+        ]
+      }
+    },
+    "required": [
+      "model"
+    ],
+    "title": "CountTokensRequest",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "ByRoleTokens": {
+        "description": "Per-role token breakdown of the `messages` array.",
+        "properties": {
+          "assistant": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "custom": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "function_result": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "user": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "assistant",
+          "custom",
+          "function_result",
+          "user"
+        ],
+        "type": "object"
+      },
+      "EstimatorName": {
+        "description": "Which estimator produced the count.",
+        "enum": [
+          "tokenizer",
+          "heuristic"
+        ],
+        "type": "string"
+      }
+    },
+    "properties": {
+      "by_role": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/ByRoleTokens"
+          },
+          {
+            "type": "null"
+          }
+        ],
+        "description": "Breakdown of the message tokens by role (system prompt and tools are not part of any role bucket)."
+      },
+      "estimator": {
+        "$ref": "#/definitions/EstimatorName"
+      },
+      "tokens": {
+        "description": "Total estimate: messages + system prompt + tools.",
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      }
+    },
+    "required": [
+      "estimator",
+      "tokens"
+    ],
+    "title": "CountTokensResponse",
+    "type": "object"
+  }
+}
diff --git a/context-manager/tests/golden/schemas/context.prune.json b/context-manager/tests/golden/schemas/context.prune.json
new file mode 100644
index 00000000..cfa7177c
--- /dev/null
+++ b/context-manager/tests/golden/schemas/context.prune.json
@@ -0,0 +1,999 @@
+{
+  "description": "Replace verbose function outputs with placeholders without summarising. A cheaper first pass before compaction.",
+  "function_id": "context::prune",
+  "request_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "AgentMessage": {
+        "description": "The canonical transcript message union, discriminated by `role`.",
+        "oneOf": [
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "role": {
+                "enum": [
+                  "user"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "error_kind": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/ErrorKind"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "error_message": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "model": {
+                "type": "string"
+              },
+              "native_stop_reason": {
+                "description": "Provider's raw finish reason, passed through untouched.",
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "provider": {
+                "type": "string"
+              },
+              "role": {
+                "enum": [
+                  "assistant"
+                ],
+                "type": "string"
+              },
+              "stop_reason": {
+                "$ref": "#/definitions/StopReason"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              },
+              "usage": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/Usage"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "warnings": {
+                "description": "Report-and-continue notices (e.g. dropped unsupported param).",
+                "items": {
+                  "type": "string"
+                },
+                "type": [
+                  "array",
+                  "null"
+                ]
+              }
+            },
+            "required": [
+              "content",
+              "model",
+              "provider",
+              "role",
+              "stop_reason",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "details": {
+                "default": null,
+                "description": "Opaque structured payload kept alongside the rendered content."
+              },
+              "function_call_id": {
+                "description": "Echoes the `function_call` block id this result answers.",
+                "type": "string"
+              },
+              "function_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "default": false,
+                "type": "boolean"
+              },
+              "role": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "function_id",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "description": "Escape hatch for app-specific transcript items (system notices, UI markers, attachments, ...). No provider wire mapping — `context::assemble` excludes these from the model-facing list.",
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "custom_type": {
+                "description": "App-defined discriminator.",
+                "type": "string"
+              },
+              "details": true,
+              "display": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "role": {
+                "enum": [
+                  "custom"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "custom_type",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ContentBlock": {
+        "description": "The atomic unit of message content. A message's `content` is an ordered array of these.",
+        "oneOf": [
+          {
+            "properties": {
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "data": {
+                "description": "Base64-encoded image bytes.",
+                "type": "string"
+              },
+              "mime": {
+                "description": "MIME type, e.g. `image/png`.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "image"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "data",
+              "mime",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "signature": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "thinking"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "arguments": {
+                "default": null,
+                "description": "Model-produced arguments (JSON)."
+              },
+              "function_id": {
+                "description": "The iii function id to invoke.",
+                "type": "string"
+              },
+              "id": {
+                "description": "Unique per call, echoed by the result.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "function_call"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "function_id",
+              "id",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "function_call_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "type": [
+                  "boolean",
+                  "null"
+                ]
+              },
+              "type": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ErrorKind": {
+        "description": "Coarse error classification carried by failed assistant messages.",
+        "enum": [
+          "auth_expired",
+          "rate_limited",
+          "context_overflow",
+          "transient",
+          "permanent"
+        ],
+        "type": "string"
+      },
+      "ModelInput": {
+        "description": "How callers identify the target model (context-manager.md § Model input). Inline `limits` keep the call standalone; `id`/`provider` alone defer limit resolution to `router::models::get`.",
+        "properties": {
+          "id": {
+            "description": "Model id, e.g. `claude-sonnet-4`.",
+            "type": "string"
+          },
+          "limits": {
+            "anyOf": [
+              {
+                "$ref": "#/definitions/ModelLimits"
+              },
+              {
+                "type": "null"
+              }
+            ],
+            "description": "Inline token limits; when present no router lookup happens."
+          },
+          "provider": {
+            "description": "Provider id, e.g. `anthropic`; disambiguates when a model id exists on multiple providers.",
+            "type": [
+              "string",
+              "null"
+            ]
+          }
+        },
+        "required": [
+          "id"
+        ],
+        "type": "object"
+      },
+      "ModelLimits": {
+        "description": "Token limits a caller may supply inline to stay fully standalone.",
+        "properties": {
+          "context_window": {
+            "description": "Total context window in tokens.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          },
+          "input_limit": {
+            "description": "Usable input budget when distinct from `context_window`.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "max_output_tokens": {
+            "description": "Maximum output tokens the model can produce per call.",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": "integer"
+          }
+        },
+        "required": [
+          "context_window",
+          "max_output_tokens"
+        ],
+        "type": "object"
+      },
+      "PruneOptions": {
+        "properties": {
+          "max_output_chars": {
+            "default": null,
+            "description": "Per-output verbosity threshold in chars (default 2000): outputs at or under it are never pruned.",
+            "format": "uint",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "min_free_tokens": {
+            "default": null,
+            "description": "Skip the pass entirely when it would free less (default 20000).",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "protect_recent_tokens": {
+            "default": null,
+            "description": "Newest function-output tokens never pruned (default 40000).",
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "protected_functions": {
+            "default": null,
+            "description": "`function_id`s whose outputs are never pruned.",
+            "items": {
+              "type": "string"
+            },
+            "type": [
+              "array",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      },
+      "StopReason": {
+        "description": "Why an assistant message stopped generating.",
+        "enum": [
+          "end",
+          "length",
+          "function_call",
+          "aborted",
+          "error"
+        ],
+        "type": "string"
+      },
+      "Usage": {
+        "description": "Token / cost accounting reported by providers.",
+        "properties": {
+          "cache_read": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "input": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      }
+    },
+    "properties": {
+      "messages": {
+        "description": "Messages to prune, oldest first.",
+        "items": {
+          "$ref": "#/definitions/AgentMessage"
+        },
+        "type": [
+          "array",
+          "null"
+        ]
+      },
+      "model": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/ModelInput"
+          },
+          {
+            "type": "null"
+          }
+        ],
+        "default": null,
+        "description": "Only used for token math; the heuristic estimator needs no model, so this is optional."
+      },
+      "options": {
+        "anyOf": [
+          {
+            "$ref": "#/definitions/PruneOptions"
+          },
+          {
+            "type": "null"
+          }
+        ]
+      }
+    },
+    "title": "PruneRequest",
+    "type": "object"
+  },
+  "response_schema": {
+    "$schema": "http://json-schema.org/draft-07/schema#",
+    "definitions": {
+      "AgentMessage": {
+        "description": "The canonical transcript message union, discriminated by `role`.",
+        "oneOf": [
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "role": {
+                "enum": [
+                  "user"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "error_kind": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/ErrorKind"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "error_message": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "model": {
+                "type": "string"
+              },
+              "native_stop_reason": {
+                "description": "Provider's raw finish reason, passed through untouched.",
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "provider": {
+                "type": "string"
+              },
+              "role": {
+                "enum": [
+                  "assistant"
+                ],
+                "type": "string"
+              },
+              "stop_reason": {
+                "$ref": "#/definitions/StopReason"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              },
+              "usage": {
+                "anyOf": [
+                  {
+                    "$ref": "#/definitions/Usage"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ]
+              },
+              "warnings": {
+                "description": "Report-and-continue notices (e.g. dropped unsupported param).",
+                "items": {
+                  "type": "string"
+                },
+                "type": [
+                  "array",
+                  "null"
+                ]
+              }
+            },
+            "required": [
+              "content",
+              "model",
+              "provider",
+              "role",
+              "stop_reason",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "details": {
+                "default": null,
+                "description": "Opaque structured payload kept alongside the rendered content."
+              },
+              "function_call_id": {
+                "description": "Echoes the `function_call` block id this result answers.",
+                "type": "string"
+              },
+              "function_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "default": false,
+                "type": "boolean"
+              },
+              "role": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "function_id",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          },
+          {
+            "description": "Escape hatch for app-specific transcript items (system notices, UI markers, attachments, ...). No provider wire mapping — `context::assemble` excludes these from the model-facing list.",
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "custom_type": {
+                "description": "App-defined discriminator.",
+                "type": "string"
+              },
+              "details": true,
+              "display": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "role": {
+                "enum": [
+                  "custom"
+                ],
+                "type": "string"
+              },
+              "timestamp": {
+                "description": "Milliseconds since epoch.",
+                "format": "int64",
+                "type": "integer"
+              }
+            },
+            "required": [
+              "content",
+              "custom_type",
+              "role",
+              "timestamp"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ContentBlock": {
+        "description": "The atomic unit of message content. A message's `content` is an ordered array of these.",
+        "oneOf": [
+          {
+            "properties": {
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "text"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "data": {
+                "description": "Base64-encoded image bytes.",
+                "type": "string"
+              },
+              "mime": {
+                "description": "MIME type, e.g. `image/png`.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "image"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "data",
+              "mime",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "signature": {
+                "type": [
+                  "string",
+                  "null"
+                ]
+              },
+              "text": {
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "thinking"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "text",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "arguments": {
+                "default": null,
+                "description": "Model-produced arguments (JSON)."
+              },
+              "function_id": {
+                "description": "The iii function id to invoke.",
+                "type": "string"
+              },
+              "id": {
+                "description": "Unique per call, echoed by the result.",
+                "type": "string"
+              },
+              "type": {
+                "enum": [
+                  "function_call"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "function_id",
+              "id",
+              "type"
+            ],
+            "type": "object"
+          },
+          {
+            "properties": {
+              "content": {
+                "items": {
+                  "$ref": "#/definitions/ContentBlock"
+                },
+                "type": "array"
+              },
+              "function_call_id": {
+                "type": "string"
+              },
+              "is_error": {
+                "type": [
+                  "boolean",
+                  "null"
+                ]
+              },
+              "type": {
+                "enum": [
+                  "function_result"
+                ],
+                "type": "string"
+              }
+            },
+            "required": [
+              "content",
+              "function_call_id",
+              "type"
+            ],
+            "type": "object"
+          }
+        ]
+      },
+      "ErrorKind": {
+        "description": "Coarse error classification carried by failed assistant messages.",
+        "enum": [
+          "auth_expired",
+          "rate_limited",
+          "context_overflow",
+          "transient",
+          "permanent"
+        ],
+        "type": "string"
+      },
+      "StopReason": {
+        "description": "Why an assistant message stopped generating.",
+        "enum": [
+          "end",
+          "length",
+          "function_call",
+          "aborted",
+          "error"
+        ],
+        "type": "string"
+      },
+      "Usage": {
+        "description": "Token / cost accounting reported by providers.",
+        "properties": {
+          "cache_read": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cache_write": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "cost_usd": {
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
+          "input": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "output": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          },
+          "reasoning": {
+            "format": "uint64",
+            "minimum": 0.0,
+            "type": [
+              "integer",
+              "null"
+            ]
+          }
+        },
+        "type": "object"
+      }
+    },
+    "properties": {
+      "messages": {
+        "description": "Same array with pruned outputs replaced by placeholders — no message or block is ever removed.",
+        "items": {
+          "$ref": "#/definitions/AgentMessage"
+        },
+        "type": "array"
+      },
+      "pruned_parts": {
+        "description": "Outputs replaced with placeholders.",
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      },
+      "pruned_tokens": {
+        "description": "Estimated tokens freed.",
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      },
+      "scanned_parts": {
+        "description": "Prunable outputs examined.",
+        "format": "uint64",
+        "minimum": 0.0,
+        "type": "integer"
+      }
+    },
+    "required": [
+      "messages",
+      "pruned_parts",
+      "pruned_tokens",
+      "scanned_parts"
+    ],
+    "title": "PruneResponse",
+    "type": "object"
+  }
+}
diff --git a/context-manager/tests/integration.rs b/context-manager/tests/integration.rs
new file mode 100644
index 00000000..46e44168
--- /dev/null
+++ b/context-manager/tests/integration.rs
@@ -0,0 +1,98 @@
+//! End-to-end test: spawn the `iii` engine + the worker binary as
+//! subprocesses, then drive the worker through `iii-sdk` as a client.
+//! Self-skips when `iii` is not on PATH.
+
+use std::process::{Child, Command, Stdio};
+use std::time::Duration;
+
+use iii_sdk::{register_worker, IIIError, InitOptions, TriggerRequest};
+use serde_json::json;
+use tokio::time::{sleep, timeout};
+
+const ENGINE_WS: &str = "ws://127.0.0.1:49134";
+
+struct Harness {
+    iii: Child,
+    worker: Child,
+}
+
+impl Drop for Harness {
+    fn drop(&mut self) {
+        let _ = self.worker.kill();
+        let _ = self.worker.wait();
+        let _ = self.iii.kill();
+        let _ = self.iii.wait();
+    }
+}
+
+async fn boot() -> Option<Harness> {
+    let iii_bin = which::which("iii").ok()?;
+
+    let iii = Command::new(&iii_bin)
+        .arg("--use-default-config")
+        .stdout(Stdio::null())
+        .stderr(Stdio::null())
+        .spawn()
+        .ok()?;
+
+    sleep(Duration::from_millis(800)).await;
+
+    let worker_bin = env!("CARGO_BIN_EXE_context-manager");
+    let manifest_dir = env!("CARGO_MANIFEST_DIR");
+    let config_path = format!("{manifest_dir}/config.yaml");
+
+    let worker = Command::new(worker_bin)
+        .args(["--url", ENGINE_WS, "--config", &config_path])
+        .stdout(Stdio::null())
+        .stderr(Stdio::null())
+        .spawn()
+        .ok()?;
+
+    Some(Harness { iii, worker })
+}
+
+#[tokio::test]
+async fn end_to_end_via_iii_sdk() {
+    let Some(_h) = boot().await else {
+        eprintln!("skipping: `iii` binary not on PATH");
+        return;
+    };
+
+    let client = register_worker(ENGINE_WS, InitOptions::default());
+
+    // Poll until the worker's functions are routable (registration
+    // flows over its own connection; no fixed sleep).
+    let deadline = std::time::Instant::now() + Duration::from_secs(20);
+    let payload = json!({
+        "messages": [{ "role": "user", "content": [{ "type": "text", "text": "hi" }],
+                       "timestamp": 1 }],
+        "model": { "id": "any-model" }
+    });
+    let result = loop {
+        let attempt = timeout(
+            Duration::from_secs(5),
+            client.trigger(TriggerRequest {
+                function_id: "context::count-tokens".into(),
+                payload: payload.clone(),
+                action: None,
+                timeout_ms: Some(4_000),
+            }),
+        )
+        .await;
+        match attempt {
+            Ok(Ok(v)) => break v,
+            Ok(Err(IIIError::NotConnected)) | Err(_) => {}
+            Ok(Err(_)) => {}
+        }
+        assert!(
+            std::time::Instant::now() < deadline,
+            "context::count-tokens did not become routable within 20s"
+        );
+        sleep(Duration::from_millis(250)).await;
+    };
+
+    assert_eq!(result["estimator"], "heuristic");
+    assert!(result["tokens"].as_u64().is_some_and(|t| t > 0));
+
+    client.shutdown_async().await;
+}
diff --git a/context-manager/tests/manifest.rs b/context-manager/tests/manifest.rs
new file mode 100644
index 00000000..19f9e134
--- /dev/null
+++ b/context-manager/tests/manifest.rs
@@ -0,0 +1,44 @@
+//! `--manifest` subprocess contract test (registry publish pipeline).
+
+use std::process::Command;
+
+use serde_json::Value;
+
+#[test]
+fn manifest_subcommand_emits_valid_json() {
+    let bin = env!("CARGO_BIN_EXE_context-manager");
+    let output = Command::new(bin)
+        .arg("--manifest")
+        .output()
+        .expect("spawn context-manager --manifest");
+
+    assert!(
+        output.status.success(),
+        "binary exited with {:?}; stderr: {}",
+        output.status,
+        String::from_utf8_lossy(&output.stderr),
+    );
+
+    let stdout = String::from_utf8(output.stdout).expect("manifest stdout is utf-8");
+    let manifest: Value = serde_json::from_str(&stdout).expect("manifest stdout is valid JSON");
+
+    assert_eq!(manifest["name"], env!("CARGO_PKG_NAME"));
+    assert_eq!(manifest["version"], env!("CARGO_PKG_VERSION"));
+    assert!(
+        manifest["description"]
+            .as_str()
+            .is_some_and(|d| !d.is_empty()),
+        "description must be non-empty"
+    );
+    assert!(
+        manifest["default_config"].is_object(),
+        "default_config must be an object"
+    );
+    assert!(
+        !manifest["supported_targets"]
+            .as_array()
+            .expect("supported_targets must be an array")
+            .is_empty(),
+        "supported_targets must not be empty"
+    );
+}
diff --git a/context-manager/tests/schemas.rs b/context-manager/tests/schemas.rs
new file mode 100644
index 00000000..3383a48f
--- /dev/null
+++ b/context-manager/tests/schemas.rs
@@ -0,0 +1,85 @@
+//! Wire-schema snapshots for the four `context::*` functions.
+//!
+//! `context_manager::functions::catalog()` is the single source of
+//! truth for each function's id, registration description, and
+//! schemars-derived request/response schemas (generated with the same
+//! `SchemaSettings::draft07()` construction iii-sdk uses at
+//! registration, from the same input/output structs). Each entry is
+//! serialized to pretty JSON and compared against
+//! `tests/golden/schemas/<id>.json` (`::` maps to `.` in filenames).
+//!
+//! These snapshots ARE the product surface consumed by callers and
+//! agents — any schema or description change must land as an explicit
+//! golden diff. Regenerate with `UPDATE_GOLDENS=1 cargo test`.
+
+mod support;
+
+use context_manager::functions::{catalog, FunctionSpec};
+
+fn golden_file_name(function_id: &str) -> String {
+    format!("schemas/{}.json", function_id.replace("::", "."))
+}
+
+fn spec_to_pretty_json(spec: &FunctionSpec) -> String {
+    let value = serde_json::json!({
+        "function_id": spec.function_id,
+        "description": spec.description,
+        "request_schema": spec.request_schema,
+        "response_schema": spec.response_schema,
+    });
+    let mut pretty = serde_json::to_string_pretty(&value).expect("spec serializes");
+    pretty.push('\n');
+    pretty
+}
+
+/// The catalog must cover exactly the four registered functions, in
+/// registration order (kept in lockstep with `register_all`).
+#[test]
+fn catalog_lists_all_four_functions_in_registration_order() {
+    let ids: Vec<&str> = catalog().iter().map(|s| s.function_id).collect();
+    assert_eq!(
+        ids,
+        vec![
+            "context::assemble",
+            "context::compact",
+            "context::prune",
+            "context::count-tokens",
+        ]
+    );
+}
+
+/// Every catalog entry matches its committed golden. Mismatches are
+/// collected across ALL functions before failing so one run shows the
+/// full drift, not just the first file.
+#[test]
+fn wire_schema_snapshots_match_goldens() {
+    let mut failures = Vec::new();
+    for spec in catalog() {
+        let rel = golden_file_name(spec.function_id);
+        let actual = spec_to_pretty_json(&spec);
+        if let Err(msg) = support::check_golden(&rel, &actual) {
+            failures.push(msg);
+        }
+    }
+    assert!(
+        failures.is_empty(),
+        "{} wire-schema golden(s) drifted:\n\n{}",
+        failures.len(),
+        failures.join("\n")
+    );
+}
+
+/// The spec strings callers are documented to rely on must appear in
+/// the schemas' doc-comment descriptions (a rename here is a breaking
+/// API change even if types still compile).
+#[test]
+fn schemas_carry_field_descriptions() {
+    for spec in catalog() {
+        let rendered = serde_json::to_string(&spec.request_schema).expect("schema serializes");
+        assert!(
+            rendered.contains("description"),
+            "{}: request schema lost its field descriptions",
+            spec.function_id
+        );
+    }
+}
diff --git a/context-manager/tests/steps/call_steps.rs b/context-manager/tests/steps/call_steps.rs
new file mode 100644
index 00000000..9f0e4e82
--- /dev/null
+++ b/context-manager/tests/steps/call_steps.rs
@@ -0,0 +1,129 @@
+//! High-level When-steps that call the four functions over the built
+//! history, with optional options/system-prompt docstrings.
+
+use cucumber::gherkin::Step;
+use cucumber::when;
+use serde_json::{json, Value};
+
+use crate::common::world::ContextWorld;
+use crate::steps::common_steps::docstring_payload;
+
+fn history(world: &ContextWorld) -> Value {
+    Value::Array(world.messages.clone())
+}
+
+#[when(regex = r#"^I count tokens with model "([^"]+)"$"#)]
+async fn count_tokens(world: &mut ContextWorld, model: String) {
+    let payload = json!({ "messages": history(world), "model": world.model_input(&model) });
+    world.call_pure("context::count-tokens", payload).await;
+}
+
+#[when(regex = r#"^I count tokens with model "([^"]+)" and system prompt "([^"]*)"$"#)]
+async fn count_tokens_system(world: &mut ContextWorld, model: String, system_prompt: String) {
+    let payload = json!({
+        "messages": history(world),
+        "model": world.model_input(&model),
+        "system_prompt": system_prompt
+    });
+    world.call_pure("context::count-tokens", payload).await;
+}
+
+#[when(regex = r#"^I count tokens with model "([^"]+)" and tools:$"#)]
+async fn count_tokens_tools(world: &mut ContextWorld, model: String, step: &Step) {
+    let tools = docstring_payload(world, step);
+    let payload = json!({
+        "messages": history(world),
+        "model": world.model_input(&model),
+        "tools": tools
+    });
+    world.call_pure("context::count-tokens", payload).await;
+}
+
+#[when("I prune the history")]
+async fn prune(world: &mut ContextWorld) {
+    let payload = json!({ "messages": history(world) });
+    world.call_pure("context::prune", payload).await;
+}
+
+#[when("I prune the history with options:")]
+async fn prune_options(world: &mut ContextWorld, step: &Step) {
+    let options = docstring_payload(world, step);
+    let payload = json!({ "messages": history(world), "options": options });
+    world.call_pure("context::prune", payload).await;
+}
+
+#[when("I prune the response messages again with the same options:")]
+async fn prune_again(world: &mut ContextWorld, step: &Step) {
+    let options = docstring_payload(world, step);
+    let messages = world
+        .last_response
+        .as_ref()
+        .and_then(|r| r.get("messages"))
+        .cloned()
+        .expect("a prior prune response with messages");
+    let payload = json!({ "messages": messages, "options": options });
+    world.call_pure("context::prune", payload).await;
+}
+
+#[when(regex = r#"^I compact the history with model "([^"]+)"$"#)]
+async fn compact(world: &mut ContextWorld, model: String) {
+    let payload = json!({ "messages": history(world), "model": world.model_input(&model) });
+    world.call_pure("context::compact", payload).await;
+}
+
+#[when(regex = r#"^I compact the history with model "([^"]+)" and options:$"#)]
+async fn compact_options(world: &mut ContextWorld, model: String, step: &Step) {
+    let options = docstring_payload(world, step);
+    let payload = json!({
+        "messages": history(world),
+        "model": world.model_input(&model),
+        "options": options
+    });
+    world.call_pure("context::compact", payload).await;
+}
+
+#[when(regex = r#"^I assemble the history with model "([^"]+)"$"#)]
+async fn assemble(world: &mut ContextWorld, model: String) {
+    let payload = json!({ "messages": history(world), "model": world.model_input(&model) });
+    world.call_pure("context::assemble", payload).await;
+}
+
+#[when(regex = r#"^I assemble the history with model "([^"]+)" and options:$"#)]
+async fn assemble_options(world: &mut ContextWorld, model: String, step: &Step) {
+    let options = docstring_payload(world, step);
+    let payload = json!({
+        "messages": history(world),
+        "model": world.model_input(&model),
+        "options": options
+    });
+    world.call_pure("context::assemble", payload).await;
+}
+
+#[when(regex = r#"^I assemble the history with model "([^"]+)" and system prompt "([^"]*)"$"#)]
+async fn assemble_system(world: &mut ContextWorld, model: String, system_prompt: String) {
+    let payload = json!({
+        "messages": history(world),
+        "model": world.model_input(&model),
+        "system_prompt": system_prompt
+    });
+    world.call_pure("context::assemble", payload).await;
+}
+
+#[when(
+    regex = r#"^I assemble the history with model "([^"]+)", system prompt "([^"]*)" and options:$"#
+)]
+async fn assemble_system_options(
+    world: &mut ContextWorld,
+    model: String,
+    system_prompt: String,
+    step: &Step,
+) {
+    let options = docstring_payload(world, step);
+    let payload = json!({
+        "messages": history(world),
+        "model": world.model_input(&model),
+        "system_prompt": system_prompt,
+        "options": options
+    });
+    world.call_pure("context::assemble", payload).await;
+}
diff --git a/context-manager/tests/steps/common_steps.rs b/context-manager/tests/steps/common_steps.rs
new file mode 100644
index 00000000..ee696812
--- /dev/null
+++ b/context-manager/tests/steps/common_steps.rs
@@ -0,0 +1,271 @@
+//! Generic call / response / error steps shared by every feature, plus
+//! the JSON helpers other step modules reuse.
+
+use cucumber::gherkin::Step;
+use cucumber::{given, then, when};
+use serde_json::Value;
+
+use crate::common::world::ContextWorld;
+
+/// Navigate a dotted path (`applied.summary`, `messages.0.role`) into a
+/// JSON value. Array segments are numeric indices.
+pub fn lookup_path<'a>(root: &'a Value, path: &str) -> Option<&'a Value> {
+    let mut current = root;
+    for segment in path.split('.') {
+        current = match current {
+            Value::Object(map) => map.get(segment)?,
+            Value::Array(items) => items.get(segment.parse::<usize>().ok()?)?,
+            _ => return None,
+        };
+    }
+    Some(current)
+}
+
+/// Parse the right-hand side of an `is <value>` assertion: JSON when it
+/// parses, raw string otherwise (so `is "ok"`, `is 3`, `is null`,
+/// `is true` all behave naturally).
+pub fn parse_expected(raw: &str) -> Value {
+    serde_json::from_str(raw.trim()).unwrap_or_else(|_| Value::String(raw.trim().to_string()))
+}
+
+pub fn docstring_payload(world: &ContextWorld, step: &Step) -> Value {
+    let text = world.substitute(step.docstring.as_deref().unwrap_or("{}"));
+    serde_json::from_str(text.trim())
+        .unwrap_or_else(|e| panic!("payload docstring must be valid JSON: {e}\n---\n{text}\n---"))
+}
+
+pub fn response_or_panic(world: &ContextWorld) -> &Value {
+    if let Some(err) = &world.last_error {
+        panic!("expected a response but the last call failed: {err}");
+    }
+    world
+        .last_response
+        .as_ref()
+        .expect("no call has been made yet")
+}
+
+/// True when an engine-dependent step already soft-skipped this
+/// scenario — generic assertions then pass vacuously.
+pub fn skipped(world: &ContextWorld) -> bool {
+    world.soft_skipped
+}
+
+// ---------------------------------------------------------------------------
+// Calls
+// ---------------------------------------------------------------------------
+
+#[given(regex = r#"^I call "([^"]+)" with:$"#)]
+#[when(regex = r#"^I call "([^"]+)" with:$"#)]
+async fn call_with(world: &mut ContextWorld, function: String, step: &Step) {
+    let payload = docstring_payload(world, step);
+    world.call_pure(&function, payload).await;
+}
+
+#[given(regex = r#"^I call "([^"]+)" with "\{\}"$"#)]
+#[when(regex = r#"^I call "([^"]+)" with "\{\}"$"#)]
+async fn call_empty(world: &mut ContextWorld, function: String) {
+    world.call_pure(&function, serde_json::json!({})).await;
+}
+
+// ---------------------------------------------------------------------------
+// Outcome assertions
+// ---------------------------------------------------------------------------
+
+#[then("the call succeeds")]
+async fn call_succeeds(world: &mut ContextWorld) {
+    if skipped(world) {
+        return;
+    }
+    if let Some(err) = &world.last_error {
+        panic!("expected success but the call failed: {err}");
+    }
+    assert!(world.last_response.is_some(), "no call has been made yet");
+}
+
+#[then(regex = r#"^the call fails with code "([^"]+)"$"#)]
+async fn call_fails_with(world: &mut ContextWorld, code: String) {
+    if skipped(world) {
+        return;
+    }
+    let err = world
+        .last_error
+        .as_ref()
+        .unwrap_or_else(|| panic!("expected failure with {code} but the call succeeded"));
+    assert!(
+        err.contains(&code),
+        "expected error containing `{code}`, got `{err}`"
+    );
+}
+
+#[then(regex = r#"^the error contains "([^"]+)"$"#)]
+async fn error_contains(world: &mut ContextWorld, needle: String) {
+    if skipped(world) {
+        return;
+    }
+    let err = world
+        .last_error
+        .as_ref()
+        .expect("expected a failure but the call succeeded");
+    assert!(
+        err.contains(&needle),
+        "expected error containing `{needle}`, got `{err}`"
+    );
+}
+
+#[then(regex = r#"^the response field "([^"]+)" is (.+)$"#)]
+async fn field_is(world: &mut ContextWorld, path: String, raw: String) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let actual = lookup_path(response, &path)
+        .unwrap_or_else(|| panic!("path `{path}` not found in {response}"));
+    let expected = parse_expected(&raw);
+    assert_eq!(
+        actual, &expected,
+        "field `{path}`: expected {expected}, got {actual}"
+    );
+}
+
+#[then(regex = r#"^the response field "([^"]+)" contains "([^"]+)"$"#)]
+async fn field_contains(world: &mut ContextWorld, path: String, needle: String) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let actual = lookup_path(response, &path)
+        .unwrap_or_else(|| panic!("path `{path}` not found in {response}"))
+        .as_str()
+        .unwrap_or_else(|| panic!("field `{path}` is not a string"));
+    assert!(
+        actual.contains(&needle),
+        "field `{path}`: expected to contain `{needle}`, got `{actual}`"
+    );
+}
+
+#[then(regex = r#"^the response field "([^"]+)" does not contain "([^"]+)"$"#)]
+async fn field_not_contains(world: &mut ContextWorld, path: String, needle: String) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let actual = lookup_path(response, &path)
+        .unwrap_or_else(|| panic!("path `{path}` not found in {response}"))
+        .as_str()
+        .unwrap_or_else(|| panic!("field `{path}` is not a string"));
+    assert!(
+        !actual.contains(&needle),
+        "field `{path}`: expected NOT to contain `{needle}`, got `{actual}`"
+    );
+}
+
+// NOTE: comparison/absence steps deliberately avoid the word "is" so
+// they can never be ambiguous with the generic `is <value>` regex
+// above (the regex crate has no lookahead to disambiguate).
+
+#[then(regex = r#"^the response field "([^"]+)" exceeds (\d+)$"#)]
+async fn field_greater(world: &mut ContextWorld, path: String, floor: u64) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let actual = lookup_path(response, &path)
+        .and_then(Value::as_u64)
+        .unwrap_or_else(|| panic!("path `{path}` is not a number in {response}"));
+    assert!(
+        actual > floor,
+        "field `{path}`: expected > {floor}, got {actual}"
+    );
+}
+
+#[then(regex = r#"^the response field "([^"]+)" does not exceed (\d+)$"#)]
+async fn field_at_most(world: &mut ContextWorld, path: String, ceiling: u64) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let actual = lookup_path(response, &path)
+        .and_then(Value::as_u64)
+        .unwrap_or_else(|| panic!("path `{path}` is not a number in {response}"));
+    assert!(
+        actual <= ceiling,
+        "field `{path}`: expected <= {ceiling}, got {actual}"
+    );
+}
+
+#[then(regex = r#"^the response has no field "([^"]+)"$"#)]
+async fn field_absent(world: &mut ContextWorld, path: String) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    assert!(
+        lookup_path(response, &path).is_none(),
+        "expected `{path}` to be absent, but it is {:?}",
+        lookup_path(response, &path)
+    );
+}
+
+#[then(regex = r#"^the response field "([^"]+)" has (\d+) items$"#)]
+async fn field_len(world: &mut ContextWorld, path: String, expected: usize) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let actual = lookup_path(response, &path)
+        .and_then(Value::as_array)
+        .unwrap_or_else(|| panic!("path `{path}` is not an array in {response}"));
+    assert_eq!(
+        actual.len(),
+        expected,
+        "field `{path}`: expected {expected} items, got {}",
+        actual.len()
+    );
+}
+
+#[then(regex = r#"^the response field "([^"]+)" equals the sum of the "([^"]+)" fields$"#)]
+async fn field_equals_object_sum(world: &mut ContextWorld, total: String, parts: String) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let total_value = lookup_path(response, &total)
+        .and_then(Value::as_u64)
+        .unwrap_or_else(|| panic!("`{total}` is not a number"));
+    let parts_value = lookup_path(response, &parts)
+        .and_then(Value::as_object)
+        .unwrap_or_else(|| panic!("`{parts}` is not an object"));
+    let parts_sum: u64 = parts_value
+        .iter()
+        .map(|(key, value)| {
+            value
+                .as_u64()
+                .unwrap_or_else(|| panic!("`{parts}.{key}` is not a non-negative integer"))
+        })
+        .sum();
+    assert_eq!(
+        total_value, parts_sum,
+        "`{total}` did not match sum of `{parts}`"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// Bookkeeping
+// ---------------------------------------------------------------------------
+
+#[given(regex = r#"^I remember the response field "([^"]+)" as "([^"]+)"$"#)]
+#[when(regex = r#"^I remember the response field "([^"]+)" as "([^"]+)"$"#)]
+#[then(regex = r#"^I remember the response field "([^"]+)" as "([^"]+)"$"#)]
+async fn remember_field(world: &mut ContextWorld, path: String, alias: String) {
+    if skipped(world) {
+        return;
+    }
+    let response = response_or_panic(world);
+    let value = lookup_path(response, &path)
+        .unwrap_or_else(|| panic!("path `{path}` not found in {response}"));
+    let rendered = match value {
+        Value::String(s) => s.clone(),
+        other => other.to_string(),
+    };
+    world.aliases.insert(alias, rendered);
+}
diff --git a/context-manager/tests/steps/compaction_steps.rs b/context-manager/tests/steps/compaction_steps.rs
new file mode 100644
index 00000000..deac6802
--- /dev/null
+++ b/context-manager/tests/steps/compaction_steps.rs
@@ -0,0 +1,212 @@
+//! Summariser scripting, lease setup, and clock control for the
+//! compaction features.
+
+use cucumber::{given, then};
+use std::sync::atomic::Ordering;
+
+use context_manager::core::lease::{default_lease_key, LEASE_SCOPE};
+use context_manager::ports::LeaseRecord;
+use context_manager::types::AgentMessage;
+
+use crate::common::fakes::SummarizerScript;
+use crate::common::world::ContextWorld;
+
+fn parsed_history(world: &ContextWorld) -> Vec<AgentMessage> {
+    world
+        .messages
+        .iter()
+        .map(|m| serde_json::from_value(m.clone()).expect("history messages parse"))
+        .collect()
+}
+
+/// The default lease key the worker derives for the current history.
+pub fn history_lease_key(world: &ContextWorld) -> String {
+    default_lease_key(&parsed_history(world))
+}
+
+// ---------------------------------------------------------------------------
+// Summariser scripting
+// ---------------------------------------------------------------------------
+
+#[given(regex = r#"^the summariser returns "([^"]*)"$"#)]
+async fn summariser_returns(world: &mut ContextWorld, summary: String) {
+    // Gherkin quoted strings carry `\n` literally; unescape so the
+    // scripted summary matches what JSON-parsing assertions expect.
+    let summary = summary.replace("\\n", "\n");
+    world
+        .summarizer
+        .set_script(SummarizerScript::Return(summary));
+}
+
+#[given("the summariser is unavailable")]
+async fn summariser_unavailable(world: &mut ContextWorld) {
+    world.summarizer.set_script(SummarizerScript::Unavailable);
+}
+
+#[given(regex = r#"^the summariser fails with "([^"]+)"$"#)]
+async fn summariser_fails(world: &mut ContextWorld, reason: String) {
+    world.summarizer.set_script(SummarizerScript::Fail(reason));
+}
+
+#[given("the summariser returns an empty summary")]
+async fn summariser_empty(world: &mut ContextWorld) {
+    world.summarizer.set_script(SummarizerScript::Empty);
+}
+
+#[then(regex = r#"^the summariser was invoked (\d+) times?$"#)]
+async fn summariser_invoked(world: &mut ContextWorld, times: usize) {
+    if world.soft_skipped {
+        return;
+    }
+    let calls = world.summarizer.calls();
+    assert_eq!(
+        calls.len(),
+        times,
+        "expected {times} summariser call(s), got {}",
+        calls.len()
+    );
+}
+
+#[then("the summariser was never invoked")]
+async fn summariser_never(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    let calls = world.summarizer.calls();
+    assert!(
+        calls.is_empty(),
+        "expected no summariser calls, got {}",
+        calls.len()
+    );
+}
+
+#[then(regex = r#"^the summariser system prompt contains "([^"]+)"$"#)]
+async fn summariser_system_contains(world: &mut ContextWorld, needle: String) {
+    let calls = world.summarizer.calls();
+    let last = calls.last().expect("the summariser was never invoked");
+    assert!(
+        last.system_prompt.contains(&needle),
+        "summariser system prompt does not contain `{needle}`:\n{}",
+        last.system_prompt
+    );
+}
+
+#[then(regex = r#"^the summariser system prompt does not contain "([^"]+)"$"#)]
+async fn summariser_system_not_contains(world: &mut ContextWorld, needle: String) {
+    let calls = world.summarizer.calls();
+    let last = calls.last().expect("the summariser was never invoked");
+    assert!(
+        !last.system_prompt.contains(&needle),
+        "summariser system prompt unexpectedly contains `{needle}`"
+    );
+}
+
+#[then(regex = r#"^the summariser user prompt contains "([^"]+)"$"#)]
+async fn summariser_user_contains(world: &mut ContextWorld, needle: String) {
+    let calls = world.summarizer.calls();
+    let last = calls.last().expect("the summariser was never invoked");
+    assert!(
+        last.user_prompt.contains(&needle),
+        "summariser user prompt does not contain `{needle}`:\n{}",
+        last.user_prompt
+    );
+}
+
+#[then(regex = r#"^the summariser user prompt does not contain "([^"]+)"$"#)]
+async fn summariser_user_not_contains(world: &mut ContextWorld, needle: String) {
+    let calls = world.summarizer.calls();
+    let last = calls.last().expect("the summariser was never invoked");
+    assert!(
+        !last.user_prompt.contains(&needle),
+        "summariser user prompt unexpectedly contains `{needle}`"
+    );
+}
+
+#[then(regex = r#"^the summariser ran on model "([^"]+)"$"#)]
+async fn summariser_model(world: &mut ContextWorld, model: String) {
+    let calls = world.summarizer.calls();
+    let last = calls.last().expect("the summariser was never invoked");
+    assert_eq!(last.model, model);
+}
+
+// ---------------------------------------------------------------------------
+// Leases and clock
+// ---------------------------------------------------------------------------
+
+#[given(regex = r#"^a foreign lease on key "([^"]+)" taken (\d+) ms ago$"#)]
+async fn foreign_lease(world: &mut ContextWorld, key: String, age_ms: i64) {
+    let now = context_manager::ports::Clock::now_ms(world.clock.as_ref());
+    world.leases.put(
+        LEASE_SCOPE,
+        &key,
+        LeaseRecord {
+            nonce: "foreign-holder".into(),
+            ts: now - age_ms,
+        },
+    );
+}
+
+#[given(regex = r#"^a foreign lease on the history's default key taken (\d+) ms ago$"#)]
+async fn foreign_lease_default_key(world: &mut ContextWorld, age_ms: i64) {
+    let key = history_lease_key(world);
+    let now = context_manager::ports::Clock::now_ms(world.clock.as_ref());
+    world.leases.put(
+        LEASE_SCOPE,
+        &key,
+        LeaseRecord {
+            nonce: "foreign-holder".into(),
+            ts: now - age_ms,
+        },
+    );
+}
+
+#[given(regex = r#"^the clock advances by (\d+) ms$"#)]
+async fn clock_advances(world: &mut ContextWorld, ms: i64) {
+    world.clock.advance(ms);
+}
+
+#[given("the lease store is unavailable")]
+async fn lease_store_unavailable(world: &mut ContextWorld) {
+    world.leases.fail_all.store(true, Ordering::SeqCst);
+}
+
+#[then("no lease claim remains")]
+async fn no_lease_remains(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    let keys = world.leases.keys();
+    assert!(
+        keys.is_empty(),
+        "expected no lease claims, found keys {keys:?}"
+    );
+}
+
+#[then(regex = r#"^the lease on key "([^"]+)" still belongs to "([^"]+)"$"#)]
+async fn lease_still_belongs(world: &mut ContextWorld, key: String, nonce: String) {
+    if world.soft_skipped {
+        return;
+    }
+    let record = world
+        .leases
+        .peek(LEASE_SCOPE, &key)
+        .unwrap_or_else(|| panic!("no lease claim on key {key}"));
+    assert_eq!(
+        record.nonce, nonce,
+        "lease on {key} changed hands: now held by {}",
+        record.nonce
+    );
+}
+
+#[then("the lease on the history's default key still belongs to the foreign holder")]
+async fn default_lease_still_foreign(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    let key = history_lease_key(world);
+    let record = world
+        .leases
+        .peek(LEASE_SCOPE, &key)
+        .unwrap_or_else(|| panic!("no lease claim on the default key"));
+    assert_eq!(record.nonce, "foreign-holder");
+}
diff --git a/context-manager/tests/steps/engine_steps.rs b/context-manager/tests/steps/engine_steps.rs
new file mode 100644
index 00000000..ecd5389a
--- /dev/null
+++ b/context-manager/tests/steps/engine_steps.rs
@@ -0,0 +1,22 @@
+//! @engine steps: drive the production surface registered in-process
+//! over a live engine. Soft-skip when no engine is reachable.
+
+use cucumber::gherkin::Step;
+use cucumber::{given, when};
+
+use crate::common::world::ContextWorld;
+use crate::steps::common_steps::docstring_payload;
+
+#[given("an engine connection")]
+async fn engine_connection(world: &mut ContextWorld) {
+    let _ = world.skip_without_engine();
+}
+
+#[when(regex = r#"^I call "([^"]+)" on the engine with:$"#)]
+async fn call_engine_with(world: &mut ContextWorld, function: String, step: &Step) {
+    if world.skip_without_engine() {
+        return;
+    }
+    let payload = docstring_payload(world, step);
+    world.call_engine(&function, payload).await;
+}
diff --git a/context-manager/tests/steps/history_steps.rs b/context-manager/tests/steps/history_steps.rs
new file mode 100644
index 00000000..022979fe
--- /dev/null
+++ b/context-manager/tests/steps/history_steps.rs
@@ -0,0 +1,148 @@
+//! Given-steps that build the candidate history. Sizes are expressed
+//! in heuristic tokens (`~N tokens` -> N*4 chars of text), so prune
+//! accounting in assertions is exact.
+
+use cucumber::given;
+use serde_json::{json, Value};
+
+use crate::common::world::ContextWorld;
+
+fn text_block(text: &str) -> Value {
+    json!({ "type": "text", "text": text })
+}
+
+fn sized_text(tokens: u64) -> String {
+    "x".repeat((tokens * 4) as usize)
+}
+
+pub fn push_user(world: &mut ContextWorld, text: &str) {
+    let ts = world.tick();
+    world
+        .messages
+        .push(json!({ "role": "user", "content": [text_block(text)], "timestamp": ts }));
+}
+
+pub fn push_assistant(world: &mut ContextWorld, content: Vec<Value>, stop_reason: &str) {
+    let ts = world.tick();
+    world.messages.push(json!({
+        "role": "assistant", "content": content, "stop_reason": stop_reason,
+        "model": "test-model", "provider": "test", "timestamp": ts
+    }));
+}
+
+#[given("an empty history")]
+async fn empty_history(world: &mut ContextWorld) {
+    world.messages.clear();
+}
+
+#[given(regex = r#"^a user message "([^"]*)"$"#)]
+async fn user_message(world: &mut ContextWorld, text: String) {
+    push_user(world, &text);
+}
+
+#[given(regex = r#"^a user message of ~(\d+) tokens$"#)]
+async fn user_message_sized(world: &mut ContextWorld, tokens: u64) {
+    push_user(world, &sized_text(tokens));
+}
+
+#[given(regex = r#"^an assistant message "([^"]*)"$"#)]
+async fn assistant_message(world: &mut ContextWorld, text: String) {
+    push_assistant(world, vec![text_block(&text)], "end");
+}
+
+#[given(regex = r#"^an assistant message of ~(\d+) tokens$"#)]
+async fn assistant_message_sized(world: &mut ContextWorld, tokens: u64) {
+    push_assistant(world, vec![text_block(&sized_text(tokens))], "end");
+}
+
+#[given(regex = r#"^an assistant function call "([^"]+)" to "([^"]+)"$"#)]
+async fn assistant_function_call(world: &mut ContextWorld, call_id: String, function_id: String) {
+    push_assistant(
+        world,
+        vec![json!({
+            "type": "function_call", "id": call_id, "function_id": function_id,
+            "arguments": { "arg": "value" }
+        })],
+        "function_call",
+    );
+}
+
+#[given(regex = r#"^a function result for call "([^"]+)" from "([^"]+)" of ~(\d+) tokens$"#)]
+async fn function_result_sized(
+    world: &mut ContextWorld,
+    call_id: String,
+    function_id: String,
+    tokens: u64,
+) {
+    let ts = world.tick();
+    world.messages.push(json!({
+        "role": "function_result", "function_call_id": call_id, "function_id": function_id,
+        "content": [text_block(&sized_text(tokens))], "details": {}, "is_error": false,
+        "timestamp": ts
+    }));
+}
+
+#[given(regex = r#"^a function result for call "([^"]+)" from "([^"]+)" saying "([^"]*)"$"#)]
+async fn function_result_text(
+    world: &mut ContextWorld,
+    call_id: String,
+    function_id: String,
+    text: String,
+) {
+    let ts = world.tick();
+    world.messages.push(json!({
+        "role": "function_result", "function_call_id": call_id, "function_id": function_id,
+        "content": [text_block(&text)], "details": {}, "is_error": false, "timestamp": ts
+    }));
+}
+
+#[given(regex = r#"^a custom message of type "([^"]+)"$"#)]
+async fn custom_message(world: &mut ContextWorld, custom_type: String) {
+    let ts = world.tick();
+    world.messages.push(json!({
+        "role": "custom", "custom_type": custom_type,
+        "content": [text_block("app-facing bookkeeping")], "timestamp": ts
+    }));
+}
+
+#[given(regex = r#"^a custom message of type "([^"]+)" of ~(\d+) tokens$"#)]
+async fn custom_message_sized(world: &mut ContextWorld, custom_type: String, tokens: u64) {
+    let ts = world.tick();
+    world.messages.push(json!({
+        "role": "custom", "custom_type": custom_type,
+        "content": [text_block(&sized_text(tokens))], "timestamp": ts
+    }));
+}
+
+#[given(regex = r#"^a user message carrying an inline function result block for call "([^"]+)"$"#)]
+async fn user_with_inline_result(world: &mut ContextWorld, call_id: String) {
+    let ts = world.tick();
+    world.messages.push(json!({
+        "role": "user",
+        "content": [
+            { "type": "function_result", "function_call_id": call_id,
+              "content": [{ "type": "text", "text": "inline result" }] }
+        ],
+        "timestamp": ts
+    }));
+}
+
+#[given(regex = r#"^a user message containing an image of (\d+) chars$"#)]
+async fn user_with_image(world: &mut ContextWorld, chars: usize) {
+    let ts = world.tick();
+    world.messages.push(json!({
+        "role": "user",
+        "content": [{ "type": "image", "mime": "image/png", "data": "A".repeat(chars) }],
+        "timestamp": ts
+    }));
+}
+
+/// `N` plain user/assistant turn pairs of `~tokens` each (the user
+/// message carries the bulk; the assistant reply is tiny).
+#[given(regex = r#"^(\d+) conversation turns of ~(\d+) tokens each$"#)]
+async fn conversation_turns(world: &mut ContextWorld, turn_count: usize, tokens: u64) {
+    for i in 0..turn_count {
+        push_user(world, &sized_text(tokens));
+        push_assistant(world, vec![text_block(&format!("reply {i}"))], "end");
+    }
+}
diff --git a/context-manager/tests/steps/invariant_steps.rs b/context-manager/tests/steps/invariant_steps.rs
new file mode 100644
index 00000000..ccb09bd9
--- /dev/null
+++ b/context-manager/tests/steps/invariant_steps.rs
@@ -0,0 +1,174 @@
+//! Structural-invariant assertions over the response messages
+//! (context-manager.md § Structural invariants). These are the checks
+//! that prove pruned/compacted output is still accepted by providers.
+
+use std::collections::HashSet;
+
+use cucumber::then;
+use serde_json::Value;
+
+use crate::common::world::ContextWorld;
+use crate::steps::common_steps::response_or_panic;
+
+fn response_messages(world: &ContextWorld) -> Vec<Value> {
+    response_or_panic(world)
+        .get("messages")
+        .and_then(Value::as_array)
+        .expect("response has a messages array")
+        .clone()
+}
+
+/// Every `function_result` (message-level or inline block) in the
+/// returned context must have its `function_call` present earlier in
+/// the same context — providers reject orphaned results.
+#[then("call/result pairing is intact in the response messages")]
+async fn pairing_intact(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    let messages = response_messages(world);
+    let mut seen_calls: HashSet<String> = HashSet::new();
+    for (idx, message) in messages.iter().enumerate() {
+        for block in message
+            .get("content")
+            .and_then(Value::as_array)
+            .into_iter()
+            .flatten()
+        {
+            match block.get("type").and_then(Value::as_str) {
+                Some("function_call") => {
+                    if let Some(id) = block.get("id").and_then(Value::as_str) {
+                        seen_calls.insert(id.to_string());
+                    }
+                }
+                Some("function_result") => {
+                    let call_id = block
+                        .get("function_call_id")
+                        .and_then(Value::as_str)
+                        .unwrap_or_default();
+                    assert!(
+                        seen_calls.contains(call_id),
+                        "message {idx}: inline function_result for `{call_id}` has no \
+                         preceding function_call in the returned context"
+                    );
+                }
+                _ => {}
+            }
+        }
+        if message.get("role").and_then(Value::as_str) == Some("function_result") {
+            let call_id = message
+                .get("function_call_id")
+                .and_then(Value::as_str)
+                .unwrap_or_default();
+            assert!(
+                seen_calls.contains(call_id),
+                "message {idx}: function_result for `{call_id}` has no preceding \
+                 function_call in the returned context"
+            );
+        }
+    }
+}
+
+#[then("the response messages contain no custom roles")]
+async fn no_custom_roles(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    for (idx, message) in response_messages(world).iter().enumerate() {
+        assert_ne!(
+            message.get("role").and_then(Value::as_str),
+            Some("custom"),
+            "message {idx} is a custom message — it must never reach the model"
+        );
+    }
+}
+
+#[then("the response messages equal the request history")]
+async fn equals_request(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    let messages = response_messages(world);
+    assert_eq!(
+        Value::Array(messages),
+        Value::Array(world.messages.clone()),
+        "response messages must be the request history, untouched"
+    );
+}
+
+#[then("the response messages have as many messages as the request")]
+async fn same_count(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    let messages = response_messages(world);
+    assert_eq!(messages.len(), world.messages.len());
+}
+
+#[then(regex = r#"^response message (\d+) text is "([^"]*)"$"#)]
+async fn message_text_is(world: &mut ContextWorld, idx: usize, expected: String) {
+    if world.soft_skipped {
+        return;
+    }
+    let messages = response_messages(world);
+    let message = messages
+        .get(idx)
+        .unwrap_or_else(|| panic!("no response message at index {idx}"));
+    let text = message
+        .get("content")
+        .and_then(Value::as_array)
+        .and_then(|blocks| blocks.first())
+        .and_then(|block| block.get("text"))
+        .and_then(Value::as_str)
+        .unwrap_or_default();
+    assert_eq!(text, expected, "message {idx} text mismatch");
+}
+
+#[then(regex = r#"^response message (\d+) text has (\d+) chars$"#)]
+async fn message_text_len(world: &mut ContextWorld, idx: usize, expected: usize) {
+    if world.soft_skipped {
+        return;
+    }
+    let messages = response_messages(world);
+    let message = messages
+        .get(idx)
+        .unwrap_or_else(|| panic!("no response message at index {idx}"));
+    let text = message
+        .get("content")
+        .and_then(Value::as_array)
+        .and_then(|blocks| blocks.first())
+        .and_then(|block| block.get("text"))
+        .and_then(Value::as_str)
+        .unwrap_or_default();
+    assert_eq!(text.len(), expected, "message {idx} text length mismatch");
+}
+
+#[then(regex = r#"^every response message keeps its function_call_id$"#)]
+async fn keeps_call_ids(world: &mut ContextWorld) {
+    if world.soft_skipped {
+        return;
+    }
+    for (idx, message) in response_messages(world).iter().enumerate() {
+        if message.get("role").and_then(Value::as_str) == Some("function_result") {
+            let call_id = message.get("function_call_id").and_then(Value::as_str);
+            assert!(
+                call_id.is_some_and(|id| !id.is_empty()),
+                "message {idx} lost its function_call_id"
+            );
+        }
+    }
+}
+
+#[then(regex = r#"^the response messages start at request message (\d+)$"#)]
+async fn starts_at_request_index(world: &mut ContextWorld, idx: usize) {
+    if world.soft_skipped {
+        return;
+    }
+    let messages = response_messages(world);
+    let expected_tail: Vec<Value> = world.messages[idx..].to_vec();
+    assert_eq!(
+        Value::Array(messages),
+        Value::Array(expected_tail),
+        "response messages must be exactly the request messages from index {idx} on"
+    );
+}
diff --git a/context-manager/tests/steps/mod.rs b/context-manager/tests/steps/mod.rs
new file mode 100644
index 00000000..04bf07d2
--- /dev/null
+++ b/context-manager/tests/steps/mod.rs
@@ -0,0 +1,7 @@
+pub mod call_steps;
+pub mod common_steps;
+pub mod compaction_steps;
+pub mod engine_steps;
+pub mod history_steps;
+pub mod invariant_steps;
+pub mod model_steps;
diff --git a/context-manager/tests/steps/model_steps.rs b/context-manager/tests/steps/model_steps.rs
new file mode 100644
index 00000000..bfa06031
--- /dev/null
+++ b/context-manager/tests/steps/model_steps.rs
@@ -0,0 +1,152 @@
+//! Model catalog (fake router), inline-limit models, and config knobs.
+
+use cucumber::{given, then};
+use std::sync::atomic::Ordering;
+
+use context_manager::types::{Model, ThinkingLevel};
+use serde_json::json;
+
+use crate::common::world::ContextWorld;
+
+fn parse_level(raw: &str) -> ThinkingLevel {
+    serde_json::from_value(json!(raw)).unwrap_or_else(|_| panic!("unknown thinking level {raw}"))
+}
+
+#[given(
+    regex = r#"^the router knows model "([^"]+)" with context window (\d+) and max output (\d+)$"#
+)]
+async fn router_model(world: &mut ContextWorld, id: String, context_window: u64, max_output: u64) {
+    world.resolver.insert(Model {
+        id: id.clone(),
+        provider: "test".into(),
+        context_window,
+        max_output_tokens: max_output,
+        input_limit: None,
+        thinking_budgets: None,
+    });
+    world.model_inputs.insert(id.clone(), json!({ "id": id }));
+}
+
+#[given(
+    regex = r#"^the router knows model "([^"]+)" with input limit (\d+), context window (\d+) and max output (\d+)$"#
+)]
+async fn router_model_input_limit(
+    world: &mut ContextWorld,
+    id: String,
+    input_limit: u64,
+    context_window: u64,
+    max_output: u64,
+) {
+    world.resolver.insert(Model {
+        id: id.clone(),
+        provider: "test".into(),
+        context_window,
+        max_output_tokens: max_output,
+        input_limit: Some(input_limit),
+        thinking_budgets: None,
+    });
+    world.model_inputs.insert(id.clone(), json!({ "id": id }));
+}
+
+#[given(
+    regex = r#"^the router model "([^"]+)" declares a thinking budget of (\d+) for "([^"]+)"$"#
+)]
+async fn router_model_thinking(world: &mut ContextWorld, id: String, budget: u64, level: String) {
+    let mut models = world
+        .resolver
+        .models
+        .lock()
+        .unwrap_or_else(|poison| poison.into_inner());
+    let model = models
+        .get_mut(&id)
+        .unwrap_or_else(|| panic!("declare the router model {id} first"));
+    model
+        .thinking_budgets
+        .get_or_insert_with(Default::default)
+        .insert(parse_level(&level), budget);
+}
+
+#[given("the router is unavailable")]
+async fn router_unavailable(world: &mut ContextWorld) {
+    world.resolver.unavailable.store(true, Ordering::SeqCst);
+}
+
+#[given(regex = r#"^inline model "([^"]+)" with context window (\d+) and max output (\d+)$"#)]
+async fn inline_model(world: &mut ContextWorld, id: String, context_window: u64, max_output: u64) {
+    world.model_inputs.insert(
+        id.clone(),
+        json!({
+            "id": id,
+            "limits": { "context_window": context_window, "max_output_tokens": max_output }
+        }),
+    );
+}
+
+#[given(
+    regex = r#"^inline model "([^"]+)" with input limit (\d+), context window (\d+) and max output (\d+)$"#
+)]
+async fn inline_model_input_limit(
+    world: &mut ContextWorld,
+    id: String,
+    input_limit: u64,
+    context_window: u64,
+    max_output: u64,
+) {
+    world.model_inputs.insert(
+        id.clone(),
+        json!({
+            "id": id,
+            "limits": {
+                "context_window": context_window,
+                "max_output_tokens": max_output,
+                "input_limit": input_limit
+            }
+        }),
+    );
+}
+
+#[given(regex = r#"^config "([^"]+)" is (\d+)$"#)]
+async fn config_numeric(world: &mut ContextWorld, field: String, value: u64) {
+    match field.as_str() {
+        "reserved_tokens_cap" => world.config.reserved_tokens_cap = value,
+        "reserved_pct" => world.config.reserved_pct = value,
+        "tail_turns" => world.config.tail_turns = value as usize,
+        "protect_recent_tokens" => world.config.protect_recent_tokens = value,
+        "min_free_tokens" => world.config.min_free_tokens = value,
+        "max_output_chars" => world.config.max_output_chars = value as usize,
+        "lease_ttl_secs" => world.config.lease_ttl_secs = value,
+        "summarizer_timeout_ms" => world.config.summarizer_timeout_ms = value,
+        other => panic!("unknown numeric config field {other}"),
+    }
+}
+
+#[given(regex = r#"^config allow_fallback_limits is (true|false)$"#)]
+async fn config_fallback(world: &mut ContextWorld, value: String) {
+    world.config.allow_fallback_limits = value == "true";
+}
+
+#[then(regex = r#"^the router was asked for model "([^"]+)"$"#)]
+async fn router_was_asked(world: &mut ContextWorld, id: String) {
+    let calls = world
+        .resolver
+        .calls
+        .lock()
+        .unwrap_or_else(|poison| poison.into_inner());
+    assert!(
+        calls.iter().any(|(_, called_id)| called_id == &id),
+        "router was never asked for {id}; calls: {calls:?}"
+    );
+}
+
+#[then("the router was never queried")]
+async fn router_never_queried(world: &mut ContextWorld) {
+    let calls = world
+        .resolver
+        .calls
+        .lock()
+        .unwrap_or_else(|poison| poison.into_inner());
+    assert!(
+        calls.is_empty(),
+        "expected no router lookups, got {calls:?}"
+    );
+}
diff --git a/context-manager/tests/support/mod.rs b/context-manager/tests/support/mod.rs
new file mode 100644
index 00000000..63e879cc
--- /dev/null
+++ b/context-manager/tests/support/mod.rs
@@ -0,0 +1,90 @@
+//! Hand-rolled golden-file harness (deliberately no `insta`/snapshot
+//! dependency). Goldens live under `tests/golden/` and are committed;
+//! any wire-surface change must show up as an explicit, reviewed diff.
+//!
+//! Workflow:
+//! - `cargo test` compares actual output against the committed goldens.
+//! - `UPDATE_GOLDENS=1 cargo test` regenerates the files; review the git
+//!   diff, then commit the new goldens alongside the change that caused
+//!   them.
+
+#![allow(dead_code)]
+
+use std::fs;
+use std::path::PathBuf;
+
+/// Root of the committed golden files.
+pub fn golden_root() -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/golden")
+}
+
+fn update_mode() -> bool {
+    std::env::var("UPDATE_GOLDENS")
+        .map(|v| v == "1")
+        .unwrap_or(false)
+}
+
+/// Compare `actual` against the golden file at `tests/golden/<rel>`.
+/// Returns `Err(readable diff hint)` on mismatch or missing golden;
+/// with `UPDATE_GOLDENS=1` the file is (re)written and the check passes.
+pub fn check_golden(rel: &str, actual: &str) -> Result<(), String> {
+    let path = golden_root().join(rel);
+    if update_mode() {
+        if let Some(parent) = path.parent() {
+            fs::create_dir_all(parent).map_err(|e| format!("create {}: {e}", parent.display()))?;
+        }
+        fs::write(&path, actual).map_err(|e| format!("write {}: {e}", path.display()))?;
+        return Ok(());
+    }
+    let expected = fs::read_to_string(&path).map_err(|e| {
+        format!(
+            "golden file {} unreadable ({e}).\n\
+             Run `UPDATE_GOLDENS=1 cargo test` to (re)generate, then review \
+             and commit the diff.",
+            path.display()
+        )
+    })?;
+    if expected == actual {
+        return Ok(());
+    }
+    Err(diff_hint(rel, &expected, actual))
+}
+
+/// Readable first-divergence diff hint: line number, expected vs actual
+/// around the mismatch, and the regeneration instructions.
+fn diff_hint(rel: &str, expected: &str, actual: &str) -> String {
+    let exp_lines: Vec<&str> = expected.lines().collect();
+    let act_lines: Vec<&str> = actual.lines().collect();
+    let first_diff = exp_lines
+        .iter()
+        .zip(act_lines.iter())
+        .position(|(e, a)| e != a)
+        .unwrap_or_else(|| exp_lines.len().min(act_lines.len()));
+
+    const CONTEXT: usize = 3;
+    let lo = first_diff.saturating_sub(CONTEXT);
+    let hi = (first_diff + CONTEXT + 1).max(first_diff + 1);
+
+    let mut out = format!(
+        "golden mismatch: tests/golden/{rel}\n\
+         first divergence at line {} (expected {} lines, actual {} lines)\n",
+        first_diff + 1,
+        exp_lines.len(),
+        act_lines.len()
+    );
+    out.push_str("--- expected (golden) ---\n");
+    for (i, line) in exp_lines.iter().enumerate().skip(lo).take(hi - lo) {
+        let marker = if i == first_diff { ">" } else { " " };
+        out.push_str(&format!("{marker} {:>4} | {line}\n", i + 1));
+    }
+    out.push_str("--- actual ---\n");
+    for (i, line) in act_lines.iter().enumerate().skip(lo).take(hi - lo) {
+        let marker = if i == first_diff { ">" } else { " " };
+        out.push_str(&format!("{marker} {:>4} | {line}\n", i + 1));
+    }
+    out.push_str(
+        "If this change is intentional, run `UPDATE_GOLDENS=1 cargo test`, \
+         review the git diff, and commit the updated goldens.\n",
+    );
+    out
+}
diff --git a/docs/sops/binary-worker.md b/docs/sops/binary-worker.md
index 5ea2a8a1..7c024c54 100644
--- a/docs/sops/binary-worker.md
+++ b/docs/sops/binary-worker.md
@@ -371,6 +371,16 @@ final messages can be lost.
 
 ## 7. Functions: registering and writing handlers
 
+### Naming
+
+Function IDs are `<worker>::<verb>` (e.g. `shell::exec`, `context::assemble`).
+Each segment is lowercase, and **multi-word segments use kebab-case, never
+snake_case** — `context::count-tokens` and `shell::on-config-change`, not
+`context::count_tokens`. The id is the public wire surface other workers and
+agents call, so renaming it later is a breaking change. The same rule applies
+to a trigger's target `function_id` and to any custom `trigger_type` string
+(§8).
+
 ### Layout
 
 One file per function (or small feature group) under `src/functions/`.
@@ -499,6 +509,11 @@ A function is callable via the SDK on its own. To make it callable from
 outside iii (HTTP, cron, queue, etc.) you also register a **trigger** that
 binds an external event source to the function.
 
+**Naming.** A trigger's `function_id` is the id the function was registered
+under, so it follows the same kebab-case rule as functions (§7 Naming). Any
+custom `trigger_type` string you introduce is also kebab-case (e.g.
+`row-change`), never snake_case.
+
 Trigger types in use across this repo and the SDK's built-ins:
 
 | Trigger type | Config keys | Used for |
diff --git a/docs/sops/new-worker.md b/docs/sops/new-worker.md
index 90c0ee72..17d72503 100644
--- a/docs/sops/new-worker.md
+++ b/docs/sops/new-worker.md
@@ -18,6 +18,9 @@ after §2.
 - Do **not** prefix with `iii-` unless the worker itself is named that way
   (e.g. `iii-directory`, `iii-lsp`).
 - Git release tags use the folder name: `<worker>/vX.Y.Z`.
+- **Function and trigger IDs** are `<worker>::<verb>`; multi-word segments use
+  kebab-case, never snake_case (e.g. `context::count-tokens`,
+  `shell::on-config-change`). See [`binary-worker.md`](binary-worker.md) §7.
 
 ## 2. Required files by deploy mode
 
diff --git a/harness/docs/workers/context-compaction.md b/harness/docs/workers/context-compaction.md
index 45dc9806..ae4c0d64 100644
--- a/harness/docs/workers/context-compaction.md
+++ b/harness/docs/workers/context-compaction.md
@@ -1,294 +1,106 @@
 # context-compaction
 
-Out-of-band session-history compactor (v2). Subscribes to the dedicated
-`agent::turn_end` stream (mirrored by the event producer) and summarises older
-turns when the session approaches the model's usable context limit — one wake
-per turn instead of one per `agent::events` frame. Also exposes a sync
-pre-turn path that the turn-orchestrator calls to compact before a turn that
-would overflow.
+A thin compatibility wrapper exposing the user-initiated **`/compact`** action
+(`context-compaction::compact_session`) on top of the standalone Rust
+**`context-manager`** worker. All real context logic — token counting,
+function-result pruning, and history summarisation — now lives in
+`context-manager` (`context::*`); see
+[context-manager/architecture/integration.md](../../../context-manager/architecture/integration.md).
+
+This worker exists only to preserve the console's existing `/compact` RPC
+contract. It owns no compaction policy and binds no triggers.
+
+> Migration note: the previous version was an out-of-band compactor with four
+> functions (`on_turn_end`, `compact_now`, `prune_tool_outputs`,
+> `compact_session`) plus a turn-orchestrator pre-flight (`preflight.ts`) and a
+> read-time provider-window reconstruction (`context-view.ts`). All of that is
+> retired. The harness now compacts **inline on the hot path** by calling
+> `context::assemble` once per turn (see
+> [turn-orchestrator.md](turn-orchestrator.md)); deciding *when* to compact
+> off the hot path is a future sibling, not this worker
+> (tech-specs/2026-06-agentic/harness.md § Out of scope).
+
+## Core invariant: the transcript is never modified
+
+`/compact` summarises the conversation **context**, not the stored
+conversation. No message entry is rewritten or deleted. The only write is an
+additive `kind:"custom"` bookkeeping entry (`custom_type: "compaction"`), which
+is invisible to `session::messages` (unless `include_custom`), never counted in
+`message_count`, and never matches a role filter. It exists so the next turn can
+anchor on the summary (`options.previous_summary`) and so the console can render
+its compaction marker.
+
+## Registered function
 
-## Purpose
-
-Keeps sessions alive indefinitely by summarising older turns into a compact
-Compaction entry. The next turn reads only the summary plus a configurable
-tail of recent turns. Two independent entry points share the same
-summarise-and-append core:
-
-- **Async** (`on_agent_event`) — fires after each `TurnEnd`; cheap if not
-  overflowing; does not block the turn in progress.
-- **Sync** (`compact_now`) — called by the turn-orchestrator pre-flight when a
-  turn would overflow; blocks until compaction completes, then reinjects the
-  user's message so the turn can proceed on a clean context.
-
-A third entry point (`prune_tool_outputs`) strips verbose tool outputs without
-summarisation — a cheaper first-pass before summarising.
-
-This worker is optional. Without it, sessions keep their full transcript.
-
-## Registered functions
-
-### `context-compaction::on_agent_event`
-
-Internal stream subscriber on `agent::turn_end` — fires once per turn (kept
-under the historical `on_agent_event` name).
-
-**Payload** (camelCase or snake_case envelope):
-```
-{ groupId | group_id: string, event: { data: EventObj } | data: EventObj }
-```
-
-Short-circuits unless the event is a `TurnEnd` carrying a `usage` object whose
-running token count (`input + output + cache_read + cache_write`) meets or
-exceeds `usable(model)`. When it does: acquires the compaction lease, runs
-`prune`, runs `summarizeAndAppend`, releases the lease.
-
-**Response:** `null`
-
----
-
-### `context-compaction::compact_now`
+### `context-compaction::compact_session`
 
-Sync pre-turn compaction called by the turn-orchestrator.
+User-initiated synchronous compaction. Invoked by the console `/compact` slash
+command.
 
 **Payload:**
 ```ts
 {
-  session_id: string;
-  projected_tokens: number;           // informational; not consumed in v2
-  last_user_message_id: string;       // entry_id of the user message to replay
-  model: {
-    id: string;
-    providerID: string;
-    limit: { context: number; input: number; output: number };
-  };
+  session_id: string;                 // required
+  model?: { id: string; providerID?: string;
+            limit?: { context: number; input?: number; output?: number } };
 }
 ```
 
-**Response** (discriminated union):
+**Response** (the console `CompactResult` wire shape — unchanged):
 ```ts
-| { status: 'ok';       tail_start_id: string | null; tokens_before: number; auto_continued: boolean }
-| { status: 'busy' }    // lease held by async path; caller should retry
-| { status: 'overflow'} // summariser itself failed with overflow
-| { status: 'empty' }   // session is empty; nothing to compact
+| { status: 'ok'; tail_start_id: string | null; tokens_before: number;
+    auto_continued: boolean; summary_text: string }
+| { status: 'busy' }      // a compaction lease is held; retry later
+| { status: 'empty' }     // nothing to compact
+| { status: 'overflow'; message: string }  // summariser unavailable / failed
+| { status: 'error'; message: string }     // model missing, or an unexpected failure
 ```
 
-Sequence: lease-with-wait → load messages → extract replay target →
-prune tool outputs → summarise → reinject user message → append synthetic
-"Continue…" prompt → release lease.
-
----
-
-### `context-compaction::prune_tool_outputs`
-
-Standalone prune without summarisation.
-
-**Payload:** `{ session_id: string }`
-
-**Response:** `{ pruned_tokens: number; pruned_parts: number; scanned_parts: number; busy?: true }`
-
-Acquires the `prune` lease (separate from `compaction` lease), walks
-`function_result` entries from newest to oldest, nulls out outputs that fall
-outside the `COMPACT_PRUNE_PROTECT` token window when freeing them would yield
-≥ `COMPACT_PRUNE_MIN_FREE` tokens. Returns immediately with `busy: true` if
-the prune lease is held.
-
----
-
-### `context-compaction::compact_session`
-
-User-initiated synchronous compaction. Intended for UI `/compact` actions.
-Takes only a `session_id`; resolves the model and last user message internally.
-
-**Payload:**
-```ts
-{ session_id: string }
-```
-
-**Response** (same discriminated union as `compact_now`):
-```ts
-| { status: 'ok';       tail_start_id: string | null; tokens_before: number; auto_continued: boolean }
-| { status: 'busy' }    // lease held; caller may retry
-| { status: 'overflow'} // summariser itself overflowed
-| { status: 'empty' }   // session is empty; nothing to compact
-```
-
-**Errors thrown** (non-2xx-style, propagated to caller):
-- `session_id is required` — when `session_id` is missing or empty.
-- `could not resolve model for session <id>` — when no assistant message with
-  provider/model fields exists in the session, or `models::get` fails.
+`auto_continued` is always `false` — `/compact` runs against a conversation at
+rest, so there is no in-flight user message to replay.
 
 **Sequence:**
-1. Resolves the model: explicit `payload.model.limit` → session-manager
-   assistant scan → orchestrator `run_request` → conservative fallback.
-   Avoids forcing `models::get` when the UI already knows the context window.
-2. Calls `handleSync` with `projected_tokens: 999_999` and
-   `last_user_message_id: ''` to force compaction unconditionally and
-   skip the replay / auto-continue branch. `/compact` runs against a
-   conversation at rest — there is no in-flight user message to re-inject.
-   Without this, single-turn sessions with one user message at index 0
-   collapsed to `truncatedMessages = []` and surfaced as `empty`.
-3. Returns the same `CompactNowResult` shape as `compact_now`, but
-   `auto_continued` is always `false` and no synthetic "Continue…" prompt
-   is appended.
-
-## Model-adaptive threshold
-
-The overflow threshold is not a flat constant. `usable(model)` computes the
-effective limit per-call:
-
-```
-usable = max(0, model.input_limit − COMPACT_RESERVED_TOKENS)
-```
-
-If `model.input_limit` is zero, it falls back to
-`model.context_window − model.output_tokens`.
-
-A session with a 200 k-token model reserves 20 k by default and triggers at
-180 k. A 32 k model triggers at 12 k with the same defaults.
-
-## Summarisation template
-
-Summaries follow a fixed Markdown template with eight sections:
-
-```
-## Goal
-## Constraints
-## Progress
-## Key Decisions
-## Tool Calls Made
-## Next Steps
-## Critical Context
-## Relevant Files
-```
-
-The system prompt requires the summariser to keep identifiers verbatim and
-fill every section. When a prior compaction exists, a second anchored prompt
-instructs the summariser to update the previous summary rather than start from
-scratch, so the summary converges rather than growing without bound.
-
-## Tool-output prune path
-
-`prune_tool_outputs` is separate from summarisation. It:
-
-1. Loads the active path via `session::messages` (paginated).
-2. Walks `function_result` entries from newest to oldest, skipping the two
-   most-recent user turns.
-3. Accumulates token estimates. Any output whose cumulative total exceeds
-   `COMPACT_PRUNE_PROTECT` goes into the prune queue.
-4. If the queue would free fewer than `COMPACT_PRUNE_MIN_FREE` tokens, it
-   skips entirely (no-op).
-5. Calls `session::update_message` per pruned entry, replacing the content
-   with `[output pruned]` and merging `{ compacted_at }` into the existing
-   details (one path load, one update per entry).
-
-Tools listed in `COMPACT_PRUNE_PROTECTED_TOOLS` are never pruned.
-
-## Replay + auto-continue (`compact_now` only)
-
-Only the turn-orchestrator overflow path uses replay. `compact_session`
-runs against a conversation at rest and passes `last_user_message_id: ''`
-so this branch is skipped — `auto_continued` is always `false` on that path.
-
-When `compact_now` runs:
-
-1. The entry matching `last_user_message_id` is extracted from the message
-   list before it is passed to the summariser (so it is not summarised away).
-2. That user message already sits on the active path as the compaction node's
-   parent, so it needs no reinjection — `context-view.ts` reconstructs the
-   window as `[summary, ...tail]`, which already ends with it.
-3. A synthetic user prompt ("Continue if you have next steps, or stop and
-   ask for clarification.") is appended via `session::append` (entry id
-   `<compaction_entry_id>-continue`, origin `{ synthetic: true }`) as a child
-   of the compaction so the model picks up where it left off.
-4. `CompactNowResult.auto_continued` is `true` when a replay target existed.
-
-## Backward compatibility
-
-Pre-v2 deployments may have free-form summaries (not Markdown-templated).
-Compaction records are read back as `custom_type: "compaction"` entries on
-the active path (`session::messages { include_custom: true }`). The last
-record's `summary` field is used as the `previousSummary` anchor regardless
-of its format, so old summaries are updated into the structured template on
-the next compaction cycle.
+1. `loadAssembleWindow(session_id)` reads the active path, finds the latest
+   compaction bookkeeping entry, and returns the post-compaction message window
+   (from `tail_start_entry_id` onward) plus its summary as `previousSummary`.
+2. `context::compact({ messages: window, model, options: { lease_key: session_id,
+   previous_summary } })` summarises the head, keeping a recent tail verbatim.
+   The summary is **returned**, not persisted — `context-manager` is
+   storage-agnostic.
+3. On `ok`, `persistCompactionRoundTrip` appends the compaction bookkeeping
+   entry (`{ summary, tokens_before, tail_start_entry_id }`), mapping the
+   worker's `tail_start_index` onto the entry id at that position in the window.
+4. The `context::compact` union is mapped onto the response above.
+
+The shared helpers live in
+[runtime/compaction.ts](../../src/runtime/compaction.ts); the wrapper is
+[context-compaction/register.ts](../../src/context-compaction/register.ts).
 
 ## Configuration
 
-All knobs are env-driven; no `config.yaml` fields are read.
-
-| Env var | Default | Purpose |
-|---|---|---|
-| `COMPACT_RESERVED_TOKENS` | `20000` | Tokens reserved for model output and overhead; subtracted from `model.input_limit` to derive the overflow threshold. |
-| `COMPACT_TAIL_TURNS` | `2` | Number of complete user+assistant turn pairs kept verbatim after the summary. |
-| `COMPACT_PRESERVE_RECENT_TOKENS` | _(adaptive)_ | Override the tail budget in tokens. When unset, defaults to 25% of `usable(model)`, clamped to [2 000, 8 000]. |
-| `COMPACT_PRUNE_PROTECT` | `40000` | Tokens of tool output to preserve from the tail before pruning. |
-| `COMPACT_PRUNE_MIN_FREE` | `20000` | Minimum tokens the prune pass must free; skips if below this. |
-| `COMPACT_TOOL_OUTPUT_MAX_CHARS` | `2000` | Per-output character cap applied before sending to the summariser. |
-| `COMPACT_BUSY_TIMEOUT_MS` | `30000` | Max ms `compact_now` / `compact_session` waits for the compaction lease before returning `{ status: 'busy' }`. Sized to cover a typical summariser stream (10–30s) so user-initiated `/compact` doesn't race the async TurnEnd path. |
-| `COMPACT_PRUNE_PROTECTED_TOOLS` | _(empty)_ | Comma-separated function IDs whose outputs are never pruned. |
-
-The summariser provider and model are always inherited from the session's
-own selection. Routing goes through `turn-orchestrator/provider-router`,
-so adding a provider there automatically covers `/compact`.
-
-## State scopes
-
-Compaction-related keys use dedicated scopes (key = `session_id`):
-
-| Scope | Purpose |
-|---|---|
-| `compaction_lease` | `{ nonce, ts }` — held for up to `LEASE_TTL_SECS = 300 s`. |
-| `prune_lease` | Same nonce-and-readback pattern, separate scope so the prune path does not block async compaction. |
-| `last_compaction_at` | Wall-clock ms of the most recent successful compaction. Stamped by `stampLastCompaction`. |
-
-Compaction appends a `custom_type: "compaction"` custom entry to session-manager (and optionally replay + synthetic continue). The turn FSM reconstructs the compacted provider window at read time via [context-view.ts](harness/src/turn-orchestrator/state-runtime/context-view.ts); there is no separate flat `messages` scope.
-
-## Observability
-
-Three OTel spans are emitted per invocation (no-op when OTel is not
-initialized):
-
-| Span name | Attributes |
-|---|---|
-| `compaction.async` | `session_id`, `tokens_before`, `used_prior_summary` |
-| `compaction.sync` | `session_id`, `tokens_before`, `replayed`, `auto_continued`, `lease_wait_ms` |
-| `compaction.prune` | `session_id`, `pruned_tokens`, `pruned_parts`, `scanned_parts` |
-
-`lease_wait_ms` measures the time spent waiting for the compaction lease in the
-sync path. All spans inherit `iii.session.id` from baggage when set by the
-outer `instrumentHandler` wrapper.
+This worker reads no configuration of its own. Token-budget knobs (reserve,
+tail turns, prune thresholds, lease TTL, summariser timeout) live in the
+`context-manager` worker's `config.yaml`; per-call options (`lease_key`,
+`previous_summary`) are passed by the wrapper.
 
 ## Dependencies
 
-session-manager endpoints used (via [runtime/session.ts](harness/src/runtime/session.ts)):
+Worker manifest deps (`iii.worker.yaml`): `session-manager ^0.1.0`,
+`context-manager ^0.1.0`.
+
+session-manager endpoints used (via [runtime/session.ts](../../src/runtime/session.ts)):
 
 | Endpoint | Purpose |
 |---|---|
-| `session::messages` | Load active path with entry IDs (paginated; `include_custom: true` for compaction records). |
-| `session::append` | Append the compaction custom entry (`custom_type: "compaction"`, data `{ summary, tokens_before, tail_start_id, details, timestamp }`) and the synthetic "Continue…" prompt. |
-| `session::update_message` | Replace pruned tool outputs in-place (`[output pruned]` + `details.compacted_at`). |
-| `models::get` | Resolve `context_window` / `max_output_tokens` for model-adaptive threshold. |
+| `session::messages` | Load the active path (paginated; `include_custom: true` to find the latest compaction entry). |
+| `session::append` | Append the additive compaction bookkeeping entry (`custom_type: "compaction"`, data `{ summary, tokens_before, tail_start_entry_id, details, timestamp }`). |
 
-Worker manifest deps (`iii.worker.yaml`):
-`session-manager ^0.1.0`, `provider-anthropic ^0.2.0`, `provider-openai ^0.2.0`.
+context-manager function used: `context::compact`.
 
 ## Source layout
 
 | File | Purpose |
 |---|---|
 | `src/context-compaction/main.ts` | Binary entry point (`iii-context-compaction`). |
-| `src/context-compaction/register.ts` | Registers the three functions + stream subscriber. |
-| `src/context-compaction/config.ts` | Reads all `COMPACT_*` env vars. |
-| `src/context-compaction/handler-async.ts` | Async TurnEnd path: envelope decode, overflow check, lease, prune, summarise. |
-| `src/context-compaction/handler-sync.ts` | Sync pre-turn path: lease-with-wait, extract replay, prune, summarise, reinject. |
-| `src/context-compaction/handler-pipeline.ts` | Shared prune → summarise pipeline used by both handlers. |
-| `src/context-compaction/flat-state.ts` | `buildSummaryMessage` helper for the compacted provider window. |
-| `src/context-compaction/model-resolver.ts` | Shared model-resolution helpers: `fetchModelLimit` (catalog lookup) and `resolveModelFromSession` (session-scan + catalog lookup). |
-| `src/context-compaction/prune.ts` | Tool-output pruning (`prune`). |
-| `src/context-compaction/summarize.ts` | `summarizeAndAppend`: load → select tail → summarise → append Compaction entry. |
-| `src/context-compaction/overflow.ts` | `usable`, `isOverflow`, `preserveRecentBudget` — model-adaptive math. |
-| `src/context-compaction/selection.ts` | `selectWithEntryIds`, `completedCompactions` — tail selection with entry ID tracking. |
-| `src/context-compaction/template.ts` | `SUMMARY_TEMPLATE` + `buildPrompt` — structured prompt construction. |
-| `src/context-compaction/replay.ts` | `extractReplayTarget` — locates the last user message so it is excluded from the summary on the sync path. |
-| `src/context-compaction/lease.ts` | `acquireLease`, `acquireLeaseWithWait`, `releaseLease`, `stampLastCompaction`. |
-| `src/context-compaction/stream-collect.ts` | Drives `provider::<name>::stream` via in-process channel and collects the final message. |
-| `src/context-compaction/strip-media.ts` | Strips images and truncates tool outputs before sending to the summariser. |
+| `src/context-compaction/register.ts` | Registers the single `compact_session` wrapper. |
+| `src/runtime/compaction.ts` | Shared round-trip helpers: `loadAssembleWindow`, `assembleContext`, `persistCompactionRoundTrip`, `compactWindow`, `modelInputFromConsole`. |
diff --git a/harness/docs/workers/turn-orchestrator.md b/harness/docs/workers/turn-orchestrator.md
index 812cd33c..780e2c38 100644
--- a/harness/docs/workers/turn-orchestrator.md
+++ b/harness/docs/workers/turn-orchestrator.md
@@ -87,9 +87,9 @@ Session-scoped iii state uses semantic scopes from
 | `run_request` | `<session_id>` | The `run::start` payload enriched by `provisioning` to include `function_schemas: [agentTriggerTool()]` and the assembled `system_prompt`. Typed as `RunRequest` ([run-request.ts](harness/src/turn-orchestrator/run-request.ts)). |
 | `event_counter` | `<session_id>` | Monotonic counter for `agent::events` sequence numbers. |
 
-Conversation history lives in the external session-manager worker (`session::*`) only. `TurnStore.loadMessages` reconstructs the provider-facing window from one paginated `session::messages { include_custom: true }` read (the latest `custom_type: "compaction"` entry on the path supplies the summary + tail boundary; empty assistant placeholders are filtered out) — see [context-view.ts](harness/src/turn-orchestrator/state-runtime/context-view.ts) and the shared client in [runtime/session.ts](harness/src/runtime/session.ts). `appendMessages` writes via `session::append`, chaining each message onto the active leaf with a deterministic `entry_id` (`<message_id>-user-<i>` for run::start seeds, `fr-<function_call_id>` for function results, `<runKey>-t<N>-assistant` for the per-turn assistant entry) — session-manager's idempotency on `entry_id` is what makes redelivered queue steps safe; there is no separate reconcile/repair path.
+Conversation history lives in the external session-manager worker (`session::*`) only. Before each provider call, `prepareStreamContext` builds the model-ready context: `loadAssembleWindow` reads the active path (`session::messages { include_custom: true }`), takes the latest `custom_type: "compaction"` bookkeeping entry as the summary anchor + tail boundary (filtering out empty assistant placeholders and this turn's just-appended assistant entry), then `context::assemble` on the [context-manager](../../../context-manager/architecture/integration.md) worker prunes/compacts the window to fit the model and renders any summary into the returned `system_prompt`. When assemble compacts, the harness persists the round trip via `persistCompactionRoundTrip` as an additive `kind:"custom"` bookkeeping entry — the durable transcript is never modified (only the model-facing context is compacted). The helpers live in [runtime/compaction.ts](harness/src/runtime/compaction.ts); the shared session client in [runtime/session.ts](harness/src/runtime/session.ts). `context-manager` is a **soft** dependency: when unavailable, assemble degrades to the raw window + base system prompt so the turn still runs. `appendMessages` writes via `session::append`, chaining each message onto the active leaf with a deterministic `entry_id` (`<message_id>-user-<i>` for run::start seeds, `fr-<function_call_id>` for function results, `<runKey>-t<N>-assistant` for the per-turn assistant entry) — session-manager's idempotency on `entry_id` is what makes redelivered queue steps safe; there is no separate reconcile/repair path.
 
-**Session-manager call discipline** (keeps the per-turn `session::` RPC count flat): `session::ensure` runs exactly once per run — `TurnStore.ensureSession` is called at the top of `run::start`, and `loadMessages`/`appendMessages` deliberately do **not** re-ensure (every later turn step is provably preceded by `run::start`, so re-ensuring on each read/write was pure overhead). Within `assistant_streaming` the just-appended assistant entry is excluded from the provider window on re-entry (`excludeEntryIds`) instead of relying on a trailing-message dup scan. **Status is driver-owned**: `run::start` flips the session to `working` before the first append; `finishSession` sets `done` (or `error` + reason when the last assistant stopped with `error`); `failTransition` sets `error` + the transition failure message. **`agent::events` carries ephemeral turn state only** (`turn_state_changed`, `function_execution_*`, `message_complete` for stop_reason notices, `turn_end`, `agent_end`); streaming deltas are no longer mirrored there — `session::message-updated` is the live token surface. `agent_end` is a turn-end **signal**: the transcript is re-read from session-manager on reload, so `finishSession`/`failTransition` emit `agent_end` with empty `messages` rather than reloading the whole session to fill a field no consumer reads.
+**Session-manager call discipline** (keeps the per-turn `session::` RPC count flat): `session::ensure` runs exactly once per run — `TurnStore.ensureSession` is called at the top of `run::start`, and the window load / `appendMessages` deliberately do **not** re-ensure (every later turn step is provably preceded by `run::start`, so re-ensuring on each read/write was pure overhead). Within `assistant_streaming` the just-appended assistant entry is excluded from the provider window on re-entry (`excludeEntryIds`) instead of relying on a trailing-message dup scan. **Status is driver-owned**: `run::start` flips the session to `working` before the first append; `finishSession` sets `done` (or `error` + reason when the last assistant stopped with `error`); `failTransition` sets `error` + the transition failure message. **`agent::events` carries ephemeral turn state only** (`turn_state_changed`, `function_execution_*`, `message_complete` for stop_reason notices, `turn_end`, `agent_end`); streaming deltas are no longer mirrored there — `session::message-updated` is the live token surface. `agent_end` is a turn-end **signal**: the transcript is re-read from session-manager on reload, so `finishSession`/`failTransition` emit `agent_end` with empty `messages` rather than reloading the whole session to fill a field no consumer reads.
 
 Keys that no longer exist: `function_prepared`, `function_executed`,
 `function_schemas` (standalone), `tool_prepared`, `tool_executed`,
@@ -157,9 +157,9 @@ From
 | [src/turn-orchestrator/function-awaiting-approval/process.ts](harness/src/turn-orchestrator/function-awaiting-approval/process.ts) | `turn::function_awaiting_approval` handler. |
 | [src/turn-orchestrator/steering-check/process.ts](harness/src/turn-orchestrator/steering-check/process.ts) | `turn::steering_check` handler. |
 | [src/turn-orchestrator/state.ts](harness/src/turn-orchestrator/state.ts) | `TurnState`, `TurnStateRecord`, `TurnWork`, `AwaitingApprovalEntry`, state-key helpers, `newRecord`, `transitionTo`. |
-| [src/turn-orchestrator/errors.ts](harness/src/turn-orchestrator/errors.ts) | `TransientError` (opt into queue retry), `ContextOverflowError`, `CompactionBusyError`. |
+| [src/turn-orchestrator/errors.ts](harness/src/turn-orchestrator/errors.ts) | `TransientError` (opt into queue retry), `TurnStateInvariantError`. |
 | [src/turn-orchestrator/events.ts](harness/src/turn-orchestrator/events.ts) | `emit(iii, sid, event)` — appends a sequenced `AgentEvent` to the `agent::events` stream. |
-| [src/turn-orchestrator/preflight.ts](harness/src/turn-orchestrator/preflight.ts) | `runPreflight` — context-compaction check before each provider call. |
+| [src/runtime/compaction.ts](harness/src/runtime/compaction.ts) | Compaction round-trip helpers (`loadAssembleWindow`, `assembleContext`, `persistCompactionRoundTrip`) — the `context::assemble` pre-flight before each provider call. |
 | [src/turn-orchestrator/provider-router.ts](harness/src/turn-orchestrator/provider-router.ts) | `decide` + `targetFunctionId` — pick `provider::<name>::stream` for the run's `provider` field. |
 | [src/turn-orchestrator/system-prompt.ts](harness/src/turn-orchestrator/system-prompt.ts) | `buildSystemPrompt` — assembles the system prompt (mode paragraph + engine-grounded identity preamble). |
 | [src/turn-orchestrator/iii.worker.yaml](harness/src/turn-orchestrator/iii.worker.yaml) | Worker manifest. |
diff --git a/harness/iii.worker.yaml b/harness/iii.worker.yaml
index 8054b744..50775a6b 100644
--- a/harness/iii.worker.yaml
+++ b/harness/iii.worker.yaml
@@ -19,3 +19,7 @@ dependencies:
   configuration: "^0.17.0"
   iii-directory: "^0.8.1"
   session-manager: "^0.1.0"
+  # Soft dependency: the turn loop degrades to raw history when absent
+  # (see harness pre-flight fallback in src/runtime/compaction.ts). Declared
+  # for install convenience; the runtime fallback is the actual contract.
+  context-manager: "^0.1.0"
diff --git a/harness/src/context-compaction/config.ts b/harness/src/context-compaction/config.ts
deleted file mode 100644
index b6241dfa..00000000
--- a/harness/src/context-compaction/config.ts
+++ /dev/null
@@ -1,60 +0,0 @@
-const DEFAULT_RESERVED_TOKENS = 20_000;
-const DEFAULT_TAIL_TURNS = 2;
-const DEFAULT_MIN_PRESERVE_RECENT_TOKENS = 2_000;
-const DEFAULT_MAX_PRESERVE_RECENT_TOKENS = 8_000;
-const DEFAULT_PRUNE_PROTECT = 40_000;
-const DEFAULT_PRUNE_MIN_FREE = 20_000;
-const DEFAULT_TOOL_OUTPUT_MAX_CHARS = 2_000;
-// Sized to cover a summariser stream (10-30s); shorter values surface
-// `busy` to users when async compaction is mid-flight.
-const DEFAULT_BUSY_TIMEOUT_MS = 30_000;
-
-export const MIN_PRESERVE_RECENT_TOKENS = DEFAULT_MIN_PRESERVE_RECENT_TOKENS;
-export const MAX_PRESERVE_RECENT_TOKENS = DEFAULT_MAX_PRESERVE_RECENT_TOKENS;
-
-export type CompactionConfig = Readonly<{
-  reservedTokens: number;
-  tailTurns: number;
-  preserveRecentTokensOverride: number | undefined;
-  pruneProtect: number;
-  pruneMinFree: number;
-  toolOutputMaxChars: number;
-  busyTimeoutMs: number;
-  pruneProtectedTools: string[];
-}>;
-
-function intEnv(name: string, def: number): number {
-  const v = process.env[name];
-  if (!v) return def;
-  const n = Number.parseInt(v, 10);
-  return Number.isFinite(n) && n > 0 ? n : def;
-}
-
-function readPreserveRecentTokensOverride(): number | undefined {
-  const v = process.env.COMPACT_PRESERVE_RECENT_TOKENS;
-  if (!v) return undefined;
-  const n = Number.parseInt(v, 10);
-  return Number.isFinite(n) && n > 0 ? n : undefined;
-}
-
-function readPruneProtectedTools(): string[] {
-  const v = process.env.COMPACT_PRUNE_PROTECTED_TOOLS;
-  if (!v) return [];
-  return v
-    .split(',')
-    .map((s) => s.trim())
-    .filter(Boolean);
-}
-
-export function compactionConfig(): CompactionConfig {
-  return {
-    reservedTokens: intEnv('COMPACT_RESERVED_TOKENS', DEFAULT_RESERVED_TOKENS),
-    tailTurns: intEnv('COMPACT_TAIL_TURNS', DEFAULT_TAIL_TURNS),
-    preserveRecentTokensOverride: readPreserveRecentTokensOverride(),
-    pruneProtect: intEnv('COMPACT_PRUNE_PROTECT', DEFAULT_PRUNE_PROTECT),
-    pruneMinFree: intEnv('COMPACT_PRUNE_MIN_FREE', DEFAULT_PRUNE_MIN_FREE),
-    toolOutputMaxChars: intEnv('COMPACT_TOOL_OUTPUT_MAX_CHARS', DEFAULT_TOOL_OUTPUT_MAX_CHARS),
-    busyTimeoutMs: intEnv('COMPACT_BUSY_TIMEOUT_MS', DEFAULT_BUSY_TIMEOUT_MS),
-    pruneProtectedTools: readPruneProtectedTools(),
-  };
-}
diff --git a/harness/src/context-compaction/flat-state.ts b/harness/src/context-compaction/flat-state.ts
deleted file mode 100644
index dbf7dc1a..00000000
--- a/harness/src/context-compaction/flat-state.ts
+++ /dev/null
@@ -1,24 +0,0 @@
-/**
- * Build the assistant summary message used in the compacted provider window.
- */
-
-import type { AssistantMessage } from '../types/agent-message.js';
-
-export function buildSummaryMessage(summary_text: string): AssistantMessage {
-  return {
-    role: 'assistant',
-    content: [
-      {
-        type: 'text',
-        text: `<conversation-summary>\n${summary_text}\n</conversation-summary>`,
-      },
-    ],
-    stop_reason: 'end',
-    error_message: null,
-    error_kind: null,
-    usage: null,
-    model: '',
-    provider: '',
-    timestamp: Date.now(),
-  };
-}
diff --git a/harness/src/context-compaction/handler-async.ts b/harness/src/context-compaction/handler-async.ts
deleted file mode 100644
index b09451d7..00000000
--- a/harness/src/context-compaction/handler-async.ts
+++ /dev/null
@@ -1,158 +0,0 @@
-/**
- * Async (turn_end-driven) compaction. Woken by a turn-orchestrator queue
- * message (`context-compaction::on_turn_end`) carrying a typed
- * `{ session_id, usage, provider, model, model_limit? }` payload.
- */
-
-import { setCurrentSpanAttribute, withSpan } from '@iii-dev/observability';
-import type { ISdk } from '../runtime/iii.js';
-import { logger } from '../runtime/otel.js';
-import { readActivePath } from '../runtime/session.js';
-import type { ModelContextLimit } from '../types/agent-event.js';
-import type { Usage } from '../types/stream-event.js';
-import { compactionConfig } from './config.js';
-import {
-  isSummarizeOk,
-  publishCompactionDone,
-  runSummarizeCompaction,
-} from './handler-pipeline.js';
-import { acquireLease, releaseLease } from './lease.js';
-import { fetchModelLimit } from './model-resolver.js';
-import { isOverflow } from './overflow.js';
-
-export type OnTurnEndPayload = {
-  session_id: string;
-  usage: Usage | null;
-  provider: string;
-  model: string;
-  /**
-   * Threaded from the turn record's `model_meta` when the catalog resolved at
-   * turn start; lets the compactor skip its `models::get` fetch.
-   */
-  model_limit?: ModelContextLimit;
-};
-
-/** Parse the queue payload enqueued by turn-orchestrator at turn_end. */
-export function parseOnTurnEnd(payload: unknown): OnTurnEndPayload | null {
-  if (!payload || typeof payload !== 'object') return null;
-  const obj = payload as Record<string, unknown>;
-  if (typeof obj.session_id !== 'string' || !obj.session_id) return null;
-  return {
-    session_id: obj.session_id,
-    usage: obj.usage && typeof obj.usage === 'object' ? (obj.usage as Usage) : null,
-    provider: typeof obj.provider === 'string' ? obj.provider : '',
-    model: typeof obj.model === 'string' ? obj.model : '',
-    ...(obj.model_limit && typeof obj.model_limit === 'object'
-      ? { model_limit: obj.model_limit as ModelContextLimit }
-      : {}),
-  };
-}
-
-type ResolvedModel = {
-  providerID: string;
-  modelID: string;
-  modelLimit: ModelContextLimit;
-} | null;
-
-/**
- * Resolve the model limit from the payload's provider/model, falling back to
- * the session's most recent assistant message when either is missing. When the
- * payload carries a threaded `model_limit` (catalog resolved at turn start),
- * the `models::get` fetch is skipped entirely.
- */
-export async function resolveModel(
-  iii: ISdk,
-  session_id: string,
-  provider: string,
-  model: string,
-  model_limit?: ModelContextLimit,
-): Promise<ResolvedModel> {
-  let providerID: string | null = provider || null;
-  let modelID: string | null = model || null;
-
-  if (providerID && modelID && model_limit) {
-    return { providerID, modelID, modelLimit: model_limit };
-  }
-
-  if (!providerID || !modelID) {
-    try {
-      const items = await readActivePath(iii, session_id, { roles: ['assistant'] });
-      for (let i = items.length - 1; i >= 0; i--) {
-        const m = items[i]?.message as Record<string, unknown> | undefined;
-        if (!m) continue;
-        if (!providerID && typeof m.provider === 'string' && m.provider) {
-          providerID = m.provider;
-        }
-        if (!modelID && typeof m.model === 'string' && m.model) {
-          modelID = m.model;
-        }
-        if (providerID && modelID) break;
-      }
-    } catch (err) {
-      logger.debug('handler-async: failed to load session messages for model resolution', {
-        session_id,
-        err: String(err),
-      });
-    }
-  }
-
-  if (!providerID || !modelID) return null;
-  return fetchModelLimit(iii, providerID, modelID);
-}
-
-export async function handleAsync(iii: ISdk, payload: unknown): Promise<void> {
-  return withSpan('compaction::async', {}, async () => {
-    const parsed = parseOnTurnEnd(payload);
-    if (!parsed?.usage) return;
-    const { session_id, usage } = parsed;
-
-    setCurrentSpanAttribute('session_id', session_id);
-
-    const model = await resolveModel(
-      iii,
-      session_id,
-      parsed.provider,
-      parsed.model,
-      parsed.model_limit,
-    );
-    if (!model) {
-      logger.debug('handler-async: could not resolve model; skipping overflow check', {
-        session_id,
-      });
-      return;
-    }
-
-    const tokens_before =
-      (usage.input ?? 0) + (usage.output ?? 0) + (usage.cache_read ?? 0) + (usage.cache_write ?? 0);
-    setCurrentSpanAttribute('tokens_before', tokens_before);
-
-    if (
-      !isOverflow({
-        tokens: usage,
-        model: { id: model.modelID, limit: model.modelLimit },
-        reserved: compactionConfig().reservedTokens,
-      })
-    ) {
-      return;
-    }
-
-    const nonce = await acquireLease(iii, session_id, 'compaction');
-    if (!nonce) {
-      logger.debug('handler-async: compaction lease held; skipping', { session_id });
-      return;
-    }
-
-    try {
-      const result = await runSummarizeCompaction(iii, session_id, { mode: 'async' }, model);
-
-      setCurrentSpanAttribute('used_prior_summary', isSummarizeOk(result));
-      if (isSummarizeOk(result)) {
-        await publishCompactionDone(iii, session_id, 'async', result);
-      }
-    } catch (err) {
-      logger.warn('handler-async: compaction failed', { session_id, err: String(err) });
-    } finally {
-      await releaseLease(iii, session_id, nonce, 'compaction');
-    }
-  });
-}
diff --git a/harness/src/context-compaction/handler-pipeline.ts b/harness/src/context-compaction/handler-pipeline.ts
deleted file mode 100644
index c2fa4cf6..00000000
--- a/harness/src/context-compaction/handler-pipeline.ts
+++ /dev/null
@@ -1,102 +0,0 @@
-/**
- * Shared prune → summarize path for sync and async compaction handlers.
- */
-
-import { logger } from '../runtime/otel.js';
-import type { ISdk } from '../runtime/iii.js';
-import { emit } from '../turn-orchestrator/events.js';
-import { compactionConfig } from './config.js';
-import type { ModelLimit } from './overflow.js';
-import { prune } from './prune.js';
-import {
-  type SummarizeOk,
-  type SummarizeOptions,
-  type SummarizeOutcome,
-  summarizeAndAppend,
-} from './summarize.js';
-
-export type CompactionMode = 'sync' | 'async';
-
-export interface CompactionDonePayload {
-  summary_text: string;
-  tokens_before: number;
-  compaction_entry_id: string;
-  /** First entry_id of the preserved tail; null when nothing was kept. */
-  tail_start_id: string | null;
-}
-
-/**
- * Best-effort: a publish failure is logged but never thrown — the
- * caller has already done the load-bearing work (tree compaction)
- * and the UI marker is a nice-to-have.
- */
-async function emitCompactionDone(
-  iii: ISdk,
-  session_id: string,
-  mode: CompactionMode,
-  payload: CompactionDonePayload,
-): Promise<void> {
-  try {
-    await emit(iii, session_id, {
-      type: 'compaction_done',
-      mode,
-      summary_text: payload.summary_text,
-      tokens_before: payload.tokens_before,
-      compaction_entry_id: payload.compaction_entry_id,
-      tail_start_id: payload.tail_start_id,
-    });
-  } catch (err) {
-    logger.warn(`handler-${mode}: compaction_done emit failed`, {
-      code: 'compaction_done_emit_failed',
-      session_id,
-      err: String(err),
-    });
-  }
-}
-
-export type CompactionModel = {
-  providerID: string;
-  modelID: string;
-  modelLimit: ModelLimit;
-};
-
-export async function pruneSessionToolOutputs(iii: ISdk, session_id: string): Promise<void> {
-  const cfg = compactionConfig();
-  await prune(iii, session_id, {
-    protectTokens: cfg.pruneProtect,
-    minFree: cfg.pruneMinFree,
-    protectedTools: cfg.pruneProtectedTools,
-  });
-}
-
-export async function runSummarizeCompaction(
-  iii: ISdk,
-  session_id: string,
-  options: SummarizeOptions,
-  model: CompactionModel,
-): Promise<SummarizeOutcome> {
-  await pruneSessionToolOutputs(iii, session_id);
-  return summarizeAndAppend(iii, session_id, options, {
-    providerID: model.providerID,
-    modelID: model.modelID,
-    modelLimit: model.modelLimit,
-  });
-}
-
-export async function publishCompactionDone(
-  iii: ISdk,
-  session_id: string,
-  mode: CompactionMode,
-  result: SummarizeOk,
-): Promise<void> {
-  await emitCompactionDone(iii, session_id, mode, {
-    summary_text: result.summary_text,
-    tokens_before: result.tokens_before,
-    compaction_entry_id: result.compaction_entry_id,
-    tail_start_id: result.tail_start_id,
-  });
-}
-
-export function isSummarizeOk(result: SummarizeOutcome): result is SummarizeOk {
-  return result !== 'empty' && result.kind === 'ok';
-}
diff --git a/harness/src/context-compaction/handler-sync.ts b/harness/src/context-compaction/handler-sync.ts
deleted file mode 100644
index 7045b183..00000000
--- a/harness/src/context-compaction/handler-sync.ts
+++ /dev/null
@@ -1,133 +0,0 @@
-/**
- * Sync (pre-turn) compaction handler — `compact_now`.
- *
- * Called by the turn-orchestrator before a turn that would overflow.
- * Sequence: lease → prune → summarise → reinject replay → auto-continue.
- */
-
-import { setCurrentSpanAttribute, withSpan } from '@iii-dev/observability';
-import type { ISdk } from '../runtime/iii.js';
-import { logger } from '../runtime/otel.js';
-import { messageItemsFromPath, readActivePath, sessionAppendMessage } from '../runtime/session.js';
-import { compactionConfig } from './config.js';
-import { publishCompactionDone, runSummarizeCompaction } from './handler-pipeline.js';
-import { acquireLeaseWithWait, releaseLease } from './lease.js';
-import type { ModelLimit } from './overflow.js';
-import { type MessageWithEntryId, extractReplayTarget } from './replay.js';
-
-export type CompactNowInput = {
-  session_id: string;
-  /** TODO(task-27): used by turn-orchestrator pre-flight to decide whether to call compact_now. Accepted now for forward-compat; not consumed by this handler. */
-  projected_tokens: number;
-  last_user_message_id: string;
-  model: { id: string; providerID: string; limit: ModelLimit };
-};
-
-export type CompactNowResult =
-  | {
-      status: 'ok';
-      tail_start_id: string | null;
-      tokens_before: number;
-      auto_continued: boolean;
-      summary_text: string;
-    }
-  | { status: 'busy' }
-  | { status: 'overflow'; message: string }
-  | { status: 'empty' };
-
-export async function handleSync(iii: ISdk, input: CompactNowInput): Promise<CompactNowResult> {
-  return withSpan('compaction.sync', {}, async () => {
-    setCurrentSpanAttribute('session_id', input.session_id);
-
-    const leaseStart = Date.now();
-    const nonce = await acquireLeaseWithWait(
-      iii,
-      input.session_id,
-      'compaction',
-      compactionConfig().busyTimeoutMs,
-    );
-    const lease_wait_ms = Date.now() - leaseStart;
-    setCurrentSpanAttribute('lease_wait_ms', lease_wait_ms);
-
-    if (!nonce) return { status: 'busy' };
-
-    try {
-      const entries: MessageWithEntryId[] = messageItemsFromPath(
-        await readActivePath(iii, input.session_id),
-      );
-
-      const { replay, truncatedMessages } = extractReplayTarget(
-        entries,
-        input.last_user_message_id,
-      );
-
-      setCurrentSpanAttribute('replayed', Boolean(replay));
-
-      const result = await runSummarizeCompaction(
-        iii,
-        input.session_id,
-        { mode: 'sync', truncatedEntries: truncatedMessages },
-        {
-          providerID: input.model.providerID,
-          modelID: input.model.id,
-          modelLimit: input.model.limit,
-        },
-      );
-
-      if (result === 'empty') return { status: 'empty' };
-      if (result.kind === 'compact') {
-        return { status: 'overflow', message: result.reason };
-      }
-
-      setCurrentSpanAttribute('tokens_before', result.tokens_before);
-
-      const auto_continued = Boolean(replay);
-      setCurrentSpanAttribute('auto_continued', auto_continued);
-
-      // The last user message is already on the path as the compaction's
-      // parent; just nudge the agent to continue past the summary.
-      if (replay) {
-        await sessionAppendMessage(iii, {
-          session_id: input.session_id,
-          message: {
-            role: 'user',
-            content: [
-              {
-                type: 'text',
-                text: 'Continue if you have next steps, or stop and ask for clarification.',
-              },
-            ],
-            timestamp: Date.now(),
-          },
-          // Deterministic per compaction record: a redelivered handler call
-          // can't double-append the nudge.
-          ...(result.compaction_entry_id
-            ? {
-                entry_id: `${result.compaction_entry_id}-continue`,
-                parent_id: result.compaction_entry_id,
-              }
-            : {}),
-          origin: { synthetic: true },
-        });
-      }
-
-      await publishCompactionDone(iii, input.session_id, 'sync', result);
-
-      return {
-        status: 'ok',
-        tail_start_id: result.tail_start_id,
-        tokens_before: result.tokens_before,
-        auto_continued,
-        summary_text: result.summary_text,
-      };
-    } catch (err) {
-      logger.warn('handler-sync: sync compaction failed', {
-        session_id: input.session_id,
-        err: String(err),
-      });
-      throw err;
-    } finally {
-      await releaseLease(iii, input.session_id, nonce, 'compaction');
-    }
-  });
-}
diff --git a/harness/src/context-compaction/iii.worker.yaml b/harness/src/context-compaction/iii.worker.yaml
index ee8f2058..101cad69 100644
--- a/harness/src/context-compaction/iii.worker.yaml
+++ b/harness/src/context-compaction/iii.worker.yaml
@@ -4,7 +4,7 @@ language: node
 deploy: binary
 manifest: package.json
 bin: iii-context-compaction
-description: Out-of-band session-history compactor. Subscribes to agent::events::TurnEnd, summarises older turns via provider::<name>::stream, and writes a session-manager compaction custom entry so the next turn reads a compressed transcript.
+description: Thin user-initiated /compact wrapper (context-compaction::compact_session) over the context-manager worker; summarises the post-compaction window via context::compact and persists the round trip as a session bookkeeping entry. The transcript is never modified.
 
 runtime:
   kind: node
@@ -15,5 +15,4 @@ scripts:
 
 dependencies:
   session-manager: "^0.1.0"
-  provider-anthropic: "^0.2.0"
-  provider-openai: "^0.2.0"
+  context-manager: "^0.1.0"
diff --git a/harness/src/context-compaction/lease.ts b/harness/src/context-compaction/lease.ts
deleted file mode 100644
index cc41edf6..00000000
--- a/harness/src/context-compaction/lease.ts
+++ /dev/null
@@ -1,71 +0,0 @@
-/**
- * Compaction/prune leases — thin (scope, ttl) adapters over the shared lease in
- * runtime/lease.ts. Maps a LeaseKind to a state scope and the 300s TTL.
- */
-
-import type { ISdk } from '../runtime/iii.js';
-import {
-  acquireLease as acquireLeaseShared,
-  acquireLeaseWithWait as acquireLeaseWithWaitShared,
-  mintLeaseNonce,
-  releaseLease as releaseLeaseShared,
-} from '../runtime/lease.js';
-import { stateSet } from '../runtime/state.js';
-
-const COMPACTION_LEASE_SCOPE = 'compaction_lease';
-const PRUNE_LEASE_SCOPE = 'prune_lease';
-const LAST_COMPACTION_AT_SCOPE = 'last_compaction_at';
-
-export type LeaseKind = 'compaction' | 'prune';
-
-export const LEASE_TTL_SECS = 300;
-const LEASE_TTL_MS = LEASE_TTL_SECS * 1000;
-
-export { mintLeaseNonce };
-
-function leaseScope(kind: LeaseKind): string {
-  return kind === 'compaction' ? COMPACTION_LEASE_SCOPE : PRUNE_LEASE_SCOPE;
-}
-
-/** Epoch-secs of a `{nonce, ts}` lease claim; 0 for any other shape. */
-export function readLeaseTimestampSecs(v: unknown): number {
-  if (!v || typeof v !== 'object') return 0;
-  const ts = (v as Record<string, unknown>).ts;
-  return typeof ts === 'number' ? Math.floor(ts / 1000) : 0;
-}
-
-export async function acquireLease(
-  iii: ISdk,
-  session_id: string,
-  kind: LeaseKind = 'compaction',
-): Promise<string | null> {
-  return acquireLeaseShared(iii, leaseScope(kind), session_id, LEASE_TTL_MS);
-}
-
-export async function releaseLease(
-  iii: ISdk,
-  session_id: string,
-  ourNonce: string,
-  kind: LeaseKind = 'compaction',
-): Promise<void> {
-  await releaseLeaseShared(iii, leaseScope(kind), session_id, ourNonce);
-}
-
-export async function acquireLeaseWithWait(
-  iii: ISdk,
-  session_id: string,
-  kind: LeaseKind,
-  totalTimeoutMs: number,
-): Promise<string | null> {
-  return acquireLeaseWithWaitShared(
-    iii,
-    leaseScope(kind),
-    session_id,
-    LEASE_TTL_MS,
-    totalTimeoutMs,
-  );
-}
-
-export async function stampLastCompaction(iii: ISdk, session_id: string): Promise<void> {
-  await stateSet(iii, LAST_COMPACTION_AT_SCOPE, session_id, Date.now());
-}
diff --git a/harness/src/context-compaction/main.ts b/harness/src/context-compaction/main.ts
index d6806876..5b1b06bf 100644
--- a/harness/src/context-compaction/main.ts
+++ b/harness/src/context-compaction/main.ts
@@ -5,6 +5,6 @@ import { register } from './register.js';
 await bootstrapWorker({
   name: 'context-compaction',
   description:
-    'Out-of-band session-history compactor. Subscribes to agent::events::TurnEnd and writes a session-manager compaction custom entry when the running token count crosses the configured threshold.',
+    'Thin user-initiated /compact wrapper (context-compaction::compact_session) over the context-manager worker; persists the compaction round trip as a session bookkeeping entry. The transcript is never modified.',
   register: (iii) => register(iii),
 });
diff --git a/harness/src/context-compaction/model-resolver.ts b/harness/src/context-compaction/model-resolver.ts
deleted file mode 100644
index 6b4f5ea6..00000000
--- a/harness/src/context-compaction/model-resolver.ts
+++ /dev/null
@@ -1,116 +0,0 @@
-import { RUN_REQUEST_SCOPE } from '../turn-orchestrator/state.js';
-import type { ISdk } from '../runtime/iii.js';
-import { logger } from '../runtime/otel.js';
-import { readActivePath } from '../runtime/session.js';
-import type { ModelLimit } from './overflow.js';
-
-export type ResolvedModel = {
-  providerID: string;
-  modelID: string;
-  modelLimit: ModelLimit;
-} | null;
-
-/**
- * Derive a {@link ModelLimit} from a catalog entry (a full {@link Model} or a
- * `models::get` result). The single mapping used by both the threaded
- * `model_meta` path and a fresh fetch, so they yield the exact same limit.
- */
-export function limitFromModel(m: {
-  context_window?: number;
-  max_output_tokens?: number;
-}): ModelLimit {
-  const context = typeof m.context_window === 'number' ? m.context_window : 0;
-  const output = typeof m.max_output_tokens === 'number' ? m.max_output_tokens : 0;
-  return { context, input: context, output };
-}
-
-export async function fetchModelLimit(
-  iii: ISdk,
-  providerID: string,
-  modelID: string,
-): Promise<ResolvedModel> {
-  try {
-    // router::models::get: payload key is `id`, result is wrapped as
-    // `{ model }`, null on a catalog miss (the cold-window signal).
-    const entry = await iii.trigger<
-      unknown,
-      {
-        model?: {
-          id?: string;
-          provider?: string;
-          context_window?: number;
-          max_output_tokens?: number;
-        } | null;
-      } | null
-    >({
-      function_id: 'router::models::get',
-      payload: { provider: providerID, id: modelID },
-      timeoutMs: 5_000,
-    });
-
-    const model = entry?.model ?? null;
-    if (!model) {
-      logger.debug('model-resolver: model not found in catalog', { providerID, modelID });
-      return null;
-    }
-
-    return {
-      providerID,
-      modelID,
-      modelLimit: limitFromModel(model),
-    };
-  } catch (err) {
-    logger.debug('model-resolver: router::models::get failed', {
-      providerID,
-      modelID,
-      err: String(err),
-    });
-    return null;
-  }
-}
-
-export async function resolveModelFromSession(
-  iii: ISdk,
-  session_id: string,
-): Promise<ResolvedModel> {
-  try {
-    const items = await readActivePath(iii, session_id, { roles: ['assistant'] });
-    let providerID: string | null = null;
-    let modelID: string | null = null;
-
-    for (let i = items.length - 1; i >= 0; i--) {
-      const m = items[i]?.message as Record<string, unknown> | undefined;
-      if (!m) continue;
-      if (!providerID && typeof m.provider === 'string' && m.provider) providerID = m.provider;
-      if (!modelID && typeof m.model === 'string' && m.model) modelID = m.model;
-      if (providerID && modelID) break;
-    }
-
-    if (!providerID || !modelID) return null;
-
-    return fetchModelLimit(iii, providerID, modelID);
-  } catch (err) {
-    logger.debug('model-resolver: session scan failed', { session_id, err: String(err) });
-    return null;
-  }
-}
-
-export async function resolveModelFromRunRequest(
-  iii: ISdk,
-  session_id: string,
-): Promise<ResolvedModel> {
-  try {
-    const req = await iii.trigger<unknown, { provider?: string; model?: string } | null>({
-      function_id: 'state::get',
-      payload: { scope: RUN_REQUEST_SCOPE, key: session_id },
-      timeoutMs: 5_000,
-    });
-    const providerID = typeof req?.provider === 'string' && req.provider ? req.provider : null;
-    const modelID = typeof req?.model === 'string' && req.model ? req.model : null;
-    if (!providerID || !modelID) return null;
-    return fetchModelLimit(iii, providerID, modelID);
-  } catch (err) {
-    logger.debug('model-resolver: run_request fetch failed', { session_id, err: String(err) });
-    return null;
-  }
-}
diff --git a/harness/src/context-compaction/overflow.ts b/harness/src/context-compaction/overflow.ts
deleted file mode 100644
index 3246100e..00000000
--- a/harness/src/context-compaction/overflow.ts
+++ /dev/null
@@ -1,65 +0,0 @@
-import type { ModelContextLimit } from '../types/agent-event.js';
-import {
-  MAX_PRESERVE_RECENT_TOKENS,
-  MIN_PRESERVE_RECENT_TOKENS,
-  compactionConfig,
-} from './config.js';
-
-/** A model's token limits. Single shape, defined in types/agent-event.ts. */
-export type ModelLimit = ModelContextLimit;
-
-export type ModelLike = {
-  id: string;
-  limit: ModelLimit;
-};
-
-export type TokensLike = {
-  input?: number;
-  output?: number;
-  cache_read?: number;
-  cache_write?: number;
-  total?: number;
-};
-
-export function usable(input: { model: ModelLike; reserved?: number }): number {
-  const { model } = input;
-  if (model.limit.context === 0) return 0;
-  const reserved = input.reserved ?? compactionConfig().reservedTokens;
-  const base =
-    model.limit.input > 0
-      ? Math.max(0, model.limit.input - reserved)
-      : Math.max(0, model.limit.context - model.limit.output);
-  return base;
-}
-
-export function isOverflow(input: {
-  tokens: TokensLike;
-  model: ModelLike;
-  reserved?: number;
-  autoDisabled?: boolean;
-}): boolean {
-  if (input.autoDisabled) return false;
-  if (input.model.limit.context === 0) return false;
-  const count =
-    typeof input.tokens.total === 'number' && input.tokens.total > 0
-      ? input.tokens.total
-      : (input.tokens.input ?? 0) +
-        (input.tokens.output ?? 0) +
-        (input.tokens.cache_read ?? 0) +
-        (input.tokens.cache_write ?? 0);
-  return count >= usable({ model: input.model, reserved: input.reserved });
-}
-
-export function preserveRecentBudget(input: {
-  model: ModelLike;
-  reserved?: number;
-  override?: number;
-}): number {
-  const ovr = input.override ?? compactionConfig().preserveRecentTokensOverride;
-  if (ovr !== undefined) return ovr;
-  const u = usable({ model: input.model, reserved: input.reserved });
-  return Math.min(
-    MAX_PRESERVE_RECENT_TOKENS,
-    Math.max(MIN_PRESERVE_RECENT_TOKENS, Math.floor(u * 0.25)),
-  );
-}
diff --git a/harness/src/context-compaction/prompts/compaction.txt b/harness/src/context-compaction/prompts/compaction.txt
deleted file mode 100644
index 75b34f07..00000000
--- a/harness/src/context-compaction/prompts/compaction.txt
+++ /dev/null
@@ -1,27 +0,0 @@
-You are an anchored context summarization assistant for an iii harness coding session.
-
-Summarize only the conversation history you are given. The newest turns are kept verbatim outside your summary, so focus on the older context that still matters for continuing the work.
-
-If the prompt includes a <previous-summary> block, treat it as the current anchored summary. Update it with the new history by preserving still-true details, removing stale details, and merging in new facts. Never drop a still-true detail just because the new turns don't mention it.
-
-Follow this exact output structure. Keep every section heading even if a section is empty (write `(none)`). Prefer terse bullets to paragraphs. Preserve exact file paths, bus function ids (e.g. `shell::fs::read`, `agent_trigger`, `state::set`), session ids, call_ids, and other identifiers verbatim. Target a total length under 10% of the original conversation token count.
-
-# Goal
-One or two bullets describing what the user is actually trying to accomplish in this session.
-
-# Worked on so far
-Chronological bullets of the meaningful actions taken. Each bullet should name the agent_trigger function used and the salient argument or result (path, command, summary of effect). Skip routine reads and noise.
-
-# Files touched
-Two sub-lists: `read:` (paths the agent inspected) and `modified:` (paths the agent created or edited).
-
-# Open questions
-Anything the user asked or implied that has not yet been answered. Keep verbatim where possible.
-
-# Plan state
-Where the agent currently is in any multi-step plan: which step just completed, which is next.
-
-# Identifiers worth remembering
-Session ids, sandbox ids, branch names, ticket ids, urls, call_ids, anything the agent will need to refer to later. One per bullet, prefixed with what kind of id it is.
-
-Do not answer the conversation itself. Do not mention that you are summarising, compacting, or merging context. Respond in the same language as the conversation.
diff --git a/harness/src/context-compaction/prune.ts b/harness/src/context-compaction/prune.ts
deleted file mode 100644
index 2dcc2288..00000000
--- a/harness/src/context-compaction/prune.ts
+++ /dev/null
@@ -1,108 +0,0 @@
-import { setCurrentSpanAttribute, withSpan } from '@iii-dev/observability';
-import type { ISdk } from '../runtime/iii.js';
-import { logger } from '../runtime/otel.js';
-import { messageItemsFromPath, readActivePath, sessionUpdateMessage } from '../runtime/session.js';
-import type { AgentMessage } from '../types/agent-message.js';
-
-export type PruneOptions = {
-  protectTokens: number;
-  minFree: number;
-  protectedTools: string[];
-};
-
-export type PruneResult = {
-  pruned_tokens: number;
-  pruned_parts: number;
-  scanned_parts: number;
-};
-
-function tokensOfText(text: string): number {
-  return Math.floor(text.length / 4);
-}
-
-export async function prune(
-  iii: ISdk,
-  session_id: string,
-  opts: PruneOptions,
-): Promise<PruneResult> {
-  return withSpan('compaction.prune', {}, async () => {
-    setCurrentSpanAttribute('session_id', session_id);
-
-    const entries = messageItemsFromPath(await readActivePath(iii, session_id));
-    if (entries.length === 0) {
-      setCurrentSpanAttribute('pruned_tokens', 0);
-      setCurrentSpanAttribute('pruned_parts', 0);
-      setCurrentSpanAttribute('scanned_parts', 0);
-      return { pruned_tokens: 0, pruned_parts: 0, scanned_parts: 0 };
-    }
-
-    let total = 0;
-    let scanned = 0;
-    let userTurns = 0;
-    const queue: Array<{ entry_id: string; tokens: number; message: AgentMessage }> = [];
-
-    for (let i = entries.length - 1; i >= 0; i--) {
-      const e = entries[i];
-      if (!e) continue;
-      if (e.message.role === 'user') {
-        userTurns++;
-        continue;
-      }
-      if (userTurns < 2) continue;
-      if (e.message.role !== 'function_result') continue;
-      if (opts.protectedTools.includes(e.message.function_id)) continue;
-      const text = e.message.content
-        .filter(
-          (b): b is { type: 'text'; text: string } => (b as { type?: string }).type === 'text',
-        )
-        .map((b) => b.text)
-        .join('');
-      const t = tokensOfText(text);
-      scanned++;
-      total += t;
-      if (total <= opts.protectTokens) continue;
-      queue.push({ entry_id: e.entry_id, tokens: t, message: e.message });
-    }
-
-    const pruned_tokens = queue.reduce((a, q) => a + q.tokens, 0);
-    if (pruned_tokens < opts.minFree) {
-      logger.debug('prune skipped (not enough to free)', { session_id, pruned_tokens, scanned });
-      setCurrentSpanAttribute('pruned_tokens', 0);
-      setCurrentSpanAttribute('pruned_parts', 0);
-      setCurrentSpanAttribute('scanned_parts', scanned);
-      return { pruned_tokens: 0, pruned_parts: 0, scanned_parts: scanned };
-    }
-
-    const compacted_at = Date.now();
-    for (const q of queue) {
-      // update_message replaces `details` wholesale for function_result
-      // entries, so merge the existing details with the compaction stamp.
-      const existing =
-        q.message.role === 'function_result' &&
-        q.message.details &&
-        typeof q.message.details === 'object'
-          ? (q.message.details as Record<string, unknown>)
-          : {};
-      try {
-        await sessionUpdateMessage(iii, {
-          session_id,
-          entry_id: q.entry_id,
-          content: [{ type: 'text', text: '[output pruned]' }],
-          details: { ...existing, compacted_at },
-        });
-      } catch (err) {
-        logger.warn('prune: session::update_message failed; skipping entry', {
-          session_id,
-          entry_id: q.entry_id,
-          err: String(err),
-        });
-      }
-    }
-
-    logger.info('pruned tool outputs', { session_id, pruned_tokens, pruned_parts: queue.length });
-    setCurrentSpanAttribute('pruned_tokens', pruned_tokens);
-    setCurrentSpanAttribute('pruned_parts', queue.length);
-    setCurrentSpanAttribute('scanned_parts', scanned);
-    return { pruned_tokens, pruned_parts: queue.length, scanned_parts: scanned };
-  });
-}
diff --git a/harness/src/context-compaction/register.ts b/harness/src/context-compaction/register.ts
index 2f7cbddc..a94bccf9 100644
--- a/harness/src/context-compaction/register.ts
+++ b/harness/src/context-compaction/register.ts
@@ -1,181 +1,101 @@
+/**
+ * `context-compaction::compact_session` — the thin user-initiated `/compact`
+ * wrapper kept for the console's existing RPC contract.
+ *
+ * All real context logic now lives in the standalone Rust `context-manager`
+ * worker (`context::*`). This handler:
+ *   1. loads the post-compaction window + anchored summary from session-manager,
+ *   2. calls `context::compact` (storage-agnostic — it only returns a summary),
+ *   3. persists the compaction round trip as an additive `kind:"custom"`
+ *      bookkeeping entry (the transcript itself is never modified),
+ * and maps the `context::compact` union back onto the console's `CompactResult`
+ * wire shape (`status` / `tail_start_id` / `tokens_before` / `summary_text` /
+ * `auto_continued`).
+ */
+
 import type { ISdk } from '../runtime/iii.js';
 import { logger } from '../runtime/otel.js';
-import { compactionConfig } from './config.js';
-import { handleAsync } from './handler-async.js';
-import { type CompactNowInput, handleSync } from './handler-sync.js';
-import { acquireLease, releaseLease } from './lease.js';
 import {
-  fetchModelLimit,
-  resolveModelFromRunRequest,
-  resolveModelFromSession,
-} from './model-resolver.js';
-import { prune } from './prune.js';
+  type AppliedReport,
+  compactWindow,
+  loadAssembleWindow,
+  modelInputFromConsole,
+  persistCompactionRoundTrip,
+} from '../runtime/compaction.js';
 
-// Sized so preserveRecentBudget clamps to its 2k minimum when the real
-// model is unknown — compaction is best-effort, not fatal.
-const FALLBACK_MODEL_LIMIT = {
-  context: 32_000,
-  input: 32_000,
-  output: 4_000,
-};
-const FALLBACK_MODEL_ID = 'unknown';
-const FALLBACK_PROVIDER_ID = 'unknown';
+type CompactSessionResult =
+  | {
+      status: 'ok';
+      tail_start_id: string | null;
+      tokens_before: number;
+      auto_continued: boolean;
+      summary_text: string;
+    }
+  | { status: 'busy' }
+  | { status: 'empty' }
+  | { status: 'overflow'; message: string }
+  | { status: 'error'; message: string };
 
-// payload.model: { id, providerID } with optional limit; null on malformed.
-async function resolveExplicitModel(
-  iii: ISdk,
-  raw: unknown,
-): Promise<{
-  providerID: string;
-  modelID: string;
-  modelLimit: { context: number; input: number; output: number };
-} | null> {
-  if (!raw || typeof raw !== 'object') return null;
-  const m = raw as Record<string, unknown>;
-  const providerID = typeof m.providerID === 'string' && m.providerID ? m.providerID : null;
-  const modelID = typeof m.id === 'string' && m.id ? m.id : null;
-  if (!providerID || !modelID) return null;
-  const lim = m.limit as { context?: number; input?: number; output?: number } | undefined;
-  if (lim && typeof lim.context === 'number' && lim.context > 0) {
-    return {
-      providerID,
-      modelID,
-      modelLimit: {
-        context: lim.context,
-        input: typeof lim.input === 'number' ? lim.input : lim.context,
-        output: typeof lim.output === 'number' ? lim.output : 0,
-      },
-    };
+async function compactSession(iii: ISdk, payload: unknown): Promise<CompactSessionResult> {
+  const obj = (payload ?? {}) as Record<string, unknown>;
+  const session_id = typeof obj.session_id === 'string' ? obj.session_id : '';
+  if (!session_id) {
+    throw new Error('context-compaction::compact_session: session_id is required');
   }
-  return fetchModelLimit(iii, providerID, modelID);
-}
 
-export async function register(iii: ISdk): Promise<void> {
-  iii.registerFunction(
-    'context-compaction::on_turn_end',
-    async (payload: unknown) => {
-      await handleAsync(iii, payload);
-      return null;
-    },
-    {
-      description:
-        'Internal: woken by a turn-orchestrator queue message at turn_end; triggers async compaction when running tokens exceed usable(model).',
-    },
-  );
-
-  iii.registerFunction(
-    'context-compaction::compact_now',
-    async (payload: unknown) => {
-      const obj = (payload ?? {}) as Record<string, unknown>;
+  const model = modelInputFromConsole(obj.model);
+  if (!model) {
+    return { status: 'error', message: 'model { id, providerID } is required' };
+  }
 
-      const session_id_raw = obj.session_id;
-      if (typeof session_id_raw !== 'string' || session_id_raw.length === 0) {
-        throw new Error('context-compaction::compact_now: session_id is required');
-      }
-      const session_id = session_id_raw;
+  try {
+    const { window, previousSummary } = await loadAssembleWindow(iii, session_id);
+    const result = await compactWindow(iii, session_id, window, model, previousSummary);
 
-      const modelObj = (obj.model ?? {}) as {
-        id?: string;
-        providerID?: string;
-        limit?: { context?: number; input?: number; output?: number };
-      };
-      const input: CompactNowInput = {
-        session_id,
-        projected_tokens: typeof obj.projected_tokens === 'number' ? obj.projected_tokens : 0,
-        last_user_message_id: String(obj.last_user_message_id ?? ''),
-        model: {
-          id: String(modelObj.id ?? ''),
-          providerID: String(modelObj.providerID ?? ''),
-          limit: {
-            context: modelObj.limit?.context ?? 0,
-            input: modelObj.limit?.input ?? 0,
-            output: modelObj.limit?.output ?? 0,
-          },
-        },
+    if (result.status === 'busy') return { status: 'busy' };
+    if (result.status === 'empty') return { status: 'empty' };
+    if (result.status === 'overflow') {
+      return {
+        status: 'overflow',
+        message: 'compaction unavailable (no llm-router or the summariser failed)',
       };
-      return handleSync(iii, input);
-    },
-    {
-      description:
-        'Sync pre-turn compaction triggered by turn-orchestrator when a turn would overflow. Performs prune+summarise+reinject+continue.',
-    },
-  );
+    }
 
-  iii.registerFunction(
-    'context-compaction::prune_tool_outputs',
-    async (payload: unknown) => {
-      const obj = (payload ?? {}) as Record<string, unknown>;
-      const session_id_raw = obj.session_id;
-      if (typeof session_id_raw !== 'string' || session_id_raw.length === 0) {
-        throw new Error('context-compaction::prune_tool_outputs: session_id is required');
-      }
-      const session_id = session_id_raw;
-      const nonce = await acquireLease(iii, session_id, 'prune');
-      if (!nonce) return { pruned_tokens: 0, pruned_parts: 0, scanned_parts: 0, busy: true };
-      try {
-        const cfg = compactionConfig();
-        return await prune(iii, session_id, {
-          protectTokens: cfg.pruneProtect,
-          minFree: cfg.pruneMinFree,
-          protectedTools: cfg.pruneProtectedTools,
-        });
-      } finally {
-        await releaseLease(iii, session_id, nonce, 'prune');
-      }
-    },
-    { description: 'Prune older tool outputs without summarisation (cheap path).' },
-  );
+    // ok — persist the round trip so the next turn anchors on this summary.
+    const applied: AppliedReport = {
+      pruned: false,
+      pruned_tokens: 0,
+      compacted: true,
+      summary: result.summary,
+      tail_start_index: result.tail_start_index,
+      tokens_before: result.tokens_before,
+    };
+    await persistCompactionRoundTrip(iii, session_id, applied, window);
+
+    const idx = result.tail_start_index;
+    const tailEntry =
+      typeof idx === 'number' && idx >= 0 && idx < window.length ? window[idx] : undefined;
+    return {
+      status: 'ok',
+      tail_start_id: tailEntry ? tailEntry.entry_id : null,
+      tokens_before: result.tokens_before,
+      // /compact runs against a conversation at rest; no synthetic continue.
+      auto_continued: false,
+      summary_text: result.summary,
+    };
+  } catch (err) {
+    logger.warn('compact_session failed', { session_id, err: String(err) });
+    return { status: 'error', message: err instanceof Error ? err.message : String(err) };
+  }
+}
 
+export async function register(iii: ISdk): Promise<void> {
   iii.registerFunction(
     'context-compaction::compact_session',
-    async (payload: unknown) => {
-      const obj = (payload ?? {}) as Record<string, unknown>;
-      const session_id_raw = obj.session_id;
-      if (typeof session_id_raw !== 'string' || session_id_raw.length === 0) {
-        throw new Error('context-compaction::compact_session: session_id is required');
-      }
-      const session_id = session_id_raw;
-
-      const explicitObj = obj.model as { id?: string; providerID?: string } | undefined;
-      let model = await resolveExplicitModel(iii, obj.model);
-      if (!model) model = await resolveModelFromSession(iii, session_id);
-      if (!model) model = await resolveModelFromRunRequest(iii, session_id);
-      if (!model) {
-        const providedId =
-          typeof explicitObj?.id === 'string' && explicitObj.id ? explicitObj.id : null;
-        const providedProvider =
-          typeof explicitObj?.providerID === 'string' && explicitObj.providerID
-            ? explicitObj.providerID
-            : null;
-        logger.warn('compact_session: model resolution failed; using fallback limit', {
-          session_id,
-          requestedProvider: providedProvider,
-          requestedModel: providedId,
-        });
-        model = {
-          providerID: providedProvider ?? FALLBACK_PROVIDER_ID,
-          modelID: providedId ?? FALLBACK_MODEL_ID,
-          modelLimit: FALLBACK_MODEL_LIMIT,
-        };
-      }
-
-      // Empty last_user_message_id skips handleSync's replay branch.
-      // Replay belongs to compact_now (orchestrator overflow); /compact
-      // runs against a conversation at rest with nothing to re-inject.
-      return handleSync(iii, {
-        session_id,
-        projected_tokens: 999_999,
-        last_user_message_id: '',
-        model: {
-          id: model.modelID,
-          providerID: model.providerID,
-          limit: model.modelLimit,
-        },
-      });
-    },
+    async (payload: unknown) => compactSession(iii, payload),
     {
       description:
-        'User-initiated synchronous compaction of a session. Required: session_id. Optional: model { id, providerID, limit? } to skip auto-resolution. If model is omitted, falls back to (1) most recent assistant message in session-manager, (2) orchestrator run_request.',
+        'User-initiated synchronous compaction of a session. Required: session_id. Optional: model { id, providerID, limit? }. Summarises the post-compaction window via context::compact and persists the compaction round trip as a custom bookkeeping entry; the transcript is left untouched.',
     },
   );
 }
diff --git a/harness/src/context-compaction/replay.ts b/harness/src/context-compaction/replay.ts
deleted file mode 100644
index 1af5280b..00000000
--- a/harness/src/context-compaction/replay.ts
+++ /dev/null
@@ -1,17 +0,0 @@
-import type { AgentMessage } from '../types/agent-message.js';
-
-export type MessageWithEntryId = { entry_id: string; message: AgentMessage };
-
-export function extractReplayTarget(
-  entries: MessageWithEntryId[],
-  lastUserMessageId: string,
-): { replay?: MessageWithEntryId; truncatedMessages: MessageWithEntryId[] } {
-  const idx = entries.findIndex((e) => e.entry_id === lastUserMessageId);
-  if (idx === -1) return { truncatedMessages: entries };
-  const entry = entries[idx];
-  if (entry?.message.role !== 'user') return { truncatedMessages: entries };
-  return {
-    replay: entry,
-    truncatedMessages: entries.slice(0, idx),
-  };
-}
diff --git a/harness/src/context-compaction/selection.ts b/harness/src/context-compaction/selection.ts
deleted file mode 100644
index 7151416f..00000000
--- a/harness/src/context-compaction/selection.ts
+++ /dev/null
@@ -1,149 +0,0 @@
-// Token-aware tail selection. Pure logic, no I/O.
-
-import type { AgentMessage } from '../types/agent-message.js';
-
-export type Turn = {
-  start: number;
-  end: number; // exclusive
-  index: number; // index of the user message
-};
-
-export type Tail = {
-  start: number;
-  index: number;
-};
-
-export type CompletedCompaction = {
-  /** Index of the user message that produced this compaction (or -1 if unknown). */
-  userIndex: number;
-  /** Index of the assistant summary message (or -1 if unknown). */
-  assistantIndex: number;
-  /** The summary text. */
-  summary: string;
-};
-
-export type SelectInput = {
-  messages: AgentMessage[];
-  budget: number;
-  tailTurns: number;
-  estimate: (messages: AgentMessage[]) => number;
-};
-
-export type SelectResult = {
-  head: AgentMessage[];
-  tail_start_index: number | undefined;
-};
-
-export type CompactionEntryLike = {
-  id: string;
-  summary: string;
-  tokens_before: number;
-  timestamp: number;
-};
-
-export type MessageWithEntryId = {
-  entry_id: string;
-  message: AgentMessage;
-};
-
-export type SelectWithEntryIdsInput = {
-  entries: MessageWithEntryId[];
-  budget: number;
-  tailTurns: number;
-  estimate: (messages: AgentMessage[]) => number;
-};
-
-export type SelectWithEntryIdsResult = {
-  head: MessageWithEntryId[];
-  tail_start_id: string | undefined;
-};
-
-export function turns(messages: AgentMessage[]): Turn[] {
-  const result: Turn[] = [];
-  for (let i = 0; i < messages.length; i++) {
-    if (messages[i]?.role !== 'user') continue;
-    result.push({ index: i, start: i, end: messages.length });
-  }
-  for (let i = 0; i < result.length - 1; i++) {
-    const current = result[i];
-    const nextStart = result[i + 1]?.start;
-    if (current && nextStart !== undefined) current.end = nextStart;
-  }
-  return result;
-}
-
-export function splitTurn(input: {
-  messages: AgentMessage[];
-  turn: Turn;
-  budget: number;
-  estimate: (m: AgentMessage[]) => number;
-}): Tail | undefined {
-  if (input.budget <= 0) return undefined;
-  if (input.turn.end - input.turn.start <= 1) return undefined;
-  for (let start = input.turn.start + 1; start < input.turn.end; start++) {
-    const size = input.estimate(input.messages.slice(start, input.turn.end));
-    if (size > input.budget) continue;
-    return { start, index: start };
-  }
-  return undefined;
-}
-
-export function select(input: SelectInput): SelectResult {
-  if (input.tailTurns <= 0) return { head: input.messages, tail_start_index: undefined };
-  const all = turns(input.messages);
-  if (all.length === 0) return { head: input.messages, tail_start_index: undefined };
-  const recent = all.slice(-input.tailTurns);
-  const sizes = recent.map((t) => input.estimate(input.messages.slice(t.start, t.end)));
-
-  let total = 0;
-  let keep: Tail | undefined;
-  for (let i = recent.length - 1; i >= 0; i--) {
-    const turn = recent[i];
-    const size = sizes[i];
-    if (turn === undefined || size === undefined) continue;
-    if (total + size <= input.budget) {
-      total += size;
-      keep = { start: turn.start, index: turn.index };
-      continue;
-    }
-    const remaining = input.budget - total;
-    const split = splitTurn({
-      messages: input.messages,
-      turn,
-      budget: remaining,
-      estimate: input.estimate,
-    });
-    if (split) keep = split;
-    break;
-  }
-  if (!keep || keep.start === 0) {
-    return { head: input.messages, tail_start_index: undefined };
-  }
-  return {
-    head: input.messages.slice(0, keep.start),
-    tail_start_index: keep.start,
-  };
-}
-
-export function selectWithEntryIds(input: SelectWithEntryIdsInput): SelectWithEntryIdsResult {
-  const messages = input.entries.map((e) => e.message);
-  const r = select({
-    messages,
-    budget: input.budget,
-    tailTurns: input.tailTurns,
-    estimate: input.estimate,
-  });
-  const headCount = r.head.length;
-  const head = input.entries.slice(0, headCount);
-  const tail_start_id =
-    r.tail_start_index !== undefined ? input.entries[r.tail_start_index]?.entry_id : undefined;
-  return { head, tail_start_id };
-}
-
-export function completedCompactions(entries: CompactionEntryLike[]): CompletedCompaction[] {
-  return entries
-    .filter((e) => e.summary && e.summary.length > 0)
-    .slice()
-    .sort((a, b) => a.timestamp - b.timestamp)
-    .map((e) => ({ summary: e.summary, userIndex: -1, assistantIndex: -1 }));
-}
diff --git a/harness/src/context-compaction/strip-media.ts b/harness/src/context-compaction/strip-media.ts
deleted file mode 100644
index 3f7fbe89..00000000
--- a/harness/src/context-compaction/strip-media.ts
+++ /dev/null
@@ -1,33 +0,0 @@
-import type { AgentMessage } from '../types/agent-message.js';
-import type { ContentBlock } from '../types/content.js';
-
-export function stripMedia(
-  messages: AgentMessage[],
-  opts: { toolOutputMaxChars: number },
-): AgentMessage[] {
-  return messages.map((m) => {
-    if (m.role === 'function_result') {
-      return {
-        ...m,
-        content: m.content
-          .map(stripImageBlock)
-          .map((b) => truncateBlock(b, opts.toolOutputMaxChars)),
-      };
-    }
-    return { ...m, content: m.content.map(stripImageBlock) } as AgentMessage;
-  });
-}
-
-function stripImageBlock(b: ContentBlock): ContentBlock {
-  if ((b as { type?: string }).type === 'image') {
-    return { type: 'text', text: '[image stripped]' } as ContentBlock;
-  }
-  return b;
-}
-
-function truncateBlock(b: ContentBlock, max: number): ContentBlock {
-  if ((b as { type?: string }).type !== 'text') return b;
-  const text = (b as { type: 'text'; text: string }).text;
-  if (text.length <= max) return b;
-  return { type: 'text', text: `${text.slice(0, max)}... [truncated]` } as ContentBlock;
-}
diff --git a/harness/src/context-compaction/summarize.ts b/harness/src/context-compaction/summarize.ts
deleted file mode 100644
index 9d815aa1..00000000
--- a/harness/src/context-compaction/summarize.ts
+++ /dev/null
@@ -1,271 +0,0 @@
-import type { ISdk } from '../runtime/iii.js';
-import { logger } from '../runtime/otel.js';
-import {
-  COMPACTION_CUSTOM_TYPE,
-  type CompactionData,
-  compactionRowsFromPath,
-  messageItemsFromPath,
-  readActivePath,
-  sessionAppendCustom,
-} from '../runtime/session.js';
-import type { AgentMessage, AssistantMessage } from '../types/agent-message.js';
-import { compactionConfig } from './config.js';
-import { stampLastCompaction } from './lease.js';
-import { type ModelLimit, preserveRecentBudget } from './overflow.js';
-import {
-  type CompactionEntryLike,
-  type MessageWithEntryId,
-  completedCompactions,
-  selectWithEntryIds,
-} from './selection.js';
-import { stripMedia } from './strip-media.js';
-import { buildPrompt } from './template.js';
-
-/** Outer trigger budget; must exceed the router's 300s stream budget. */
-const SUMMARIZER_TIMEOUT_MS = 320_000;
-
-export type SummarizeMode = 'async' | 'sync';
-
-export type SummarizeOptions = {
-  mode: SummarizeMode;
-  /** Sync path passes its already-truncated set so we don't double-load. */
-  truncatedEntries?: MessageWithEntryId[];
-};
-
-export type SummarizeOk = {
-  kind: 'ok';
-  tail_start_id: string | null;
-  tokens_before: number;
-  compaction_entry_id: string;
-  summary_text: string;
-  tail_messages: AgentMessage[];
-};
-
-export type SummarizeFailure = { kind: 'compact'; reason: string };
-export type SummarizeOutcome = SummarizeOk | SummarizeFailure | 'empty';
-
-// Auth / 4xx / malformed requests won't fix on retry. Everything else
-// (5xx, 429, network, timeout, unknown) gets one retry. Conservative
-// default: when in doubt, retry.
-export function isRetryableStreamError(err: unknown): boolean {
-  if (!(err instanceof Error)) return true;
-  const msg = err.message.toLowerCase();
-  if (
-    /\b(40[01345]|auth|unauthorized|forbidden|invalid[_ ]?(api[_ ]?key|key|request)|malformed)\b/.test(
-      msg,
-    )
-  ) {
-    return false;
-  }
-  return true;
-}
-
-export function renderUserPrompt(older: AgentMessage[]): string {
-  let out =
-    'Summarise the conversation below following the system-prompt structure exactly. Keep identifiers verbatim.\n\n';
-  out += '<conversation>\n';
-  for (const msg of older) {
-    out += `\n[${msg.role}]\n`;
-    for (const block of msg.content) {
-      if ((block as { type?: string }).type === 'text') {
-        out += `${(block as { type: 'text'; text: string }).text}\n`;
-      } else if ((block as { type?: string }).type === 'function_call') {
-        const fc = block as unknown as { function_id: string; arguments: unknown };
-        out += `[tool_call] ${fc.function_id} ${JSON.stringify(fc.arguments)}\n`;
-      }
-    }
-  }
-  out += '</conversation>\n';
-  return out;
-}
-
-export function estimateTokenCount(messages: AgentMessage[]): number {
-  let chars = 0;
-  for (const m of messages) chars += JSON.stringify(m).length;
-  return Math.floor(chars / 4);
-}
-
-async function loadActiveWithIds(iii: ISdk, session_id: string): Promise<MessageWithEntryId[]> {
-  return messageItemsFromPath(await readActivePath(iii, session_id));
-}
-
-async function loadCompactionEntries(
-  iii: ISdk,
-  session_id: string,
-): Promise<CompactionEntryLike[]> {
-  const items = await readActivePath(iii, session_id, { include_custom: true });
-  return compactionRowsFromPath(items).map((row) => ({
-    id: row.entry_id,
-    summary: row.summary,
-    tokens_before: row.tokens_before,
-    timestamp: row.timestamp,
-  }));
-}
-
-async function appendCompaction(
-  iii: ISdk,
-  session_id: string,
-  summary: string,
-  tokens_before: number,
-  tail_start_id: string | null,
-): Promise<string> {
-  const data: CompactionData = {
-    summary,
-    tokens_before,
-    tail_start_id,
-    details: { read_files: [], modified_files: [] },
-    timestamp: Date.now(),
-  };
-  const resp = await sessionAppendCustom(iii, {
-    session_id,
-    custom_type: COMPACTION_CUSTOM_TYPE,
-    data,
-  });
-  return resp.entry_id;
-}
-
-export async function summarizeAndAppend(
-  iii: ISdk,
-  session_id: string,
-  opts: SummarizeOptions,
-  model: { providerID: string; modelID: string; modelLimit: ModelLimit },
-): Promise<SummarizeOutcome> {
-  const entries = opts.truncatedEntries ?? (await loadActiveWithIds(iii, session_id));
-  const prior = completedCompactions(await loadCompactionEntries(iii, session_id));
-  const previousSummary = prior.at(-1)?.summary;
-
-  const cfg = compactionConfig();
-  const budget = preserveRecentBudget({
-    model: { id: model.modelID, limit: model.modelLimit },
-    override: cfg.preserveRecentTokensOverride,
-  });
-
-  const sel = selectWithEntryIds({
-    entries,
-    budget,
-    tailTurns: cfg.tailTurns,
-    estimate: estimateTokenCount,
-  });
-  if (sel.head.length === 0) {
-    return 'empty';
-  }
-
-  const head_messages = sel.head.map((e) => e.message);
-  const tail_messages: AgentMessage[] = entries.slice(sel.head.length).map((e) => e.message);
-  const tokens_before = estimateTokenCount(head_messages);
-  const stripped = stripMedia(head_messages, { toolOutputMaxChars: cfg.toolOutputMaxChars });
-
-  const systemPrompt = buildPrompt({ previousSummary, context: [] });
-  const userPrompt = renderUserPrompt(stripped);
-
-  // One 250ms retry for transient failures (429/5xx/network). Permanent
-  // failures (auth, malformed request) skip the retry to release the lease
-  // sooner. The session's own provider/model are pinned explicitly so
-  // /compact runs on exactly the provider the session streams on.
-  const completeInput = {
-    model: model.modelID,
-    provider: model.providerID,
-    system_prompt: systemPrompt,
-    messages: [
-      {
-        role: 'user' as const,
-        content: [{ type: 'text' as const, text: userPrompt }],
-        timestamp: Date.now(),
-      },
-    ],
-    tools: [],
-  };
-  let final: AssistantMessage;
-  try {
-    final = await routerComplete(iii, completeInput);
-  } catch (firstErr) {
-    const firstReason = firstErr instanceof Error ? firstErr.message : String(firstErr);
-    if (!isRetryableStreamError(firstErr)) {
-      logger.warn('summariser stream failed (non-retryable)', { session_id, err: firstReason });
-      return { kind: 'compact', reason: firstReason };
-    }
-    logger.warn('summariser stream failed; retrying once', { session_id, err: firstReason });
-    await new Promise<void>((resolve) => {
-      setTimeout(resolve, 250);
-    });
-    try {
-      final = await routerComplete(iii, completeInput);
-    } catch (secondErr) {
-      const reason = secondErr instanceof Error ? secondErr.message : String(secondErr);
-      logger.warn('summariser stream failed after retry', {
-        session_id,
-        err: reason,
-        firstErr: firstReason,
-      });
-      return { kind: 'compact', reason };
-    }
-  }
-
-  let summary = '';
-  for (const block of final.content) {
-    if ((block as { type?: string }).type === 'text') {
-      const t = (block as { type: 'text'; text: string }).text;
-      summary += summary.length > 0 ? `\n${t}` : t;
-    }
-  }
-  if (summary.length === 0) {
-    logger.warn('summariser produced empty summary; skipping append', { session_id });
-    return 'empty';
-  }
-
-  const tail_start_id = sel.tail_start_id ?? null;
-  const compaction_entry_id = await appendCompaction(
-    iii,
-    session_id,
-    summary,
-    tokens_before,
-    tail_start_id,
-  );
-  await stampLastCompaction(iii, session_id);
-  return {
-    kind: 'ok',
-    tail_start_id,
-    tokens_before,
-    compaction_entry_id,
-    summary_text: summary,
-    tail_messages,
-  };
-}
-
-/**
- * Run the summariser turn through `router::complete`. Pre-stream failures
- * (unknown provider, not configured) throw from the trigger; a mid-stream
- * failure comes back as an error-shaped AssistantMessage, surfaced here as a
- * throw so the caller's retry logic treats both uniformly.
- */
-async function routerComplete(
-  iii: ISdk,
-  input: Record<string, unknown>,
-): Promise<AssistantMessage> {
-  const resp = await iii.trigger<unknown, { message?: AssistantMessage }>({
-    function_id: 'router::complete',
-    payload: input,
-    timeoutMs: SUMMARIZER_TIMEOUT_MS,
-  });
-  const message = resp?.message;
-  if (!message) {
-    throw new Error('router::complete returned no message');
-  }
-  if (message.stop_reason === 'error') {
-    const detail =
-      typeof message.error_message === 'string' && message.error_message.length > 0
-        ? message.error_message
-        : extractTextFromMessage(message);
-    throw new Error(`summariser stream error: ${detail || 'unknown provider error'}`);
-  }
-  return message;
-}
-
-function extractTextFromMessage(msg: AssistantMessage): string {
-  for (const block of msg.content ?? []) {
-    if ((block as { type?: string }).type === 'text') {
-      return (block as { type: 'text'; text: string }).text;
-    }
-  }
-  return '';
-}
diff --git a/harness/src/context-compaction/template.ts b/harness/src/context-compaction/template.ts
deleted file mode 100644
index ba0670be..00000000
--- a/harness/src/context-compaction/template.ts
+++ /dev/null
@@ -1,54 +0,0 @@
-// "Tool Calls Made" section retains tool-call awareness even though prune
-// strips the actual outputs before summarisation.
-export const SUMMARY_TEMPLATE = `Output exactly the Markdown structure shown inside <template> and keep the section order unchanged. Do not include the <template> tags in your response.
-<template>
-## Goal
-- [single-sentence task summary]
-
-## Constraints & Preferences
-- [user constraints, preferences, specs, or "(none)"]
-
-## Progress
-### Done
-- [completed work or "(none)"]
-
-### In Progress
-- [current work or "(none)"]
-
-### Blocked
-- [blockers or "(none)"]
-
-## Key Decisions
-- [decision and why, or "(none)"]
-
-## Tool Calls Made
-- [notable tool calls with their arguments (terse), or "(none)"]
-
-## Next Steps
-- [ordered next actions or "(none)"]
-
-## Critical Context
-- [important technical facts, errors, open questions, or "(none)"]
-
-## Relevant Files
-- [file or directory path: why it matters, or "(none)"]
-</template>
-
-Rules:
-- Keep every section, even when empty.
-- Use terse bullets, not prose paragraphs.
-- Preserve exact file paths, commands, error strings, and identifiers when known.
-- Do not mention the summary process or that context was compacted.`;
-
-export function buildPrompt(input: { previousSummary?: string; context: string[] }): string {
-  const anchor = input.previousSummary
-    ? [
-        'Update the anchored summary below using the conversation history above.',
-        'Preserve still-true details, remove stale details, and merge in the new facts.',
-        '<previous-summary>',
-        input.previousSummary,
-        '</previous-summary>',
-      ].join('\n')
-    : 'Create a new anchored summary from the conversation history above.';
-  return [anchor, SUMMARY_TEMPLATE, ...input.context].join('\n\n');
-}
diff --git a/harness/src/index.ts b/harness/src/index.ts
index 2abd37d1..b8d5258d 100644
--- a/harness/src/index.ts
+++ b/harness/src/index.ts
@@ -64,7 +64,7 @@ const WORKERS: readonly WorkerDefinition[] = [
   {
     name: 'context-compaction',
     description:
-      'Out-of-band session-history compactor. Subscribes to agent::events::TurnEnd and writes a session-manager compaction custom entry when the running token count crosses the configured threshold.',
+      'Thin user-initiated /compact wrapper (context-compaction::compact_session) over the context-manager worker; persists the compaction round trip as a session bookkeeping entry. The transcript is never modified.',
     register: (iii) => registerContextCompaction(iii),
   },
   {
diff --git a/harness/src/runtime/compaction.ts b/harness/src/runtime/compaction.ts
new file mode 100644
index 00000000..44f0dfcc
--- /dev/null
+++ b/harness/src/runtime/compaction.ts
@@ -0,0 +1,307 @@
+/**
+ * Compaction round-trip helpers — the harness side of the `context-manager`
+ * contract (context-manager/architecture/integration.md §5, harness.md
+ * § Compaction persistence).
+ *
+ * context-manager is stateless: it returns a budgeted context plus, when it
+ * summarised, a summary + tail boundary. The harness OWNS persistence:
+ *
+ * 1. `loadAssembleWindow` reads the active path, finds the latest compaction
+ *    bookkeeping entry, and returns the post-compaction message window (from
+ *    `tail_start_entry_id` onward) plus its summary as `previousSummary`.
+ *    Compaction custom entries are NEVER part of the window.
+ * 2. `assembleContext` calls `context::assemble` with that window; on any
+ *    failure it degrades to the raw window + base prompt (context-manager is a
+ *    SOFT dependency — harness.md § Dependencies).
+ * 3. `persistCompactionRoundTrip` writes a `kind:"custom"` bookkeeping entry
+ *    (`custom_type: "compaction"`) when `applied.compacted`, mapping
+ *    `applied.tail_start_index` onto the entry id at that position in the
+ *    window we sent. The durable transcript itself is never modified.
+ */
+
+import type { ISdk } from './iii.js';
+import { logger } from './otel.js';
+import {
+  COMPACTION_CUSTOM_TYPE,
+  type CompactionData,
+  type CompactionRow,
+  type MessageWithEntryId,
+  compactionRowsFromPath,
+  messageItemsFromPath,
+  readActivePath,
+  sessionAppendCustom,
+  type TranscriptItem,
+} from './session.js';
+import type { AgentMessage } from '../types/agent-message.js';
+import type { Model } from '../types/model.js';
+
+/** Outer budget: assemble/compact may make a summariser LLM call. */
+const CONTEXT_TIMEOUT_MS = 320_000;
+
+/** Inline limits accepted by `context::assemble` / `context::compact`. */
+type ModelLimits = {
+  context_window: number;
+  max_output_tokens: number;
+  input_limit?: number;
+};
+
+/** `context-manager` ModelInput (types.rs `ModelInput`). */
+type ModelInput = {
+  id: string;
+  provider?: string;
+  limits?: ModelLimits;
+};
+
+/** The subset of `context::assemble`'s `applied` block the harness reads. */
+export type AppliedReport = {
+  pruned: boolean;
+  pruned_tokens: number;
+  compacted: boolean;
+  summary?: string;
+  /** Index into the REQUEST messages where the tail begins; null = all summarised. */
+  tail_start_index?: number | null;
+  tokens_before?: number;
+};
+
+export type AssembleResult = {
+  system_prompt: string;
+  messages: AgentMessage[];
+  applied: AppliedReport;
+};
+
+export type LoadWindowOptions = {
+  /** Entry ids dropped from the window (this turn's assistant placeholder). */
+  excludeEntryIds?: string[];
+};
+
+export type AssembleWindow = {
+  window: MessageWithEntryId[];
+  previousSummary?: string;
+};
+
+/**
+ * Empty assistant messages are streaming placeholders (or the residue of a
+ * run that died before its first delta); providers reject empty assistant
+ * turns, so they never enter the window.
+ */
+function isEmptyAssistant(m: AgentMessage): boolean {
+  return m.role === 'assistant' && m.content.length === 0;
+}
+
+/**
+ * Latest compaction by path order — each record is appended at the
+ * then-current leaf, so path order is chronological.
+ */
+function latestCompaction(compactions: CompactionRow[]): CompactionRow | null {
+  return compactions.length > 0 ? (compactions.at(-1) ?? null) : null;
+}
+
+/**
+ * Where the candidate window starts inside the path `items`:
+ * - no compaction → 0 (the whole history).
+ * - compaction with a `tail_start_entry_id` still on the path → that entry
+ *   (the verbatim tail + everything appended after the compaction record).
+ * - compaction with a null/missing boundary → right after the compaction
+ *   record (the summary already covers everything before it; replaying it
+ *   would double-count against `previous_summary`).
+ */
+function windowStartIndex(items: TranscriptItem[], compaction: CompactionRow): number {
+  const afterCompaction = () => {
+    const cIdx = items.findIndex((it) => it.entry_id === compaction.entry_id);
+    return cIdx >= 0 ? cIdx + 1 : 0;
+  };
+  if (!compaction.tail_start_entry_id) return afterCompaction();
+  const tailIdx = items.findIndex((it) => it.entry_id === compaction.tail_start_entry_id);
+  return tailIdx >= 0 ? tailIdx : afterCompaction();
+}
+
+/**
+ * The post-compaction message window plus the anchored summary. The window is
+ * the raw messages (with entry ids) the harness sends to `context::assemble`;
+ * the summary returns as `previousSummary` so a re-compaction updates it in
+ * place instead of starting over.
+ */
+export async function loadAssembleWindow(
+  iii: ISdk,
+  session_id: string,
+  opts?: LoadWindowOptions,
+): Promise<AssembleWindow> {
+  const items = await readActivePath(iii, session_id, { include_custom: true });
+  const exclude = new Set(opts?.excludeEntryIds ?? []);
+  const compaction = latestCompaction(compactionRowsFromPath(items));
+  const start = compaction ? windowStartIndex(items, compaction) : 0;
+  const window = messageItemsFromPath(items.slice(start)).filter(
+    (e) => !exclude.has(e.entry_id) && !isEmptyAssistant(e.message),
+  );
+  return compaction ? { window, previousSummary: compaction.summary } : { window };
+}
+
+/** Build a `context-manager` ModelInput, preferring inline limits when known. */
+export function buildModelInput(modelId: string, provider?: string, meta?: Model): ModelInput {
+  const model: ModelInput = { id: modelId };
+  if (provider) model.provider = provider;
+  if (meta && typeof meta.context_window === 'number' && meta.context_window > 0) {
+    model.limits = {
+      context_window: meta.context_window,
+      max_output_tokens: typeof meta.max_output_tokens === 'number' ? meta.max_output_tokens : 0,
+      ...(typeof meta.input_limit === 'number' ? { input_limit: meta.input_limit } : {}),
+    };
+  }
+  return model;
+}
+
+export type AssembleParams = {
+  session_id: string;
+  window: MessageWithEntryId[];
+  baseSystemPrompt: string;
+  modelId: string;
+  provider?: string;
+  modelMeta?: Model;
+  thinkingLevel?: string;
+  previousSummary?: string;
+};
+
+type AssembleResponse = {
+  system_prompt?: string;
+  messages?: AgentMessage[];
+  applied?: AppliedReport;
+};
+
+/**
+ * Build the model-ready context via `context::assemble`. Best-effort: any
+ * failure (worker absent, trigger error, malformed response) degrades to the
+ * raw window + base prompt with no compaction, so the turn still runs without
+ * `context-manager` (harness.md § Dependencies — soft).
+ */
+export async function assembleContext(iii: ISdk, params: AssembleParams): Promise<AssembleResult> {
+  const raw: AssembleResult = {
+    system_prompt: params.baseSystemPrompt,
+    messages: params.window.map((w) => w.message),
+    applied: { pruned: false, pruned_tokens: 0, compacted: false },
+  };
+
+  const options: Record<string, unknown> = { lease_key: params.session_id };
+  if (params.previousSummary) options.previous_summary = params.previousSummary;
+  if (params.thinkingLevel && params.thinkingLevel !== 'off') {
+    options.thinking_level = params.thinkingLevel;
+  }
+
+  try {
+    const resp = await iii.trigger<unknown, AssembleResponse>({
+      function_id: 'context::assemble',
+      payload: {
+        messages: raw.messages,
+        model: buildModelInput(params.modelId, params.provider, params.modelMeta),
+        system_prompt: params.baseSystemPrompt,
+        options,
+      },
+      timeoutMs: CONTEXT_TIMEOUT_MS,
+    });
+    if (!resp || !Array.isArray(resp.messages) || typeof resp.system_prompt !== 'string') {
+      logger.warn('context::assemble returned an unexpected shape; using raw window', {
+        session_id: params.session_id,
+      });
+      return raw;
+    }
+    return {
+      system_prompt: resp.system_prompt,
+      messages: resp.messages,
+      applied: resp.applied ?? raw.applied,
+    };
+  } catch (err) {
+    logger.warn('context::assemble failed; falling back to raw window', {
+      session_id: params.session_id,
+      err: String(err),
+    });
+    return raw;
+  }
+}
+
+/**
+ * Persist the compaction round trip as an additive `kind:"custom"` bookkeeping
+ * entry. Returns the new entry id, or null when nothing was compacted. The
+ * transcript messages are never touched — this is loop bookkeeping the harness
+ * owns (harness.md § Compaction persistence).
+ */
+export async function persistCompactionRoundTrip(
+  iii: ISdk,
+  session_id: string,
+  applied: AppliedReport,
+  window: MessageWithEntryId[],
+): Promise<string | null> {
+  if (!applied.compacted) return null;
+  const idx = applied.tail_start_index;
+  const tailEntry =
+    typeof idx === 'number' && idx >= 0 && idx < window.length ? window[idx] : undefined;
+  const tail_start_entry_id = tailEntry ? tailEntry.entry_id : null;
+  const data: CompactionData = {
+    summary: applied.summary ?? '',
+    tokens_before: applied.tokens_before ?? 0,
+    tail_start_entry_id,
+    details: { read_files: [], modified_files: [] },
+    timestamp: Date.now(),
+  };
+  const resp = await sessionAppendCustom(iii, {
+    session_id,
+    custom_type: COMPACTION_CUSTOM_TYPE,
+    data,
+  });
+  return resp.entry_id;
+}
+
+/** Map console's `{ id, providerID, limit? }` shape onto a ModelInput. */
+export function modelInputFromConsole(raw: unknown): ModelInput | null {
+  if (!raw || typeof raw !== 'object') return null;
+  const m = raw as {
+    id?: unknown;
+    providerID?: unknown;
+    limit?: { context?: unknown; input?: unknown; output?: unknown };
+  };
+  if (typeof m.id !== 'string' || m.id.length === 0) return null;
+  const model: ModelInput = { id: m.id };
+  if (typeof m.providerID === 'string' && m.providerID.length > 0) model.provider = m.providerID;
+  const lim = m.limit;
+  if (lim && typeof lim.context === 'number' && lim.context > 0) {
+    model.limits = {
+      context_window: lim.context,
+      max_output_tokens: typeof lim.output === 'number' ? lim.output : 0,
+      ...(typeof lim.input === 'number' ? { input_limit: lim.input } : {}),
+    };
+  }
+  return model;
+}
+
+export type CompactResponse =
+  | {
+      status: 'ok';
+      summary: string;
+      tail_start_index: number | null;
+      tokens_before: number;
+      tokens_after: number;
+      used_prior_summary: boolean;
+    }
+  | { status: 'busy' }
+  | { status: 'empty' }
+  | { status: 'overflow' };
+
+/** Call `context::compact` directly (used by the `/compact` wrapper). */
+export async function compactWindow(
+  iii: ISdk,
+  session_id: string,
+  window: MessageWithEntryId[],
+  model: ModelInput,
+  previousSummary?: string,
+): Promise<CompactResponse> {
+  const options: Record<string, unknown> = { lease_key: session_id };
+  if (previousSummary) options.previous_summary = previousSummary;
+  const resp = await iii.trigger<unknown, CompactResponse>({
+    function_id: 'context::compact',
+    payload: {
+      messages: window.map((w) => w.message),
+      model,
+      options,
+    },
+    timeoutMs: CONTEXT_TIMEOUT_MS,
+  });
+  return resp ?? { status: 'overflow' };
+}
diff --git a/harness/src/runtime/lease.ts b/harness/src/runtime/lease.ts
index 4e4676c3..0fe26d5f 100644
--- a/harness/src/runtime/lease.ts
+++ b/harness/src/runtime/lease.ts
@@ -3,8 +3,8 @@
  * iii exposes (there's no native lock). Acquire writes a `{nonce, ts}` claim and
  * reads the prior value in one atomic op, so exactly one concurrent acquirer
  * sees a free/expired prior and wins; a claim older than `ttlMs` reads inactive,
- * folding crash recovery into the same op. session-lease and
- * context-compaction/lease are thin (scope, ttl) adapters.
+ * folding crash recovery into the same op. session-lease is a thin (scope, ttl)
+ * adapter over this.
  */
 
 import { randomUUID } from 'node:crypto';
diff --git a/harness/src/runtime/session.ts b/harness/src/runtime/session.ts
index 206d659f..72cee9d1 100644
--- a/harness/src/runtime/session.ts
+++ b/harness/src/runtime/session.ts
@@ -45,7 +45,13 @@ export const COMPACTION_CUSTOM_TYPE = 'compaction';
 export type CompactionData = {
   summary: string;
   tokens_before: number;
-  tail_start_id: string | null;
+  /**
+   * Entry id where the verbatim tail begins (spec name, harness.md
+   * § Compaction persistence). `null` when the whole head was summarised.
+   * Read back with a legacy `tail_start_id` fallback for pre-migration
+   * sessions (see {@link compactionRowsFromPath}).
+   */
+  tail_start_entry_id: string | null;
   details: Record<string, unknown>;
   /** Wall-clock at write time; path order remains the authoritative order. */
   timestamp: number;
@@ -243,13 +249,19 @@ export function compactionRowsFromPath(items: TranscriptItem[]): CompactionRow[]
   const rows: CompactionRow[] = [];
   for (const item of items) {
     if (item.custom?.custom_type !== COMPACTION_CUSTOM_TYPE) continue;
-    const data = (item.custom.data ?? {}) as Partial<CompactionData>;
+    // Legacy `tail_start_id` is read for sessions written before the
+    // rename to the spec's `tail_start_entry_id`.
+    const data = (item.custom.data ?? {}) as Partial<CompactionData> & {
+      tail_start_id?: unknown;
+    };
     if (typeof data.summary !== 'string') continue;
+    const tailNew = typeof data.tail_start_entry_id === 'string' ? data.tail_start_entry_id : null;
+    const tailLegacy = typeof data.tail_start_id === 'string' ? data.tail_start_id : null;
     rows.push({
       entry_id: item.entry_id,
       summary: data.summary,
       tokens_before: typeof data.tokens_before === 'number' ? data.tokens_before : 0,
-      tail_start_id: typeof data.tail_start_id === 'string' ? data.tail_start_id : null,
+      tail_start_entry_id: tailNew ?? tailLegacy,
       details:
         data.details && typeof data.details === 'object'
           ? (data.details as Record<string, unknown>)
diff --git a/harness/src/turn-orchestrator/assistant-streaming/ports.ts b/harness/src/turn-orchestrator/assistant-streaming/ports.ts
index 62eb54f3..f4774212 100644
--- a/harness/src/turn-orchestrator/assistant-streaming/ports.ts
+++ b/harness/src/turn-orchestrator/assistant-streaming/ports.ts
@@ -5,7 +5,21 @@
 import { z } from 'zod';
 import { logger } from '../../runtime/otel.js';
 import type { ISdk } from '../../runtime/iii.js';
-import { sessionAppendMessage, sessionUpdateMessage } from '../../runtime/session.js';
+import {
+  sessionAppendMessage,
+  sessionUpdateMessage,
+  type MessageWithEntryId,
+} from '../../runtime/session.js';
+import {
+  assembleContext,
+  loadAssembleWindow,
+  persistCompactionRoundTrip,
+  type AppliedReport,
+  type AssembleParams,
+  type AssembleResult,
+  type AssembleWindow,
+  type LoadWindowOptions,
+} from '../../runtime/compaction.js';
 import {
   emptyAssistant,
   type AgentMessage,
@@ -16,10 +30,8 @@ import type { AgentFunction } from '../../types/function.js';
 import type { AssistantMessageEvent } from '../../types/stream-event.js';
 import { AgentFunctionSchema } from '../../types/function.js';
 import { emit } from '../events.js';
-import { runPreflight } from '../preflight.js';
 import { streamProviderTurn } from '../provider-stream.js';
 import type { RunRequest } from '../run-request.js';
-import type { Model } from '../../types/model.js';
 import { createTurnStatePorts, type TurnStatePorts } from '../state-runtime/ports.js';
 
 export type StreamContext = {
@@ -71,13 +83,23 @@ export type DeltaHandler = (
 
 export type AssistantStreamingPorts = TurnStatePorts & {
   loadRunRequest(session_id: string): Promise<RunRequest>;
-  runPreflight(
+  /**
+   * Load the post-compaction message window (raw messages from the latest
+   * compaction's tail boundary onward) plus the anchored summary, for
+   * `context::assemble`. Compaction bookkeeping entries are never included.
+   */
+  loadAssembleWindow(session_id: string, opts?: LoadWindowOptions): Promise<AssembleWindow>;
+  /** Build the model-ready context via `context::assemble` (best-effort). */
+  assembleContext(params: AssembleParams): Promise<AssembleResult>;
+  /**
+   * Persist the compaction round trip as an additive `kind:"custom"`
+   * bookkeeping entry when assemble compacted. The transcript is untouched.
+   */
+  persistCompaction(
     session_id: string,
-    messages: AgentMessage[],
-    provider: string,
-    model: string,
-    model_meta?: Model,
-  ): Promise<'ok' | 'compacted'>;
+    applied: AppliedReport,
+    window: MessageWithEntryId[],
+  ): Promise<void>;
   streamTurn(
     ctx: StreamContext,
     onDelta: DeltaHandler,
@@ -124,8 +146,16 @@ export function createStreamingPorts(iii: ISdk): AssistantStreamingPorts {
   return {
     ...base,
 
-    async runPreflight(session_id, messages, provider, model, model_meta) {
-      return runPreflight(iii, session_id, messages, provider, model, model_meta);
+    async loadAssembleWindow(session_id, opts) {
+      return loadAssembleWindow(iii, session_id, opts);
+    },
+
+    async assembleContext(params) {
+      return assembleContext(iii, params);
+    },
+
+    async persistCompaction(session_id, applied, window) {
+      await persistCompactionRoundTrip(iii, session_id, applied, window);
     },
 
     async streamTurn(ctx, onDelta) {
diff --git a/harness/src/turn-orchestrator/assistant-streaming/run.ts b/harness/src/turn-orchestrator/assistant-streaming/run.ts
index a019baa9..6000f45c 100644
--- a/harness/src/turn-orchestrator/assistant-streaming/run.ts
+++ b/harness/src/turn-orchestrator/assistant-streaming/run.ts
@@ -42,37 +42,44 @@ export async function prepareStreamContext(
   assistantEntry: string,
 ): Promise<StreamContext> {
   const request = await ports.loadRunRequest(rec.session_id);
-  // Exclude this turn's assistant entry: on step re-entry the placeholder
-  // already sits on the path (possibly with partial content) and must not be
-  // replayed to the provider as history.
-  const loadOpts = { excludeEntryIds: [assistantEntry] };
-  let messages = await ports.loadMessages(rec.session_id, loadOpts);
   const { provider, model, system_prompt, function_schemas, thinking_level } = request;
   // Provisioning pinned the routed provider; fall back to the raw request
   // provider for records provisioned before the router cutover.
   const routedProvider = request.routed_provider ?? provider;
   const tools = parseFunctionSchemas(function_schemas);
-  const model_meta = rec.model_meta;
-
-  if (
-    (await ports.runPreflight(
-      rec.session_id,
-      messages,
-      routedProvider || provider,
-      model,
-      model_meta,
-    )) === 'compacted'
-  ) {
-    messages = await ports.loadMessages(rec.session_id, loadOpts);
+
+  // Load the post-compaction window. Exclude this turn's assistant entry: on
+  // step re-entry the placeholder already sits on the path (possibly with
+  // partial content) and must not be replayed to the provider as history.
+  const { window, previousSummary } = await ports.loadAssembleWindow(rec.session_id, {
+    excludeEntryIds: [assistantEntry],
+  });
+
+  // context::assemble prunes/compacts the window to fit the model and renders
+  // any summary into the system prompt. When it compacts, the harness owns
+  // persistence of the round trip (an additive compaction bookkeeping entry;
+  // the transcript itself is never modified).
+  const assembled = await ports.assembleContext({
+    session_id: rec.session_id,
+    window,
+    baseSystemPrompt: system_prompt,
+    modelId: model,
+    provider: routedProvider || provider,
+    modelMeta: rec.model_meta,
+    thinkingLevel: thinking_level,
+    previousSummary,
+  });
+  if (assembled.applied.compacted) {
+    await ports.persistCompaction(rec.session_id, assembled.applied, window);
   }
 
   return {
     session_id: rec.session_id,
     provider: routedProvider,
     model,
-    system_prompt,
+    system_prompt: assembled.system_prompt,
     tools,
-    messages,
+    messages: assembled.messages,
     ...(thinking_level ? { thinking_level } : {}),
     request_id: `${rec.session_id}:${rec.started_at_ms}`,
   };
diff --git a/harness/src/turn-orchestrator/errors.ts b/harness/src/turn-orchestrator/errors.ts
index f3166320..8b708e86 100644
--- a/harness/src/turn-orchestrator/errors.ts
+++ b/harness/src/turn-orchestrator/errors.ts
@@ -1,6 +1,5 @@
 /**
- * Turn-orchestrator pre-flight error types surfaced when context-compaction
- * cannot recover a session before a provider call.
+ * Turn-orchestrator FSM error types.
  */
 
 /** Thrown by a handler for a genuinely retryable failure. runTransition
@@ -13,26 +12,6 @@ export class TransientError extends Error {
   }
 }
 
-export class ContextOverflowError extends Error {
-  constructor(message: string) {
-    super(message);
-    this.name = 'ContextOverflowError';
-  }
-}
-
-/** Another compaction holds the session's compaction lease (e.g. the async
- *  post-turn summarize of a large session outlives `busyTimeoutMs`). This is
- *  transient by construction — the lease TTL (300s) bounds it — so it extends
- *  {@link TransientError}: runTransition re-throws and the turn-step queue
- *  retries the step once the in-flight compaction releases, instead of
- *  killing the session with a terminal `failed`. */
-export class CompactionBusyError extends TransientError {
-  constructor(message: string) {
-    super(message);
-    this.name = 'CompactionBusyError';
-  }
-}
-
 /** Persisted turn_state is missing fields required for the current FSM step. */
 export class TurnStateInvariantError extends Error {
   constructor(message: string) {
diff --git a/harness/src/turn-orchestrator/events.ts b/harness/src/turn-orchestrator/events.ts
index 1775d7af..8135dc02 100644
--- a/harness/src/turn-orchestrator/events.ts
+++ b/harness/src/turn-orchestrator/events.ts
@@ -1,10 +1,6 @@
 /**
  * Emit AgentEvent frames on `agent::events`, one per call with a per-session
- * monotonic sequence number. `turn_end` frames additionally enqueue a
- * compaction wake (`context-compaction::on_turn_end`) on the default queue, so
- * the out-of-band compactor runs once per turn instead of consuming the full
- * `agent::events` firehose. The enqueue is best-effort: a failure (e.g. the
- * compactor isn't deployed) is logged, never fatal.
+ * monotonic sequence number.
  *
  * The sequence number is kept IN-PROCESS (not persisted): the old per-event
  * `state::update` increment cost one engine state write — and thus one
@@ -19,21 +15,12 @@
  */
 
 import { uuidLike } from '../runtime/ids.js';
-import { TriggerAction, type ISdk } from '../runtime/iii.js';
+import type { ISdk } from '../runtime/iii.js';
 import { logger } from '../runtime/otel.js';
 import type { AgentEvent } from '../types/agent-event.js';
 
 export const EVENTS_STREAM = 'agent::events';
 
-/**
- * Out-of-band compactor woken once per turn. The turn-orchestrator enqueues a
- * typed `{ session_id, usage, provider, model, model_limit? }` payload here at
- * `turn_end` instead of mirroring the event onto a dedicated stream — a 1:1
- * channel is a queue message, not pub/sub.
- */
-const COMPACTION_ON_TURN_END = 'context-compaction::on_turn_end';
-const COMPACTION_QUEUE = 'default';
-
 /** Unique per process run; prefixes every item_id so a restart can't collide. */
 const PROCESS_EPOCH = uuidLike();
 /**
@@ -58,10 +45,6 @@ function formatItemId(session_id: string, seq: number): string {
   return `${session_id}-${PROCESS_EPOCH}-${seq.toString().padStart(8, '0')}`;
 }
 
-function isTurnEnd(event: AgentEvent): boolean {
-  return (event as { type?: string }).type === 'turn_end';
-}
-
 async function setStream(
   iii: ISdk,
   stream_name: string,
@@ -84,42 +67,8 @@ async function setStream(
   }
 }
 
-/**
- * Enqueue an out-of-band compaction wake from a `turn_end` event. The assistant
- * message carries everything the compactor needs to decide overflow — token
- * usage plus the provider/model it ran on — so we pass them as a typed payload
- * rather than making the compactor re-derive them from a stream frame.
- */
-async function enqueueCompaction(iii: ISdk, session_id: string, event: AgentEvent): Promise<void> {
-  const ev = event as {
-    message?: { usage?: unknown; provider?: unknown; model?: unknown };
-    model_limit?: unknown;
-  };
-  const message = ev.message;
-  try {
-    await iii.trigger({
-      function_id: COMPACTION_ON_TURN_END,
-      payload: {
-        session_id,
-        usage: message?.usage ?? null,
-        provider: typeof message?.provider === 'string' ? message.provider : '',
-        model: typeof message?.model === 'string' ? message.model : '',
-        ...(ev.model_limit && typeof ev.model_limit === 'object'
-          ? { model_limit: ev.model_limit }
-          : {}),
-      },
-      action: TriggerAction.Enqueue({ queue: COMPACTION_QUEUE }),
-    });
-  } catch (err) {
-    logger.warn('compaction on_turn_end enqueue failed', { session_id, err: String(err) });
-  }
-}
-
 export async function emit(iii: ISdk, session_id: string, event: AgentEvent): Promise<void> {
   const seq = nextSeq(session_id);
   const item_id = formatItemId(session_id, seq);
   await setStream(iii, EVENTS_STREAM, session_id, item_id, event);
-  if (isTurnEnd(event)) {
-    await enqueueCompaction(iii, session_id, event);
-  }
 }
diff --git a/harness/src/turn-orchestrator/preflight.ts b/harness/src/turn-orchestrator/preflight.ts
deleted file mode 100644
index 6065d426..00000000
--- a/harness/src/turn-orchestrator/preflight.ts
+++ /dev/null
@@ -1,97 +0,0 @@
-/**
- * Pre-flight overflow check. Returns 'compacted' when compact_now ran;
- * caller must then reload messages from persistence.
- *
- * @throws ContextOverflowError when the session is too large to compact (terminal).
- * @throws CompactionBusyError  when another compaction is in progress — a
- *   TransientError subclass, so the turn-step queue retries the step after
- *   the in-flight compaction releases the lease instead of failing the run.
- */
-
-import { fetchModelLimit, limitFromModel } from '../context-compaction/model-resolver.js';
-import { usable as computeUsable } from '../context-compaction/overflow.js';
-import type { Model } from '../types/model.js';
-import type { ISdk } from '../runtime/iii.js';
-import { logger } from '../runtime/otel.js';
-import { readActivePath } from '../runtime/session.js';
-import type { AgentMessage } from '../types/agent-message.js';
-import { CompactionBusyError, ContextOverflowError } from './errors.js';
-
-/** Cheap chars/4 token estimate — same heuristic as context-compaction's estimateTokenCount. */
-function estimateMessages(messages: AgentMessage[]): number {
-  let chars = 0;
-  for (const m of messages) chars += JSON.stringify(m).length;
-  return Math.floor(chars / 4);
-}
-
-export async function runPreflight(
-  iii: ISdk,
-  session_id: string,
-  messages: AgentMessage[],
-  providerID: string,
-  modelID: string,
-  modelMeta?: Model,
-): Promise<'ok' | 'compacted'> {
-  const model = modelMeta
-    ? { providerID, modelID, modelLimit: limitFromModel(modelMeta) }
-    : await fetchModelLimit(iii, providerID, modelID);
-  if (!model || model.modelLimit.context === 0) {
-    if (model && model.modelLimit.context === 0) {
-      logger.debug('preflight: model has no context_window; skipping pre-flight check', {
-        providerID,
-        modelID,
-      });
-    }
-    return 'ok';
-  }
-
-  const usable = computeUsable({ model: { id: modelID, limit: model.modelLimit } });
-  const projected = estimateMessages(messages);
-
-  if (projected < usable) return 'ok';
-
-  logger.info('preflight: projected overflow detected; triggering compact_now', {
-    session_id,
-    projected,
-    usable,
-    model: modelID,
-  });
-
-  let lastUserEntryId = '';
-  try {
-    const items = await readActivePath(iii, session_id, { roles: ['user'] });
-    lastUserEntryId = items.at(-1)?.entry_id ?? '';
-  } catch (err) {
-    logger.debug('preflight: session::messages failed; using empty last_user_entry_id', {
-      session_id,
-      err: String(err),
-    });
-  }
-
-  const res = await iii.trigger<unknown, { status: string }>({
-    function_id: 'context-compaction::compact_now',
-    payload: {
-      session_id,
-      projected_tokens: projected,
-      last_user_message_id: lastUserEntryId,
-      model: { id: model.modelID, providerID: model.providerID, limit: model.modelLimit },
-    },
-    timeoutMs: 60_000,
-  });
-
-  if (res?.status === 'overflow') {
-    throw new ContextOverflowError('session too large to compact');
-  }
-  if (res?.status === 'busy') {
-    throw new CompactionBusyError('compaction already in progress');
-  }
-  if (res?.status !== 'ok') {
-    logger.warn('preflight: compact_now returned non-ok status; proceeding without reload', {
-      session_id,
-      status: res?.status,
-    });
-    return 'ok';
-  }
-
-  return 'compacted';
-}
diff --git a/harness/src/turn-orchestrator/state-runtime/context-view.ts b/harness/src/turn-orchestrator/state-runtime/context-view.ts
deleted file mode 100644
index ee7bfd25..00000000
--- a/harness/src/turn-orchestrator/state-runtime/context-view.ts
+++ /dev/null
@@ -1,74 +0,0 @@
-/**
- * Reconstruct the provider-facing message window from session-manager state.
- *
- * No compaction: the raw active path. With compaction: the latest summary
- * followed by the preserved tail (everything from `tail_start_id` onward).
- */
-
-import { buildSummaryMessage } from '../../context-compaction/flat-state.js';
-import type { ISdk } from '../../runtime/iii.js';
-import {
-  compactionRowsFromPath,
-  messageItemsFromPath,
-  readActivePath,
-  type CompactionRow,
-  type MessageWithEntryId,
-} from '../../runtime/session.js';
-import type { AgentMessage } from '../../types/agent-message.js';
-
-export type ContextViewCompaction = Pick<CompactionRow, 'summary' | 'tail_start_id' | 'timestamp'>;
-
-export type ContextViewOptions = {
-  /**
-   * Entry ids excluded from the window — this turn's just-appended assistant
-   * placeholder on step re-entry, which must not be replayed to the provider.
-   */
-  excludeEntryIds?: string[];
-};
-
-function latestCompaction(compactions: ContextViewCompaction[]): ContextViewCompaction | null {
-  if (compactions.length === 0) return null;
-  // Stable sort: ties keep path order, which is the authoritative order.
-  return [...compactions].sort((a, b) => a.timestamp - b.timestamp).at(-1) ?? null;
-}
-
-/**
- * Empty assistant messages are streaming placeholders (or the residue of a
- * run that died before its first delta); providers reject empty assistant
- * turns, so they never enter the window.
- */
-function isEmptyAssistant(m: AgentMessage): boolean {
-  return m.role === 'assistant' && m.content.length === 0;
-}
-
-/** Pure reconstruction from path-ordered messages and compaction rows. */
-export function buildContextView(
-  messages: MessageWithEntryId[],
-  compactions: ContextViewCompaction[],
-): AgentMessage[] {
-  const usable = messages.filter((m) => !isEmptyAssistant(m.message));
-  const compaction = latestCompaction(compactions);
-  if (!compaction) {
-    return usable.map((m) => m.message);
-  }
-  const found = compaction.tail_start_id
-    ? usable.findIndex((m) => m.entry_id === compaction.tail_start_id)
-    : 0;
-  const tailStart = found >= 0 ? found : 0;
-  return [
-    buildSummaryMessage(compaction.summary),
-    ...usable.slice(tailStart).map((m) => m.message),
-  ];
-}
-
-export async function loadContextView(
-  iii: ISdk,
-  session_id: string,
-  opts?: ContextViewOptions,
-): Promise<AgentMessage[]> {
-  const items = await readActivePath(iii, session_id, { include_custom: true });
-  const exclude = new Set(opts?.excludeEntryIds ?? []);
-  const entries = messageItemsFromPath(items).filter((e) => !exclude.has(e.entry_id));
-  const compactions = compactionRowsFromPath(items);
-  return buildContextView(entries, compactions);
-}
diff --git a/harness/src/turn-orchestrator/state-runtime/ports.ts b/harness/src/turn-orchestrator/state-runtime/ports.ts
index a74db099..07441357 100644
--- a/harness/src/turn-orchestrator/state-runtime/ports.ts
+++ b/harness/src/turn-orchestrator/state-runtime/ports.ts
@@ -5,19 +5,12 @@
 import { emit } from '../events.js';
 import type { RunRequest } from '../run-request.js';
 import type { ISdk } from '../../runtime/iii.js';
-import type { ModelContextLimit } from '../../types/agent-event.js';
 import { sessionSetStatus } from '../../runtime/session.js';
 import type { AgentMessage, FunctionResultMessage } from '../../types/agent-message.js';
 import { transitionTo, type TurnStateRecord } from '../state.js';
-import {
-  createTurnStore,
-  type AppendOptions,
-  type LoadMessagesOptions,
-  type TurnStore,
-} from './store.js';
+import { createTurnStore, type AppendOptions, type TurnStore } from './store.js';
 
 export type TurnStatePorts = {
-  loadMessages(session_id: string, opts?: LoadMessagesOptions): Promise<AgentMessage[]>;
   appendMessages(session_id: string, msgs: AgentMessage[], opts?: AppendOptions): Promise<void>;
   checkpoint(rec: TurnStateRecord): Promise<void>;
   loadRunRequest(session_id: string): Promise<RunRequest>;
@@ -26,7 +19,6 @@ export type TurnStatePorts = {
     session_id: string,
     message: AgentMessage,
     function_results: FunctionResultMessage[],
-    model_limit?: ModelContextLimit,
   ): Promise<void>;
   finishSession(rec: TurnStateRecord): Promise<void>;
 };
@@ -44,10 +36,6 @@ export function createTurnStatePorts(iii: ISdk, store?: TurnStore): TurnStatePor
   const s = store ?? createTurnStore(iii);
 
   return {
-    loadMessages(session_id, opts) {
-      return s.loadMessages(session_id, opts);
-    },
-
     appendMessages(session_id, msgs, opts) {
       return s.appendMessages(session_id, msgs, opts);
     },
@@ -64,12 +52,11 @@ export function createTurnStatePorts(iii: ISdk, store?: TurnStore): TurnStatePor
       return s.saveRunRequest(session_id, request);
     },
 
-    async emitTurnEnd(session_id, message, function_results, model_limit) {
+    async emitTurnEnd(session_id, message, function_results) {
       await emit(iii, session_id, {
         type: 'turn_end',
         message,
         function_results,
-        ...(model_limit ? { model_limit } : {}),
       });
     },
 
diff --git a/harness/src/turn-orchestrator/state-runtime/store.ts b/harness/src/turn-orchestrator/state-runtime/store.ts
index 1069ab90..30285a65 100644
--- a/harness/src/turn-orchestrator/state-runtime/store.ts
+++ b/harness/src/turn-orchestrator/state-runtime/store.ts
@@ -17,7 +17,6 @@ import { emit } from '../events.js';
 import { type RunRequest, defaultRunRequest } from '../run-request.js';
 import { toView, type TurnStateView } from '../schemas.js';
 import { type TurnState, type TurnStateRecord } from '../state.js';
-import { loadContextView } from './context-view.js';
 
 /**
  * Turn-step wakes go to the engine's `default` queue. NOTE: engine.config.yaml
@@ -82,11 +81,6 @@ export type AppendOptions = {
   entryIdFor?: (msg: AgentMessage, index: number) => string | undefined;
 };
 
-export type LoadMessagesOptions = {
-  /** Entry ids dropped from the provider window (this turn's placeholder). */
-  excludeEntryIds?: string[];
-};
-
 export type TurnStore = {
   loadRecord(session_id: string): Promise<TurnStateRecord | null>;
   /**
@@ -99,7 +93,6 @@ export type TurnStore = {
   saveRecord(rec: TurnStateRecord, previous?: TurnStateRecord | null): Promise<void>;
   writeRecord(rec: TurnStateRecord): Promise<void>;
   ensureSession(session_id: string): Promise<void>;
-  loadMessages(session_id: string, opts?: LoadMessagesOptions): Promise<AgentMessage[]>;
   appendMessages(session_id: string, msgs: AgentMessage[], opts?: AppendOptions): Promise<void>;
   loadRunRequest(session_id: string): Promise<RunRequest>;
   saveRunRequest(session_id: string, request: RunRequest): Promise<void>;
@@ -197,10 +190,6 @@ export function createTurnStore(iii: ISdk): TurnStore {
       await sessionEnsure(iii, session_id);
     },
 
-    async loadMessages(session_id, opts) {
-      return loadContextView(iii, session_id, opts);
-    },
-
     /*
      * Sequential appends without explicit parent_id: each append chains from
      * the active leaf and moves it, preserving order. Deterministic entry ids
diff --git a/harness/src/turn-orchestrator/state-runtime/turn-end.ts b/harness/src/turn-orchestrator/state-runtime/turn-end.ts
index a01d6714..5c125685 100644
--- a/harness/src/turn-orchestrator/state-runtime/turn-end.ts
+++ b/harness/src/turn-orchestrator/state-runtime/turn-end.ts
@@ -2,8 +2,7 @@
  * Shared turn-end helper for step outcome application.
  */
 
-import { limitFromModel } from '../../context-compaction/model-resolver.js';
-import type { AgentEvent, ModelContextLimit } from '../../types/agent-event.js';
+import type { AgentEvent } from '../../types/agent-event.js';
 import {
   emptyAssistant,
   type AgentMessage,
@@ -18,7 +17,6 @@ export type TurnEndEmitter = {
     session_id: string,
     message: AssistantMessage,
     function_results: FunctionResultMessage[],
-    model_limit?: ModelContextLimit,
   ): Promise<void>;
 };
 
@@ -30,8 +28,7 @@ export async function emitTurnEndOnce(
 ): Promise<void> {
   if (rec.turn_end_emitted) return;
   const last = message ?? rec.last_assistant ?? emptyAssistant();
-  const model_limit = rec.model_meta ? limitFromModel(rec.model_meta) : undefined;
-  await ports.emitTurnEnd(rec.session_id, last, function_results, model_limit);
+  await ports.emitTurnEnd(rec.session_id, last, function_results);
   rec.turn_end_emitted = true;
 }
 
diff --git a/harness/src/turn-orchestrator/state.ts b/harness/src/turn-orchestrator/state.ts
index ef69ab30..273a5716 100644
--- a/harness/src/turn-orchestrator/state.ts
+++ b/harness/src/turn-orchestrator/state.ts
@@ -55,11 +55,12 @@ type TurnStateRecordCore = {
   assistant_body_streamed?: boolean;
   /**
    * Full catalog entry for this turn's model, resolved once at `provisioning`
-   * and read by preflight, provider streaming, and turn-end compaction instead
-   * of each re-fetching `models::get`. Optional: absent on records persisted
-   * before this field existed (and on a cold catalog) — every reader falls back
-   * to a live fetch. The model is fixed for a turn, so no invalidation is
-   * needed. Excluded from {@link toView}, so it never bloats turn_state events.
+   * and read by the assemble pre-flight (inline limits for `context::assemble`)
+   * and provider streaming instead of re-fetching `models::get`. Optional:
+   * absent on records persisted before this field existed (and on a cold
+   * catalog) — readers fall back accordingly. The model is fixed for a turn, so
+   * no invalidation is needed. Excluded from {@link toView}, so it never bloats
+   * turn_state events.
    */
   model_meta?: Model;
   error?: { kind: string; message: string };
diff --git a/harness/src/types/agent-event.ts b/harness/src/types/agent-event.ts
index 5c8eb377..8d33492b 100644
--- a/harness/src/types/agent-event.ts
+++ b/harness/src/types/agent-event.ts
@@ -10,22 +10,12 @@ import type { AgentMessage, FunctionResultMessage } from './agent-message.js';
 import type { FunctionResult } from './function.js';
 import type { AssistantMessageEvent } from './stream-event.js';
 
-/** A model's context limits, in tokens. Produced once per turn by the orchestrator. */
-export type ModelContextLimit = { context: number; input: number; output: number };
-
 export type AgentEvent =
   | { type: 'agent_end'; messages: AgentMessage[] }
   | {
       type: 'turn_end';
       message: AgentMessage;
       function_results: FunctionResultMessage[];
-      /**
-       * Turn's model context limit, threaded from the orchestrator so the
-       * compaction worker reads it inline instead of re-fetching `models::get`.
-       * Optional: absent only when the catalog had no entry to resolve at turn
-       * start, where compaction falls back to resolving the model itself.
-       */
-      model_limit?: ModelContextLimit;
     }
   | {
       type: 'message_update';
diff --git a/harness/tests/_helpers/fakeSessionManager.ts b/harness/tests/_helpers/fakeSessionManager.ts
index 04af56f0..98fc741a 100644
--- a/harness/tests/_helpers/fakeSessionManager.ts
+++ b/harness/tests/_helpers/fakeSessionManager.ts
@@ -91,12 +91,18 @@ export class FakeSessionManager {
     data: {
       summary: string;
       tokens_before?: number;
-      tail_start_id?: string | null;
+      tail_start_entry_id?: string | null;
       timestamp?: number;
+      /**
+       * Write the legacy `tail_start_id` key instead of `tail_start_entry_id`,
+       * to cover the back-compat read path for pre-migration sessions.
+       */
+      legacyTailField?: boolean;
     },
   ): string {
     const s = this.session(session_id);
     const entry_id = `compaction-${++autoId}`;
+    const tail = data.tail_start_entry_id ?? null;
     s.entries.push({
       entry_id,
       parent_id: s.entries.at(-1)?.entry_id ?? null,
@@ -106,7 +112,7 @@ export class FakeSessionManager {
         data: {
           summary: data.summary,
           tokens_before: data.tokens_before ?? 0,
-          tail_start_id: data.tail_start_id ?? null,
+          ...(data.legacyTailField ? { tail_start_id: tail } : { tail_start_entry_id: tail }),
           details: {},
           timestamp: data.timestamp ?? Date.now(),
         },
@@ -209,3 +215,39 @@ export class FakeSessionManager {
     return { status: s.status, previous_status };
   }
 }
+
+/**
+ * In-memory fake of the `context::*` surface — `assemble` and `compact`.
+ * Tests route `context::*` trigger calls into `handle()`; responses are
+ * configurable via {@link assemble} / {@link compact}.
+ */
+export class FakeContextManager {
+  calls: Array<{ function_id: string; payload: Record<string, unknown> }> = [];
+
+  /** Default assemble: echo the request (no prune/compaction). Override per test. */
+  assemble: (p: Record<string, unknown>) => unknown = (p) => ({
+    system_prompt: typeof p.system_prompt === 'string' ? p.system_prompt : '',
+    messages: Array.isArray(p.messages) ? p.messages : [],
+    token_count: 0,
+    usable: 1_000_000,
+    model_resolved: 'inline',
+    applied: { pruned: false, pruned_tokens: 0, compacted: false },
+  });
+
+  /** Default compact: nothing to compact. Override per test. */
+  compact: (p: Record<string, unknown>) => unknown = () => ({ status: 'empty' });
+
+  /** Route a trigger call. Returns undefined when not a context::* id. */
+  async handle(function_id: string, payload: unknown): Promise<unknown> {
+    if (!function_id.startsWith('context::')) return undefined;
+    const p = (payload ?? {}) as Record<string, unknown>;
+    this.calls.push({ function_id, payload: p });
+    if (function_id === 'context::assemble') return this.assemble(p);
+    if (function_id === 'context::compact') return this.compact(p);
+    throw new Error(`fake context-manager: unhandled ${function_id}`);
+  }
+
+  callsTo(function_id: string): Array<Record<string, unknown>> {
+    return this.calls.filter((c) => c.function_id === function_id).map((c) => c.payload);
+  }
+}
diff --git a/harness/tests/context-compaction/compact-session-registered.test.ts b/harness/tests/context-compaction/compact-session-registered.test.ts
index 947d44d1..29fde734 100644
--- a/harness/tests/context-compaction/compact-session-registered.test.ts
+++ b/harness/tests/context-compaction/compact-session-registered.test.ts
@@ -1,241 +1,173 @@
 /**
- * End-to-end test for `context-compaction::compact_session` invoked via the
- * REGISTERED handler (not handleSync directly). Mirrors the UI `/compact`
- * slash-command path: provider/model in payload, session-manager mock that
- * returns a realistic transcript.
- *
- * The bug we're chasing: the UI reports `compacted — 0 tokens summarised
- * (continued)`. `(continued)` proves a replay happened, so summarize was
- * NOT in the 'empty' branch. So tokens_before should be > 0. This test
- * asserts that.
+ * `context-compaction::compact_session` — the thin `/compact` wrapper over the
+ * context-manager worker. Exercises the registered handler end to end:
+ * session-manager + context-manager are faked, and we assert the wrapper maps
+ * the `context::compact` union onto the console's `CompactResult` wire shape and
+ * persists the round trip as an additive compaction bookkeeping entry (the
+ * transcript is never rewritten).
  */
 
 import { describe, expect, it, vi } from 'vitest';
 import { register } from '../../src/context-compaction/register.js';
 import type { ISdk } from '../../src/runtime/iii.js';
 import type { AgentMessage } from '../../src/types/agent-message.js';
+import { FakeContextManager, FakeSessionManager } from '../_helpers/fakeSessionManager.js';
 
-type Handler = (payload: unknown) => Promise<unknown>;
+type Handler = (payload: unknown) => Promise<unknown> | unknown;
 
-function userEntry(id: string, text: string): { entry_id: string; message: AgentMessage } {
-  return {
-    entry_id: id,
-    message: { role: 'user', content: [{ type: 'text', text }], timestamp: 0 },
-  };
-}
-
-function asstEntry(id: string, text: string): { entry_id: string; message: AgentMessage } {
-  return {
-    entry_id: id,
-    message: {
-      role: 'assistant',
-      content: [{ type: 'text', text }],
-      stop_reason: 'end',
-      error_message: null,
-      error_kind: null,
-      usage: null,
-      model: 'fake-model',
-      provider: 'anthropic',
-      timestamp: 0,
-    },
-  };
-}
-
-function buildSdk(opts: {
-  sessionEntries: Array<{ entry_id: string; message: AgentMessage }>;
-  summaryText?: string;
-}): {
+function buildIii(): {
   iii: ISdk;
   handlers: Map<string, Handler>;
-  stateStore: Map<string, unknown>;
+  sessions: FakeSessionManager;
+  context: FakeContextManager;
 } {
   const handlers = new Map<string, Handler>();
-  const stateStore = new Map<string, unknown>();
-
-  const trigger = vi.fn(async (req: { function_id: string; payload?: unknown }) => {
-    const fn = req.function_id;
-    const payload = req.payload;
-    const p = (payload ?? {}) as Record<string, unknown>;
-
-    if (fn === 'state::get') {
-      const v = stateStore.get(p.key as string);
-      return v !== undefined ? v : null;
-    }
-    if (fn === 'state::set') {
-      if (p.value === null || p.value === undefined) stateStore.delete(p.key as string);
-      else stateStore.set(p.key as string, p.value);
-      return { ok: true };
-    }
-    if (fn === 'state::update') {
-      const key = p.key as string;
-      const ops = (p.ops ?? []) as Array<{ type: string; value?: unknown }>;
-      const oldValue = stateStore.has(key) ? stateStore.get(key) : null;
-      let newValue: unknown = oldValue;
-      for (const op of ops) {
-        if (op.type === 'set') newValue = op.value;
-      }
-      if (newValue === null || newValue === undefined) stateStore.delete(key);
-      else stateStore.set(key, newValue);
-      return { old_value: oldValue ?? null, new_value: newValue ?? null };
-    }
-    if (fn === 'session::messages') {
-      return { messages: opts.sessionEntries };
-    }
-    if (fn === 'session::append') {
-      return { entry_id: `appended-${Date.now()}`, parent_id: null, timestamp: Date.now() };
-    }
-    if (fn === 'session::update_message') {
-      return { updated: true, revision: 1 };
-    }
-    if (fn === 'router::models::get') {
-      return { model: { context_window: 200_000, max_output_tokens: 4_096 } };
-    }
-    if (fn === 'router::complete') {
-      const summary = opts.summaryText ?? 'summary text here';
-      return {
-        message: {
-          role: 'assistant',
-          content: [{ type: 'text', text: summary }],
-          stop_reason: 'end',
-          model: 'fake-model',
-          provider: 'anthropic',
-          timestamp: Date.now(),
-        },
-      };
-    }
-    return null;
-  });
-
-  const registerFunction = vi.fn((id: string, h: Handler) => {
-    handlers.set(id, h);
-  });
-
+  const sessions = new FakeSessionManager();
+  const context = new FakeContextManager();
   const iii = {
-    trigger,
-    registerFunction,
+    trigger: vi.fn(async (req: { function_id: string; payload?: unknown }) => {
+      const s = await sessions.handle(req.function_id, req.payload);
+      if (s !== undefined) return s;
+      const c = await context.handle(req.function_id, req.payload);
+      if (c !== undefined) return c;
+      return null;
+    }),
+    registerFunction: vi.fn((id: string, h: Handler) => {
+      handlers.set(id, h);
+    }),
     registerTrigger: vi.fn(),
     publish: vi.fn(),
     subscribe: vi.fn(),
     enqueue: vi.fn(),
   } as unknown as ISdk;
+  return { iii, handlers, sessions, context };
+}
 
-  return { iii, handlers, stateStore };
+function seed(sessions: FakeSessionManager): string[] {
+  const msgs: AgentMessage[] = [
+    { role: 'user', content: [{ type: 'text', text: 'q1' }], timestamp: 1 },
+    {
+      role: 'assistant',
+      content: [{ type: 'text', text: 'a1' }],
+      stop_reason: 'end',
+      model: 'm',
+      provider: 'p',
+      timestamp: 2,
+    },
+    { role: 'user', content: [{ type: 'text', text: 'q2' }], timestamp: 3 },
+  ];
+  return sessions.seedMessages('s1', msgs, ['u1', 'a1', 'u2']);
 }
 
 describe('context-compaction::compact_session via registered handler', () => {
-  it('returns ok with tokens_before > 0 and auto_continued=false for a multi-turn session', async () => {
-    // 2 user + 2 assistant turns. compact_session does NOT extract a
-    // replay target (the replay mechanism is for compact_now's overflow
-    // pre-flight only), so all 4 entries feed into selection. The summary
-    // covers turn 1, turn 2 stays as tail.
-    const entries = [
-      userEntry('e1', 'first question'),
-      asstEntry('e2', 'first answer'),
-      userEntry('e3', 'second question — the most recent user message'),
-      asstEntry('e4', 'second answer'),
-    ];
-
-    const { iii, handlers } = buildSdk({
-      sessionEntries: entries,
-      summaryText: 'summary covering turns 1-2',
+  it('maps context::compact ok onto CompactResult and persists the round trip', async () => {
+    const { iii, handlers, sessions, context } = buildIii();
+    seed(sessions);
+    context.compact = () => ({
+      status: 'ok',
+      summary: 'SUMMARY',
+      tail_start_index: 2,
+      tokens_before: 123,
+      tokens_after: 10,
+      used_prior_summary: false,
     });
 
     await register(iii);
-
     const handler = handlers.get('context-compaction::compact_session');
-    expect(handler).toBeDefined();
-
     const result = (await handler?.({
-      session_id: 'test-session',
-      model: { id: 'fake-model', providerID: 'anthropic' },
-    })) as {
-      status: string;
-      tokens_before?: number;
-      auto_continued?: boolean;
-      tail_start_id?: string | null;
-      summary_text?: string;
-    };
-
-    expect(result.status).toBe('ok');
-    // /compact is user-initiated; no in-flight turn to auto-continue.
-    expect(result.auto_continued).toBe(false);
-    expect(typeof result.tokens_before).toBe('number');
-    expect(result.tokens_before).toBeGreaterThan(0);
-    // Surface every field so we can spot wire-format weirdness.
+      session_id: 's1',
+      model: { id: 'm', providerID: 'p' },
+    })) as Record<string, unknown>;
+
+    // Console wire contract (unchanged): legacy `tail_start_id` key in the RPC.
+    expect(result).toMatchObject({
+      status: 'ok',
+      tokens_before: 123,
+      auto_continued: false,
+      summary_text: 'SUMMARY',
+      tail_start_id: 'u2',
+    });
     expect(Object.keys(result).sort()).toEqual(
       ['auto_continued', 'status', 'summary_text', 'tail_start_id', 'tokens_before'].sort(),
     );
-    // summary_text MUST be present and non-empty so the UI marker can ship
-    // it as the next-turn <conversation-summary> block.
-    expect(typeof result.summary_text).toBe('string');
-    expect((result.summary_text ?? '').length).toBeGreaterThan(0);
+
+    // Persisted as an additive compaction custom entry using the spec field name.
+    const appended = sessions
+      .callsTo('session::append')
+      .filter((p) => (p.custom as { custom_type?: string })?.custom_type === 'compaction');
+    expect(appended).toHaveLength(1);
+    const data = (appended[0]?.custom as { data: Record<string, unknown> }).data;
+    expect(data.tail_start_entry_id).toBe('u2');
+    expect(data.summary).toBe('SUMMARY');
+
+    // The transcript is untouched: the 3 seeded messages are all still present.
+    expect(sessions.messageEntries('s1')).toHaveLength(3);
   });
 
-  it('summarises a session with only one user message and nothing else', async () => {
-    // Before the fix this returned 'empty' because compact_session
-    // extracted u1 as the replay target, leaving truncatedMessages = [].
-    // After the fix the single user message is summarised normally.
-    const entries = [userEntry('only-user', 'just one message')];
-    const { iii, handlers } = buildSdk({
-      sessionEntries: entries,
-      summaryText: 'summary of the lone user message',
+  it('anchors on the latest stored summary as previous_summary', async () => {
+    const { iii, handlers, sessions, context } = buildIii();
+    seed(sessions);
+    sessions.seedCompaction('s1', { summary: 'PRIOR', tail_start_entry_id: 'u1' });
+    context.compact = () => ({
+      status: 'ok',
+      summary: 'NEW',
+      tail_start_index: null,
+      tokens_before: 5,
+      tokens_after: 5,
+      used_prior_summary: true,
     });
-    await register(iii);
-    const handler = handlers.get('context-compaction::compact_session');
 
-    const result = (await handler?.({
-      session_id: 'one-msg-session',
-      model: { id: 'fake-model', providerID: 'anthropic' },
-    })) as { status: string; tokens_before?: number; auto_continued?: boolean };
+    await register(iii);
+    await handlers.get('context-compaction::compact_session')?.({
+      session_id: 's1',
+      model: { id: 'm', providerID: 'p' },
+    });
 
-    expect(result.status).toBe('ok');
-    expect(result.tokens_before).toBeGreaterThan(0);
-    expect(result.auto_continued).toBe(false);
+    expect(context.callsTo('context::compact')[0]?.options).toMatchObject({
+      lease_key: 's1',
+      previous_summary: 'PRIOR',
+    });
   });
 
-  it('summarises a single-user-turn session with many subsequent entries', async () => {
-    // Regression: a long-running task with one user message + lots of
-    // assistant/tool output ends up at ~30k+ tokens. Before the fix,
-    // compact_session always extracted the last user message via
-    // extractReplayTarget; with only one user message at idx=0,
-    // truncatedMessages was [] and summarize returned 'empty'. The user
-    // saw "compact: session is too small to summarise" despite a 30%
-    // full context window.
-    //
-    // After the fix, compact_session does NOT extract a replay target
-    // (that mechanism is for the compact_now overflow path), so the full
-    // entry list is summarised normally.
-    const longText = 'x'.repeat(8_000);
-    const entries = [
-      userEntry('u1', 'the only user question — start a long task'),
-      asstEntry('a1', `${longText} first assistant chunk`),
-      asstEntry('a2', `${longText} second assistant chunk`),
-      asstEntry('a3', `${longText} third assistant chunk`),
-      asstEntry('a4', `${longText} fourth assistant chunk`),
-    ];
-
-    const { iii, handlers } = buildSdk({
-      sessionEntries: entries,
-      summaryText: 'summary of the long task',
-    });
+  it('maps busy / empty / overflow statuses', async () => {
+    for (const status of ['busy', 'empty'] as const) {
+      const { iii, handlers, context } = buildIii();
+      context.compact = () => ({ status });
+      await register(iii);
+      const r = await handlers.get('context-compaction::compact_session')?.({
+        session_id: 's1',
+        model: { id: 'm', providerID: 'p' },
+      });
+      expect(r).toEqual({ status });
+    }
 
+    const { iii, handlers, context } = buildIii();
+    context.compact = () => ({ status: 'overflow' });
     await register(iii);
-    const handler = handlers.get('context-compaction::compact_session');
+    const r = (await handlers.get('context-compaction::compact_session')?.({
+      session_id: 's1',
+      model: { id: 'm', providerID: 'p' },
+    })) as Record<string, unknown>;
+    expect(r.status).toBe('overflow');
+    expect(typeof r.message).toBe('string');
+  });
 
-    const result = (await handler?.({
-      session_id: 'one-turn-large-session',
-      model: { id: 'fake-model', providerID: 'anthropic' },
-    })) as {
-      status: string;
-      tokens_before?: number;
-      auto_continued?: boolean;
-      summary_text?: string;
-    };
-
-    expect(result.status).toBe('ok');
-    expect(result.tokens_before).toBeGreaterThan(0);
-    // /compact (user-initiated) does NOT auto-continue. Replay is only
-    // appropriate for the orchestrator's overflow pre-flight.
-    expect(result.auto_continued).toBe(false);
+  it('returns an error result when no model is supplied', async () => {
+    const { iii, handlers } = buildIii();
+    await register(iii);
+    const r = (await handlers.get('context-compaction::compact_session')?.({
+      session_id: 's1',
+    })) as Record<string, unknown>;
+    expect(r.status).toBe('error');
+  });
+
+  it('throws when session_id is missing', async () => {
+    const { iii, handlers } = buildIii();
+    await register(iii);
+    await expect(
+      handlers.get('context-compaction::compact_session')?.({
+        model: { id: 'm', providerID: 'p' },
+      }),
+    ).rejects.toThrow('session_id is required');
   });
 });
diff --git a/harness/tests/context-compaction/compact-session.test.ts b/harness/tests/context-compaction/compact-session.test.ts
deleted file mode 100644
index 311aee44..00000000
--- a/harness/tests/context-compaction/compact-session.test.ts
+++ /dev/null
@@ -1,323 +0,0 @@
-/**
- * Smoke tests for the compact_session wrapper (register.ts logic).
- *
- * We test three things:
- *   1. Throws when session_id is missing/empty.
- *   2. Throws when the session has no assistant messages (cannot resolve model).
- *   3. Happy path: valid session → resolveModelFromSession succeeds → handleSync
- *      returns a deterministic CompactNowResult.
- */
-
-import { describe, expect, it, vi } from 'vitest';
-import { handleSync } from '../../src/context-compaction/handler-sync.js';
-import {
-  fetchModelLimit,
-  resolveModelFromRunRequest,
-  resolveModelFromSession,
-} from '../../src/context-compaction/model-resolver.js';
-import type { ISdk } from '../../src/runtime/iii.js';
-
-// ---------------------------------------------------------------------------
-// ISdk stub factory (mirrors handler-sync.test.ts)
-// ---------------------------------------------------------------------------
-
-function makeStubIii(triggerOverrides: Record<string, unknown> = {}): ISdk {
-  const stateStore = new Map<string, unknown>();
-
-  return {
-    trigger: vi.fn(async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-      if (Object.hasOwn(triggerOverrides, function_id)) {
-        const v = triggerOverrides[function_id];
-        return typeof v === 'function' ? (v as () => unknown)() : v;
-      }
-
-      const p = (payload ?? {}) as Record<string, unknown>;
-
-      if (function_id === 'state::get') {
-        const v = stateStore.get(p.key as string);
-        return v !== undefined ? v : null;
-      }
-      if (function_id === 'state::set') {
-        const v = p.value;
-        if (v === null || v === undefined) {
-          stateStore.delete(p.key as string);
-        } else {
-          stateStore.set(p.key as string, v);
-        }
-        return { ok: true };
-      }
-      if (function_id === 'state::update') {
-        const key = p.key as string;
-        const ops = (p.ops ?? []) as Array<{ type: string; value?: unknown }>;
-        const oldValue = stateStore.has(key) ? stateStore.get(key) : null;
-        let newValue: unknown = oldValue;
-        for (const op of ops) {
-          if (op.type === 'set') newValue = op.value;
-        }
-        if (newValue === null || newValue === undefined) {
-          stateStore.delete(key);
-        } else {
-          stateStore.set(key, newValue);
-        }
-        return { old_value: oldValue ?? null, new_value: newValue ?? null };
-      }
-
-      return null;
-    }),
-    registerFunction: vi.fn(),
-    registerTrigger: vi.fn(),
-    publish: vi.fn(),
-    subscribe: vi.fn(),
-    enqueue: vi.fn(),
-  } as unknown as ISdk;
-}
-
-// ---------------------------------------------------------------------------
-// Fixture helpers — content must be a proper ContentBlock array
-// ---------------------------------------------------------------------------
-
-const userEntry = (entry_id: string, text = 'hello') => ({
-  entry_id,
-  message: {
-    role: 'user' as const,
-    content: [{ type: 'text', text }],
-    timestamp: 0,
-  },
-});
-
-const assistantEntry = (entry_id: string, provider: string, model: string, text = 'reply') => ({
-  entry_id,
-  message: {
-    role: 'assistant' as const,
-    provider,
-    model,
-    content: [{ type: 'text', text }],
-    stop_reason: 'end',
-    timestamp: 0,
-  },
-});
-
-// ---------------------------------------------------------------------------
-// compact_session inline implementation (same logic as register.ts)
-// Keeps the test self-contained without depending on register.ts directly.
-// ---------------------------------------------------------------------------
-
-async function resolveExplicitModel(iii: ISdk, raw: unknown) {
-  if (!raw || typeof raw !== 'object') return null;
-  const m = raw as Record<string, unknown>;
-  const providerID = typeof m.providerID === 'string' && m.providerID ? m.providerID : null;
-  const modelID = typeof m.id === 'string' && m.id ? m.id : null;
-  if (!providerID || !modelID) return null;
-  const lim = m.limit as { context?: number; input?: number; output?: number } | undefined;
-  if (lim && typeof lim.context === 'number' && lim.context > 0) {
-    return {
-      providerID,
-      modelID,
-      modelLimit: {
-        context: lim.context,
-        input: typeof lim.input === 'number' ? lim.input : lim.context,
-        output: typeof lim.output === 'number' ? lim.output : 0,
-      },
-    };
-  }
-  return fetchModelLimit(iii, providerID, modelID);
-}
-
-async function runCompactSession(iii: ISdk, session_id: unknown, payloadModel?: unknown) {
-  if (typeof session_id !== 'string' || session_id.length === 0) {
-    throw new Error('context-compaction::compact_session: session_id is required');
-  }
-
-  // Mirrors register.ts fallback chain so the test catches drift. When
-  // all three sources fail, register.ts now falls back to a conservative
-  // model limit rather than throwing — compaction is best-effort.
-  const FALLBACK = {
-    providerID: 'unknown',
-    modelID: 'unknown',
-    modelLimit: { context: 32_000, input: 32_000, output: 4_000 },
-  };
-  let model = await resolveExplicitModel(iii, payloadModel);
-  if (!model) model = await resolveModelFromSession(iii, session_id);
-  if (!model) model = await resolveModelFromRunRequest(iii, session_id);
-  if (!model) model = FALLBACK;
-
-  // Mirrors register.ts: /compact does NOT extract a replay target.
-  // The replay mechanism is for compact_now (turn-orchestrator overflow
-  // pre-flight) only. Passing '' tells extractReplayTarget that no entry
-  // matches, so the full entries list is summarised.
-  return handleSync(iii, {
-    session_id,
-    projected_tokens: 999_999,
-    last_user_message_id: '',
-    model: {
-      id: model.modelID,
-      providerID: model.providerID,
-      limit: model.modelLimit,
-    },
-  });
-}
-
-// ---------------------------------------------------------------------------
-// Tests
-// ---------------------------------------------------------------------------
-
-describe('compact_session smoke', () => {
-  it('throws when session_id is missing', async () => {
-    const iii = makeStubIii();
-    await expect(runCompactSession(iii, '')).rejects.toThrow(
-      'context-compaction::compact_session: session_id is required',
-    );
-  });
-
-  it('throws when session_id is undefined', async () => {
-    const iii = makeStubIii();
-    await expect(runCompactSession(iii, undefined)).rejects.toThrow(
-      'context-compaction::compact_session: session_id is required',
-    );
-  });
-
-  it('uses the fallback model limit when no source can resolve a model', async () => {
-    // No assistant messages AND no run_request. compact_session used to
-    // throw; it now degrades to a conservative fallback so /compact still
-    // runs (with a small preserve-recent budget) rather than failing.
-    const iii = makeStubIii({
-      'session::messages': { messages: [] },
-    });
-    const result = await runCompactSession(iii, 'no-messages-session');
-    expect(['ok', 'empty', 'overflow', 'busy']).toContain(result.status);
-  });
-
-  it('falls back to run_request when the session has no assistant messages', async () => {
-    // This is the `/compact` UI scenario: the session hasn't yet mirrored
-    // any assistant message into the tree, but run_request from run::start
-    // carries provider/model.
-    const stateStore = new Map<string, unknown>();
-    stateStore.set('session/ui-session/run_request', {
-      provider: 'anthropic',
-      model: 'claude-haiku-4-5',
-    });
-
-    const iii = {
-      trigger: vi.fn(
-        async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-          const p = (payload ?? {}) as Record<string, unknown>;
-          if (function_id === 'session::messages') return { messages: [] };
-          if (function_id === 'router::models::get') {
-            return { model: { context_window: 200_000, max_output_tokens: 4_096 } };
-          }
-          if (function_id === 'state::get') {
-            const v = stateStore.get(p.key as string);
-            return v !== undefined ? v : null;
-          }
-          if (function_id === 'state::set') {
-            const v = p.value;
-            if (v === null || v === undefined) stateStore.delete(p.key as string);
-            else stateStore.set(p.key as string, v);
-            return { ok: true };
-          }
-          if (function_id === 'state::update') {
-            const key = p.key as string;
-            const ops = (p.ops ?? []) as Array<{ type: string; value?: unknown }>;
-            const oldValue = stateStore.has(key) ? stateStore.get(key) : null;
-            let newValue: unknown = oldValue;
-            for (const op of ops) {
-              if (op.type === 'set') newValue = op.value;
-            }
-            if (newValue === null || newValue === undefined) stateStore.delete(key);
-            else stateStore.set(key, newValue);
-            return { old_value: oldValue ?? null, new_value: newValue ?? null };
-          }
-          return null;
-        },
-      ),
-      registerFunction: vi.fn(),
-      registerTrigger: vi.fn(),
-      publish: vi.fn(),
-      subscribe: vi.fn(),
-      enqueue: vi.fn(),
-    } as unknown as ISdk;
-
-    const result = await runCompactSession(iii, 'ui-session');
-    expect(['ok', 'empty', 'overflow', 'busy']).toContain(result.status);
-  });
-
-  it('uses explicit payload.model when provided (skips session scan)', async () => {
-    // Even when the session returns no assistant messages AND no run_request
-    // exists, an explicit model in the payload takes precedence.
-    const stateStore = new Map<string, unknown>();
-    const iii = {
-      trigger: vi.fn(
-        async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-          const p = (payload ?? {}) as Record<string, unknown>;
-          if (function_id === 'session::messages') return { messages: [] };
-          if (function_id === 'state::get') {
-            const v = stateStore.get(p.key as string);
-            return v !== undefined ? v : null;
-          }
-          if (function_id === 'state::set') {
-            const v = p.value;
-            if (v === null || v === undefined) stateStore.delete(p.key as string);
-            else stateStore.set(p.key as string, v);
-            return { ok: true };
-          }
-          if (function_id === 'state::update') {
-            const key = p.key as string;
-            const ops = (p.ops ?? []) as Array<{ type: string; value?: unknown }>;
-            const oldValue = stateStore.has(key) ? stateStore.get(key) : null;
-            let newValue: unknown = oldValue;
-            for (const op of ops) {
-              if (op.type === 'set') newValue = op.value;
-            }
-            if (newValue === null || newValue === undefined) stateStore.delete(key);
-            else stateStore.set(key, newValue);
-            return { old_value: oldValue ?? null, new_value: newValue ?? null };
-          }
-          return null;
-        },
-      ),
-      registerFunction: vi.fn(),
-      registerTrigger: vi.fn(),
-      publish: vi.fn(),
-      subscribe: vi.fn(),
-      enqueue: vi.fn(),
-    } as unknown as ISdk;
-
-    const result = await runCompactSession(iii, 'explicit-model-session', {
-      id: 'claude-opus-4-7',
-      providerID: 'anthropic',
-      limit: { context: 200_000, input: 200_000, output: 4_096 },
-    });
-    expect(['ok', 'empty', 'overflow', 'busy']).toContain(result.status);
-    // Confirm router::models::get was NOT called (limits supplied inline).
-    expect(
-      (iii.trigger as ReturnType<typeof vi.fn>).mock.calls.filter(
-        (c) => (c[0] as { function_id: string }).function_id === 'router::models::get',
-      ).length,
-    ).toBe(0);
-  });
-
-  it('happy path: valid session returns a deterministic CompactNowResult', async () => {
-    // Session has: user message → assistant message (with provider/model) → user message.
-    // resolveModelFromSession scans backwards, finds the assistant message, then fetches
-    // limits via models::get.  handleSync runs on this stub and returns a result.
-    const sessionMessages = {
-      messages: [
-        userEntry('e1', 'first message'),
-        assistantEntry('e2', 'anthropic', 'claude-haiku-4-5'),
-        userEntry('e3', 'please compact'),
-      ],
-    };
-
-    const iii = makeStubIii({
-      'session::messages': sessionMessages,
-      'models::get': { context_window: 200_000, max_output_tokens: 4_096 },
-    });
-
-    const result = await runCompactSession(iii, 'happy-session');
-
-    // With only 3 messages the summariser finds a very small / empty head and
-    // returns 'ok' (tokens_before: 0) or 'empty'. Both are valid deterministic
-    // outcomes given the stub.
-    expect(['ok', 'empty', 'overflow', 'busy']).toContain(result.status);
-  });
-});
diff --git a/harness/tests/context-compaction/compaction-done-emit.test.ts b/harness/tests/context-compaction/compaction-done-emit.test.ts
deleted file mode 100644
index 22cfd005..00000000
--- a/harness/tests/context-compaction/compaction-done-emit.test.ts
+++ /dev/null
@@ -1,221 +0,0 @@
-/**
- * Asserts that both the sync (handleSync) and async (handleAsync) compaction
- * paths publish a `compaction_done` AgentEvent on `agent::events` after a
- * successful tree compaction. The UI consumes this event to drop the
- * post-compaction CTX bar to its correct value.
- *
- * Strategy: vi.mock the summarize + prune + replay helpers so
- * the handler reaches the emit call deterministically without needing a
- * real provider stream. We capture stream::set payloads via a stub ISdk.
- */
-
-import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
-import type { ISdk } from '../../src/runtime/iii.js';
-
-vi.mock('../../src/context-compaction/summarize.js', () => ({
-  summarizeAndAppend: vi.fn(async () => ({
-    kind: 'ok',
-    tail_start_id: 'entry-tail-1',
-    tokens_before: 12_345,
-    compaction_entry_id: 'entry-compact-1',
-    summary_text: 'condensed summary of older turns',
-    tail_messages: [],
-  })),
-}));
-
-vi.mock('../../src/context-compaction/prune.js', () => ({
-  prune: vi.fn(async () => ({ pruned_tokens: 0, pruned_parts: 0, scanned_parts: 0 })),
-}));
-
-vi.mock('../../src/context-compaction/replay.js', () => ({
-  extractReplayTarget: () => ({ replay: null, truncatedMessages: [] }),
-}));
-
-vi.mock('../../src/context-compaction/model-resolver.js', () => ({
-  fetchModelLimit: vi.fn(async () => ({
-    providerID: 'anthropic',
-    modelID: 'claude-haiku-4-5',
-    modelLimit: { context: 200_000, input: 200_000, output: 4_096 },
-  })),
-  resolveModelFromSession: vi.fn(),
-  resolveModelFromRunRequest: vi.fn(),
-}));
-
-type StreamSetCall = {
-  stream_name: string;
-  group_id: string;
-  item_id: string;
-  data: unknown;
-};
-
-function makeStubIii(extra: Record<string, unknown> = {}): {
-  iii: ISdk;
-  streamSetCalls: StreamSetCall[];
-} {
-  const state = new Map<string, unknown>();
-  const streamSetCalls: StreamSetCall[] = [];
-
-  const trigger = vi.fn(
-    async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-      if (Object.hasOwn(extra, function_id)) {
-        const v = extra[function_id];
-        return typeof v === 'function' ? (v as () => unknown)() : v;
-      }
-      const p = (payload ?? {}) as Record<string, unknown>;
-
-      if (function_id === 'stream::set') {
-        streamSetCalls.push(p as unknown as StreamSetCall);
-        return { ok: true };
-      }
-      if (function_id === 'state::get') {
-        const v = state.get(p.key as string);
-        return v !== undefined ? v : null;
-      }
-      if (function_id === 'state::set') {
-        const v = p.value;
-        if (v === null || v === undefined) state.delete(p.key as string);
-        else state.set(p.key as string, v);
-        return { ok: true };
-      }
-      if (function_id === 'state::update') {
-        const key = p.key as string;
-        const ops = (p.ops ?? []) as Array<{ type: string; by?: number; value?: unknown }>;
-        const oldValue = state.has(key) ? state.get(key) : null;
-        let newValue: unknown = oldValue;
-        for (const op of ops) {
-          if (op.type === 'set') newValue = op.value;
-          if (op.type === 'increment') {
-            const prev = typeof oldValue === 'number' ? oldValue : 0;
-            newValue = prev + (op.by ?? 1);
-          }
-        }
-        if (newValue === null || newValue === undefined) state.delete(key);
-        else state.set(key, newValue);
-        return { old_value: oldValue ?? null, new_value: newValue ?? null };
-      }
-      if (function_id === 'session::messages') return { messages: [] };
-      if (function_id === 'session::append')
-        return { entry_id: 'entry-appended-1', parent_id: null, timestamp: 0 };
-      return null;
-    },
-  );
-
-  return {
-    iii: {
-      trigger,
-      registerFunction: vi.fn(),
-      registerTrigger: vi.fn(),
-      publish: vi.fn(),
-      subscribe: vi.fn(),
-      enqueue: vi.fn(),
-    } as unknown as ISdk,
-    streamSetCalls,
-  };
-}
-
-const model = {
-  id: 'claude-haiku-4-5',
-  providerID: 'anthropic',
-  limit: { context: 200_000, input: 200_000, output: 4_096 },
-};
-
-describe('handleSync emits compaction_done', () => {
-  beforeEach(() => {
-    vi.clearAllMocks();
-  });
-
-  afterEach(() => {
-    delete process.env.COMPACT_BUSY_TIMEOUT_MS;
-  });
-
-  it('publishes a compaction_done event on agent::events with mode="sync" after successful compaction', async () => {
-    const { handleSync } = await import('../../src/context-compaction/handler-sync.js');
-    const { iii, streamSetCalls } = makeStubIii();
-
-    const result = await handleSync(iii, {
-      session_id: 'sess-1',
-      projected_tokens: 250_000,
-      last_user_message_id: '',
-      model,
-    });
-
-    expect(result.status).toBe('ok');
-    const events = streamSetCalls.filter((c) => c.stream_name === 'agent::events');
-    expect(events).toHaveLength(1);
-    expect(events[0]?.group_id).toBe('sess-1');
-    const data = events[0]?.data as Record<string, unknown>;
-    expect(data.type).toBe('compaction_done');
-    expect(data.mode).toBe('sync');
-    expect(data.summary_text).toBe('condensed summary of older turns');
-    expect(data.tokens_before).toBe(12_345);
-    expect(data.compaction_entry_id).toBe('entry-compact-1');
-    expect(data.tail_start_id).toBe('entry-tail-1');
-  });
-
-  it('swallows a publish failure and still returns ok to the orchestrator', async () => {
-    const { handleSync } = await import('../../src/context-compaction/handler-sync.js');
-    const { iii } = makeStubIii({
-      'stream::set': () => {
-        throw new Error('stream worker offline');
-      },
-    });
-
-    const result = await handleSync(iii, {
-      session_id: 'sess-publish-fail',
-      projected_tokens: 250_000,
-      last_user_message_id: '',
-      model,
-    });
-
-    // Publish failure must not bubble — the orchestrator's turn is
-    // blocked on this RPC return.
-    expect(result.status).toBe('ok');
-  });
-});
-
-describe('handleAsync emits compaction_done', () => {
-  beforeEach(() => {
-    vi.clearAllMocks();
-  });
-
-  it('publishes a compaction_done event on agent::events with mode="async" when an overflowing TurnEnd triggers compaction', async () => {
-    const { handleAsync } = await import('../../src/context-compaction/handler-async.js');
-    const { iii, streamSetCalls } = makeStubIii();
-
-    // 200k usable (= 200k input − 0 reserved for this test) → overflow at 250k.
-    process.env.COMPACT_RESERVED_TOKENS = '0';
-
-    try {
-      await handleAsync(iii, {
-        session_id: 'sess-async-1',
-        provider: 'anthropic',
-        model: 'claude-haiku-4-5',
-        usage: { input: 200_000, output: 50_000 },
-      });
-
-      const events = streamSetCalls.filter((c) => c.stream_name === 'agent::events');
-      expect(events).toHaveLength(1);
-      const data = events[0]?.data as Record<string, unknown>;
-      expect(data.type).toBe('compaction_done');
-      expect(data.mode).toBe('async');
-      expect(events[0]?.group_id).toBe('sess-async-1');
-    } finally {
-      delete process.env.COMPACT_RESERVED_TOKENS;
-    }
-  });
-
-  it('does NOT emit when the turn did not overflow (under threshold)', async () => {
-    const { handleAsync } = await import('../../src/context-compaction/handler-async.js');
-    const { iii, streamSetCalls } = makeStubIii();
-
-    await handleAsync(iii, {
-      session_id: 'sess-async-noop',
-      provider: 'anthropic',
-      model: 'claude-haiku-4-5',
-      usage: { input: 100, output: 50 },
-    });
-
-    const events = streamSetCalls.filter((c) => c.stream_name === 'agent::events');
-    expect(events).toHaveLength(0);
-  });
-});
diff --git a/harness/tests/context-compaction/e2e/full-session.test.ts b/harness/tests/context-compaction/e2e/full-session.test.ts
deleted file mode 100644
index 5f32507a..00000000
--- a/harness/tests/context-compaction/e2e/full-session.test.ts
+++ /dev/null
@@ -1,363 +0,0 @@
-/**
- * End-to-end compaction flow test.
- *
- * Drives the sync-compaction path through `runPreflight` against an
- * overflowing 30-turn fixture, with the in-memory session-manager fake
- * wired up behind `session::*`. Verifies the three structural guarantees
- * that the unit/integration tests cannot observe in isolation:
- *
- * 1. The reconstructed provider window (active path + compaction custom
- *    entries) is reduced to: [summary-as-asst-msg, ...tail, continue-nudge].
- * 2. The session's active path stays connected — the compaction entry and
- *    continue-prompt are chained via parent_id back to the pre-compaction
- *    tail.
- * 3. The downstream provider input (what would land in `buildInput`)
- *    no longer exceeds the model's usable token budget.
- */
-
-import { describe, expect, it, vi } from 'vitest';
-import { payloadStoreKey } from '../../_helpers/stateStoreKey.js';
-import { FakeSessionManager } from '../../_helpers/fakeSessionManager.js';
-import { handleSync } from '../../../src/context-compaction/handler-sync.js';
-import type { ISdk } from '../../../src/runtime/iii.js';
-import { loadContextView } from '../../../src/turn-orchestrator/state-runtime/context-view.js';
-import { runPreflight } from '../../../src/turn-orchestrator/preflight.js';
-import type { AgentMessage } from '../../../src/types/agent-message.js';
-
-// ---------------------------------------------------------------------------
-// Fixture: a 30-turn conversation that easily exceeds a small usable budget.
-// ---------------------------------------------------------------------------
-
-const SESSION_ID = 'e2e-session-001';
-const PROVIDER_ID = 'anthropic';
-const MODEL_ID = 'fake-haiku-test';
-// Pick limits so the 30-turn fixture overflows: usable ~= input - reserved.
-const MODEL_LIMITS = { context: 8_000, input: 8_000, output: 1_024 };
-
-function userMsg(text: string): AgentMessage {
-  return { role: 'user', content: [{ type: 'text', text }], timestamp: 0 };
-}
-
-function asstMsg(text: string): AgentMessage {
-  return {
-    role: 'assistant',
-    content: [{ type: 'text', text }],
-    stop_reason: 'end',
-    error_message: null,
-    error_kind: null,
-    usage: null,
-    model: MODEL_ID,
-    provider: PROVIDER_ID,
-    timestamp: 0,
-  };
-}
-
-/** Builds a 30-turn conversation with hefty messages so total > usable. */
-function buildOverflowingMessages(turns: number): AgentMessage[] {
-  const out: AgentMessage[] = [];
-  const filler = 'lorem ipsum dolor sit amet '.repeat(40); // ~1000 chars => ~250 tokens
-  for (let i = 0; i < turns; i++) {
-    out.push(userMsg(`turn ${i} question: ${filler}`));
-    out.push(asstMsg(`turn ${i} answer: ${filler}`));
-  }
-  // The final user message is the replay target — the one that "triggered" the turn.
-  out.push(userMsg('final user message that triggers compaction'));
-  return out;
-}
-
-// ---------------------------------------------------------------------------
-// Test ISdk: routes session::* into the FakeSessionManager, plus a stub
-// `models::get`, a stub provider stream (returns a known summary), and an
-// in-memory `state::*` so leases work without the network.
-// ---------------------------------------------------------------------------
-
-type FunctionHandler = (payload: unknown) => Promise<unknown>;
-
-function buildTestSdk(opts: {
-  summaryText: string;
-  /** Capture every provider stream invocation's `messages` payload. */
-  providerInvocations: AgentMessage[][];
-  session_id: string;
-}): { iii: ISdk; sessions: FakeSessionManager; stateStore: Map<string, unknown> } {
-  const stateStore = new Map<string, unknown>();
-  const handlers = new Map<string, FunctionHandler>();
-  const sessions = new FakeSessionManager();
-
-  const trigger = vi.fn(async (req: { function_id: string; payload?: unknown }) => {
-    const fn = req.function_id;
-    const payload = req.payload;
-
-    // 1) state::* — back the lease with stateStore.
-    if (fn === 'state::get') {
-      const p = (payload ?? {}) as { scope: string; key: string };
-      const v = stateStore.get(payloadStoreKey(p));
-      return v !== undefined ? v : null;
-    }
-    if (fn === 'state::set') {
-      const p = (payload ?? {}) as { scope: string; key: string; value: unknown };
-      const storeKey = payloadStoreKey(p);
-      if (p.value === null || p.value === undefined) stateStore.delete(storeKey);
-      else stateStore.set(storeKey, p.value);
-      return { ok: true };
-    }
-    if (fn === 'state::update') {
-      const p = (payload ?? {}) as {
-        scope: string;
-        key: string;
-        ops: Array<{ type: string; value?: unknown }>;
-      };
-      const storeKey = payloadStoreKey(p);
-      const oldValue = stateStore.has(storeKey) ? stateStore.get(storeKey) : null;
-      let newValue: unknown = oldValue;
-      for (const op of p.ops ?? []) {
-        if (op.type === 'set') newValue = op.value;
-      }
-      if (newValue === null || newValue === undefined) stateStore.delete(storeKey);
-      else stateStore.set(storeKey, newValue);
-      return { old_value: oldValue ?? null, new_value: newValue ?? null };
-    }
-
-    // 2) router::models::get — a small-context model so the 30-turn fixture overflows.
-    if (fn === 'router::models::get') {
-      return {
-        model: {
-          id: MODEL_ID,
-          provider: PROVIDER_ID,
-          context_window: MODEL_LIMITS.input,
-          max_output_tokens: MODEL_LIMITS.output,
-        },
-      };
-    }
-
-    // 3) Summariser — router::complete returns the canned summary message.
-    if (fn === 'router::complete') {
-      // Capture what the summariser was asked to look at.
-      const completeInput = (payload ?? {}) as { messages?: AgentMessage[] };
-      if (completeInput.messages) opts.providerInvocations.push(completeInput.messages);
-      return {
-        message: {
-          role: 'assistant',
-          content: [{ type: 'text', text: opts.summaryText }],
-          stop_reason: 'end',
-          model: MODEL_ID,
-          provider: PROVIDER_ID,
-          timestamp: Date.now(),
-        },
-      };
-    }
-
-    // 4) session::* — the in-memory session-manager fake.
-    const handled = await sessions.handle(fn, payload);
-    if (handled !== undefined) return handled;
-
-    const handler = handlers.get(fn);
-    if (handler) return handler(payload);
-
-    return null;
-  });
-
-  const registerFunction = vi.fn((id: string, handler: FunctionHandler) => {
-    handlers.set(id, handler);
-  });
-
-  const iii = {
-    trigger,
-    registerFunction,
-    registerTrigger: vi.fn(),
-    publish: vi.fn(),
-    subscribe: vi.fn(),
-    enqueue: vi.fn(),
-  } as unknown as ISdk;
-
-  // Wire compact_now to call handleSync directly (mirrors register.ts).
-  handlers.set('context-compaction::compact_now', async (payload: unknown) => {
-    const obj = (payload ?? {}) as Record<string, unknown>;
-    const m = obj.model as { id: string; providerID: string; limit: typeof MODEL_LIMITS };
-    return handleSync(iii, {
-      session_id: String(obj.session_id),
-      projected_tokens: typeof obj.projected_tokens === 'number' ? obj.projected_tokens : 0,
-      last_user_message_id: String(obj.last_user_message_id ?? ''),
-      model: { id: m.id, providerID: m.providerID, limit: m.limit },
-    });
-  });
-
-  return { iii, sessions, stateStore };
-}
-
-// ---------------------------------------------------------------------------
-// Helpers for path introspection.
-// ---------------------------------------------------------------------------
-
-function activePathIds(sessions: FakeSessionManager, session_id: string): string[] {
-  const entries = sessions.session(session_id).entries;
-  const last = entries[entries.length - 1];
-  if (!last) return [];
-  const byId = new Map(entries.map((e) => [e.entry_id, e] as const));
-  const path: string[] = [];
-  let cursor: string | null = last.entry_id;
-  while (cursor !== null) {
-    path.push(cursor);
-    cursor = byId.get(cursor)?.parent_id ?? null;
-  }
-  return path.reverse();
-}
-
-async function seedSessionFrom(
-  iii: ISdk,
-  session_id: string,
-  messages: AgentMessage[],
-): Promise<string[]> {
-  // Ensure the session exists.
-  await iii.trigger<unknown, unknown>({
-    function_id: 'session::ensure',
-    payload: { session_id },
-  });
-  const ids: string[] = [];
-  for (const m of messages) {
-    const resp = (await iii.trigger<unknown, { entry_id?: string }>({
-      function_id: 'session::append',
-      payload: { session_id, message: m },
-    })) as { entry_id?: string } | null;
-    ids.push(resp?.entry_id ?? '');
-  }
-  return ids;
-}
-
-// ---------------------------------------------------------------------------
-// E2E tests
-// ---------------------------------------------------------------------------
-
-describe('e2e full-session compaction', () => {
-  it('preflight + sync compaction reduces provider window AND keeps the path connected', async () => {
-    const SUMMARY = 'EARLIER TURNS SUMMARY: discussed lorem and friends.';
-    const overflowing = buildOverflowingMessages(30);
-    const providerInvocations: AgentMessage[][] = [];
-
-    const { iii, sessions } = buildTestSdk({
-      session_id: SESSION_ID,
-      summaryText: SUMMARY,
-      providerInvocations,
-    });
-
-    // Seed the session with the transcript so replay extraction works.
-    const entryIds = await seedSessionFrom(iii, SESSION_ID, overflowing);
-    const lastUserId = entryIds[entryIds.length - 1] ?? ''; // the final user msg
-
-    const beforeView = await loadContextView(iii, SESSION_ID);
-    expect(beforeView.length).toBe(overflowing.length);
-
-    // Run preflight. The 30-turn fixture should overflow the 8k usable budget.
-    const result = await runPreflight(iii, SESSION_ID, overflowing, PROVIDER_ID, MODEL_ID);
-    expect(result).toBe('compacted');
-
-    // --- Assertion 1: reconstructed window is reduced and shaped correctly. ---
-    const afterView = await loadContextView(iii, SESSION_ID);
-    expect(afterView.length).toBeLessThan(overflowing.length);
-
-    // First message must be the summary-as-assistant-msg containing SUMMARY.
-    const first = afterView[0];
-    expect(first?.role).toBe('assistant');
-    if (first?.role === 'assistant') {
-      const text = first.content[0];
-      expect(text?.type).toBe('text');
-      if (text?.type === 'text') expect(text.text).toContain(SUMMARY);
-    }
-
-    // The window keeps the final user message (it stays on the path as the
-    // compaction's parent) and now ends with the continue nudge.
-    const finalUserInView = afterView.some(
-      (m) =>
-        m.role === 'user' &&
-        m.content[0]?.type === 'text' &&
-        m.content[0].text.includes('final user message'),
-    );
-    expect(finalUserInView).toBe(true);
-    const last = afterView[afterView.length - 1];
-    expect(last?.role).toBe('user');
-    if (last?.role === 'user') {
-      const text = last.content[0];
-      if (text?.type === 'text') {
-        expect(text.text).toMatch(/Continue if you have next steps/);
-      }
-    }
-
-    // --- Assertion 2: the active path is connected, no orphans. ---
-    const path = activePathIds(sessions, SESSION_ID);
-    expect(path.length).toBeGreaterThan(2);
-
-    const entries = sessions.session(SESSION_ID).entries;
-    const byId = new Map(entries.map((e) => [e.entry_id, e] as const));
-
-    // The just-appended continue nudge should be the active leaf.
-    const leafId = path[path.length - 1];
-    const leaf = leafId ? byId.get(leafId) : undefined;
-    expect(leaf?.kind).toBe('message');
-    if (leaf?.kind === 'message' && leaf.message) {
-      const m = leaf.message;
-      expect(m.role).toBe('user');
-      if (m.role === 'user') {
-        const text = m.content[0];
-        if (text?.type === 'text') {
-          expect(text.text).toMatch(/Continue if you have next steps/);
-        }
-      }
-    }
-
-    // Walking back from the leaf must hit a compaction custom entry.
-    const hasCompactionOnPath = path.some(
-      (id) => byId.get(id)?.custom?.custom_type === 'compaction',
-    );
-    expect(hasCompactionOnPath).toBe(true);
-
-    // The final user message remains on the active path as the compaction's
-    // parent (no reinjection needed in the chain-only model).
-    const userReplayOnPath = path
-      .map((id) => byId.get(id))
-      .some((e) => {
-        if (e?.kind !== 'message' || e.message?.role !== 'user') return false;
-        const block = e.message.content[0];
-        return block?.type === 'text' && block.text.includes('final user message');
-      });
-    expect(userReplayOnPath).toBe(true);
-
-    // --- Assertion 3: the provider summariser saw a non-empty head. ---
-    expect(providerInvocations.length).toBeGreaterThan(0);
-    expect(providerInvocations[0]?.length).toBeGreaterThan(0);
-
-    // Sanity: lastUserId from the seed is set.
-    expect(lastUserId).toBeTruthy();
-  });
-
-  it('preflight short-circuits to ok when usable budget is not exceeded', async () => {
-    const SUMMARY = 'unused';
-    const tinyMessages: AgentMessage[] = [userMsg('hi'), asstMsg('hello'), userMsg('what is 2+2?')];
-    const providerInvocations: AgentMessage[][] = [];
-    const SID = `${SESSION_ID}-small`;
-    const { iii } = buildTestSdk({
-      session_id: SID,
-      summaryText: SUMMARY,
-      providerInvocations,
-    });
-    await seedSessionFrom(iii, SID, tinyMessages);
-
-    // The default COMPACT_RESERVED_TOKENS (20_000) would zero out the 8k
-    // model's usable budget and force compaction; shrink reservation for
-    // this test so the tiny transcript stays well below usable.
-    const prev = process.env.COMPACT_RESERVED_TOKENS;
-    process.env.COMPACT_RESERVED_TOKENS = '100';
-    try {
-      const result = await runPreflight(iii, SID, tinyMessages, PROVIDER_ID, MODEL_ID);
-      expect(result).toBe('ok');
-    } finally {
-      if (prev === undefined) process.env.COMPACT_RESERVED_TOKENS = undefined;
-      else process.env.COMPACT_RESERVED_TOKENS = prev;
-    }
-
-    // Provider window is untouched (no compaction entry).
-    const after = await loadContextView(iii, SID);
-    expect(after.length).toBe(tinyMessages.length);
-
-    // Summariser was never invoked.
-    expect(providerInvocations.length).toBe(0);
-  });
-});
diff --git a/harness/tests/context-compaction/handler-async.test.ts b/harness/tests/context-compaction/handler-async.test.ts
deleted file mode 100644
index 23c03f87..00000000
--- a/harness/tests/context-compaction/handler-async.test.ts
+++ /dev/null
@@ -1,85 +0,0 @@
-import { describe, expect, it, vi } from 'vitest';
-import { parseOnTurnEnd, resolveModel } from '../../src/context-compaction/handler-async.js';
-import type { ISdk } from '../../src/runtime/iii.js';
-
-describe('parseOnTurnEnd', () => {
-  it('parses a well-formed turn_end payload', () => {
-    const out = parseOnTurnEnd({
-      session_id: 'sess-1',
-      usage: { input: 100, output: 50, cache_read: 800 },
-      provider: 'anthropic',
-      model: 'claude-haiku-4-5',
-    });
-    expect(out).toEqual({
-      session_id: 'sess-1',
-      usage: { input: 100, output: 50, cache_read: 800 },
-      provider: 'anthropic',
-      model: 'claude-haiku-4-5',
-    });
-  });
-
-  it('defaults provider/model to empty strings when absent (handler falls back to session-tree)', () => {
-    const out = parseOnTurnEnd({ session_id: 'sess-2', usage: { input: 10 } });
-    expect(out).toEqual({ session_id: 'sess-2', usage: { input: 10 }, provider: '', model: '' });
-  });
-
-  it('returns null usage when usage is missing or malformed', () => {
-    expect(parseOnTurnEnd({ session_id: 'sess-3' })?.usage).toBeNull();
-    expect(parseOnTurnEnd({ session_id: 'sess-3', usage: 'nope' })?.usage).toBeNull();
-  });
-
-  it('returns null when session_id is missing', () => {
-    expect(parseOnTurnEnd({ usage: { input: 1 } })).toBeNull();
-    expect(parseOnTurnEnd({ session_id: '' })).toBeNull();
-    expect(parseOnTurnEnd(null)).toBeNull();
-    expect(parseOnTurnEnd(42)).toBeNull();
-  });
-
-  it('passes a threaded model_limit through and drops a malformed one', () => {
-    const limit = { context: 200_000, input: 200_000, output: 64_000 };
-    expect(parseOnTurnEnd({ session_id: 'sess-4', model_limit: limit })?.model_limit).toEqual(
-      limit,
-    );
-    expect(parseOnTurnEnd({ session_id: 'sess-4', model_limit: 'nope' })).not.toHaveProperty(
-      'model_limit',
-    );
-  });
-});
-
-describe('resolveModel', () => {
-  function trackingIii(): { iii: ISdk; calls: string[] } {
-    const calls: string[] = [];
-    const iii = {
-      trigger: vi.fn(async (req: { function_id: string }) => {
-        calls.push(req.function_id);
-        return { model: { context_window: 200_000, max_output_tokens: 8_096 } };
-      }),
-    } as unknown as ISdk;
-    return { iii, calls };
-  }
-
-  it('uses the threaded limit and skips router::models::get', async () => {
-    const { iii, calls } = trackingIii();
-
-    const resolved = await resolveModel(iii, 'sess-1', 'anthropic', 'claude-sonnet-4-6', {
-      context: 1_000_000,
-      input: 1_000_000,
-      output: 64_000,
-    });
-
-    expect(calls).not.toContain('router::models::get');
-    expect(resolved).toEqual({
-      providerID: 'anthropic',
-      modelID: 'claude-sonnet-4-6',
-      modelLimit: { context: 1_000_000, input: 1_000_000, output: 64_000 },
-    });
-  });
-
-  it('falls back to router::models::get when no threaded limit is present', async () => {
-    const { iii, calls } = trackingIii();
-
-    await resolveModel(iii, 'sess-1', 'anthropic', 'claude-sonnet-4-6');
-
-    expect(calls).toContain('router::models::get');
-  });
-});
diff --git a/harness/tests/context-compaction/handler-sync.test.ts b/harness/tests/context-compaction/handler-sync.test.ts
deleted file mode 100644
index 2085edfd..00000000
--- a/harness/tests/context-compaction/handler-sync.test.ts
+++ /dev/null
@@ -1,130 +0,0 @@
-import { describe, expect, it, vi } from 'vitest';
-import { handleSync } from '../../src/context-compaction/handler-sync.js';
-import type { ISdk } from '../../src/runtime/iii.js';
-
-/**
- * In-memory state store that correctly simulates state::get / state::set
- * so the lease nonce round-trip succeeds.
- */
-function makeStubIii(triggerOverrides: Record<string, unknown> = {}): ISdk {
-  const stateStore = new Map<string, unknown>();
-
-  return {
-    trigger: vi.fn(async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-      // Honour explicit overrides first
-      if (Object.hasOwn(triggerOverrides, function_id)) {
-        const v = triggerOverrides[function_id];
-        return typeof v === 'function' ? (v as () => unknown)() : v;
-      }
-
-      const p = (payload ?? {}) as Record<string, unknown>;
-
-      // In-memory state simulation so lease nonce round-trip works
-      if (function_id === 'state::get') {
-        const v = stateStore.get(p.key as string);
-        return v !== undefined ? v : null;
-      }
-      if (function_id === 'state::set') {
-        const v = p.value;
-        if (v === null || v === undefined) {
-          stateStore.delete(p.key as string);
-        } else {
-          stateStore.set(p.key as string, v);
-        }
-        return { ok: true };
-      }
-      if (function_id === 'state::update') {
-        const key = p.key as string;
-        const ops = (p.ops ?? []) as Array<{ type: string; value?: unknown }>;
-        const oldValue = stateStore.has(key) ? stateStore.get(key) : null;
-        let newValue: unknown = oldValue;
-        for (const op of ops) {
-          if (op.type === 'set') newValue = op.value;
-        }
-        if (newValue === null || newValue === undefined) {
-          stateStore.delete(key);
-        } else {
-          stateStore.set(key, newValue);
-        }
-        return { old_value: oldValue ?? null, new_value: newValue ?? null };
-      }
-
-      return null;
-    }),
-    registerFunction: vi.fn(),
-    registerTrigger: vi.fn(),
-    publish: vi.fn(),
-    subscribe: vi.fn(),
-    enqueue: vi.fn(),
-  } as unknown as ISdk;
-}
-
-const defaultModel = {
-  id: 'claude-haiku-4-5',
-  providerID: 'anthropic',
-  limit: { context: 200_000, input: 200_000, output: 4_096 },
-};
-
-describe('handleSync smoke', () => {
-  it('returns empty when session has no messages', async () => {
-    const iii = makeStubIii({
-      // No messages → selectWithEntryIds head will be empty → 'empty'
-      'session::messages': { messages: [] },
-    });
-
-    const result = await handleSync(iii, {
-      session_id: 'test-session',
-      projected_tokens: 50_000,
-      last_user_message_id: 'msg-1',
-      model: defaultModel,
-    });
-
-    // Empty session → select returns empty head → summarizeAndAppend returns
-    // 'empty' → handler-sync surfaces it as { status: 'empty' }.
-    expect(result.status).toBe('empty');
-  });
-
-  it('returns busy when lease cannot be acquired within timeout', async () => {
-    // Pre-fill state with a fresh lease to block acquisition
-    const stateStore = new Map<string, unknown>();
-    const leaseKey = 'session/test-session-busy/compaction_lease';
-    stateStore.set(leaseKey, { nonce: 'held-by-other', ts: Date.now() - 1000 });
-
-    const iii = {
-      trigger: vi.fn(
-        async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-          const p = (payload ?? {}) as Record<string, unknown>;
-          if (function_id === 'state::get') {
-            const v = stateStore.get(p.key as string);
-            return v !== undefined ? v : null;
-          }
-          if (function_id === 'state::set') {
-            const v = p.value;
-            if (v === null || v === undefined) {
-              stateStore.delete(p.key as string);
-            } else {
-              stateStore.set(p.key as string, v);
-            }
-            return { ok: true };
-          }
-          return null;
-        },
-      ),
-      registerFunction: vi.fn(),
-      registerTrigger: vi.fn(),
-    } as unknown as ISdk;
-
-    process.env.COMPACT_BUSY_TIMEOUT_MS = '1';
-    try {
-      const result = await handleSync(iii, {
-        session_id: 'test-session-busy',
-        projected_tokens: 50_000,
-        last_user_message_id: 'msg-1',
-        model: defaultModel,
-      });
-      expect(result.status).toBe('busy');
-    } finally {
-      process.env.COMPACT_BUSY_TIMEOUT_MS = undefined;
-    }
-  });
-});
diff --git a/harness/tests/context-compaction/integration/backward-compat.test.ts b/harness/tests/context-compaction/integration/backward-compat.test.ts
deleted file mode 100644
index ac05ba87..00000000
--- a/harness/tests/context-compaction/integration/backward-compat.test.ts
+++ /dev/null
@@ -1,199 +0,0 @@
-/**
- * Backward-compatibility integration test.
- *
- * Verifies that when the path carries a prior compaction custom entry whose
- * data omits tail_start_id (old schema, anchored update path), the
- * summariser receives a system_prompt containing <previous-summary> with the
- * prior summary text.
- *
- * Also verifies that a second compaction custom entry is appended.
- */
-import { describe, expect, it, vi } from 'vitest';
-import { handleAsync } from '../../../src/context-compaction/handler-async.js';
-import type { AssistantMessage } from '../../../src/types/agent-message.js';
-import { loadFixture } from '../../fixtures/load.js';
-
-// ---------------------------------------------------------------------------
-// Helpers
-// ---------------------------------------------------------------------------
-
-type MockTriggerReq = { function_id: string; payload: Record<string, unknown>; timeoutMs?: number };
-
-const PRIOR_SUMMARY = 'Prior compaction summary: user asked about X and we explored Y.';
-
-function buildBackwardCompatMock(opts: {
-  fixtureMessages: Array<{ entry_id: string; message: unknown }>;
-  capturedSystemPrompts: string[];
-  compactPayloads: unknown[];
-}) {
-  const { fixtureMessages, capturedSystemPrompts, compactPayloads } = opts;
-
-  const stateStore = new Map<string, unknown>();
-
-  const trigger = vi.fn(async (req: MockTriggerReq) => {
-    const { function_id, payload } = req;
-
-    if (function_id === 'session::messages') {
-      // Prior compaction rides the path as a custom entry; its data omits
-      // tail_start_id (old schema) — the backward-compat scenario. Only
-      // include it on include_custom reads, like the real worker.
-      const prior = payload.include_custom
-        ? [
-            {
-              entry_id: 'comp-legacy-1',
-              custom: {
-                custom_type: 'compaction',
-                data: {
-                  summary: PRIOR_SUMMARY,
-                  tokens_before: 80_000,
-                  timestamp: Date.now() - 120_000,
-                  // Intentionally omitting tail_start_id (old schema)
-                },
-              },
-            },
-          ]
-        : [];
-      return { messages: [...prior, ...fixtureMessages] };
-    }
-    if (function_id === 'session::append') {
-      if ((payload as { custom?: unknown }).custom) compactPayloads.push(payload);
-      return { entry_id: `appended-${Date.now()}`, parent_id: null, timestamp: Date.now() };
-    }
-    if (function_id === 'session::update_message') {
-      return { updated: true, revision: 1 };
-    }
-    if (function_id === 'models::get') {
-      return {
-        id: 'claude-haiku-4-5',
-        context_window: 200_000,
-        max_output_tokens: 4_096,
-      };
-    }
-    if (function_id === 'state::get') {
-      const v = stateStore.get((payload as { key: string }).key);
-      return v !== undefined ? v : null;
-    }
-    if (function_id === 'state::set') {
-      const p = payload as { key: string; value: unknown };
-      if (p.value === null || p.value === undefined) {
-        stateStore.delete(p.key);
-      } else {
-        stateStore.set(p.key, p.value);
-      }
-      return { ok: true };
-    }
-    if (function_id === 'state::update') {
-      const p = payload as { key: string; ops: Array<{ type: string; value?: unknown }> };
-      const oldValue = stateStore.has(p.key) ? stateStore.get(p.key) : null;
-      let newValue: unknown = oldValue;
-      for (const op of p.ops ?? []) {
-        if (op.type === 'set') newValue = op.value;
-      }
-      if (newValue === null || newValue === undefined) {
-        stateStore.delete(p.key);
-      } else {
-        stateStore.set(p.key, newValue);
-      }
-      return { old_value: oldValue ?? null, new_value: newValue ?? null };
-    }
-    if (function_id === 'router::models::get') {
-      return { model: { context_window: 200_000, max_output_tokens: 4_096 } };
-    }
-    if (function_id === 'router::complete') {
-      // Capture the system_prompt sent to the summariser
-      const systemPrompt = payload.system_prompt;
-      if (typeof systemPrompt === 'string') {
-        capturedSystemPrompts.push(systemPrompt);
-      }
-
-      const msg: AssistantMessage = {
-        role: 'assistant',
-        content: [{ type: 'text', text: 'Updated anchored summary.' }],
-        stop_reason: 'end',
-        model: 'claude-haiku-4-5',
-        provider: 'anthropic',
-        timestamp: Date.now(),
-      };
-      return { message: msg };
-    }
-
-    return undefined;
-  });
-
-  const iii = { trigger } as unknown as import('../../../src/runtime/iii.js').ISdk;
-
-  return { iii, trigger };
-}
-
-// ---------------------------------------------------------------------------
-// Tests
-// ---------------------------------------------------------------------------
-
-const largeFixture = loadFixture('large-with-media');
-
-describe('backward-compat: prior compaction (anchored update path)', () => {
-  it('includes <previous-summary> in system_prompt when a legacy compaction entry exists', async () => {
-    const capturedSystemPrompts: string[] = [];
-    const compactPayloads: unknown[] = [];
-
-    const fixtureMessages = largeFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    const { iii } = buildBackwardCompatMock({
-      fixtureMessages,
-      capturedSystemPrompts,
-      compactPayloads,
-    });
-
-    // Fire a turn_end that triggers overflow
-    const frame = {
-      session_id: `${largeFixture.session_id}-compat`,
-      provider: 'anthropic',
-      model: 'claude-haiku-4-5',
-      usage: { input: 185_000, output: 0, cache_read: 0, cache_write: 0 },
-    };
-
-    await handleAsync(iii, frame);
-
-    // The system_prompt must contain the anchored <previous-summary> block
-    expect(capturedSystemPrompts).toHaveLength(1);
-    const sysPrompt = capturedSystemPrompts[0];
-    expect(sysPrompt).toBeDefined();
-    expect(sysPrompt).toContain('<previous-summary>');
-    expect(sysPrompt).toContain(PRIOR_SUMMARY);
-    expect(sysPrompt).toContain('</previous-summary>');
-  });
-
-  it('appends a compaction custom entry for the second (updated) compaction', async () => {
-    const capturedSystemPrompts: string[] = [];
-    const compactPayloads: unknown[] = [];
-
-    const fixtureMessages = largeFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    const { iii } = buildBackwardCompatMock({
-      fixtureMessages,
-      capturedSystemPrompts,
-      compactPayloads,
-    });
-
-    const frame = {
-      session_id: `${largeFixture.session_id}-compat2`,
-      provider: 'anthropic',
-      model: 'claude-haiku-4-5',
-      usage: { input: 185_000, output: 0, cache_read: 0, cache_write: 0 },
-    };
-
-    await handleAsync(iii, frame);
-
-    // The second (updated) compaction entry must be appended
-    expect(compactPayloads.length).toBeGreaterThan(0);
-    const cp = compactPayloads[0] as { custom: { data: { summary: string } } };
-    expect(typeof cp.custom.data.summary).toBe('string');
-    expect(cp.custom.data.summary.length).toBeGreaterThan(0);
-  });
-});
diff --git a/harness/tests/context-compaction/integration/flow-async.test.ts b/harness/tests/context-compaction/integration/flow-async.test.ts
deleted file mode 100644
index e3c3d306..00000000
--- a/harness/tests/context-compaction/integration/flow-async.test.ts
+++ /dev/null
@@ -1,229 +0,0 @@
-/**
- * Integration tests: async (TurnEnd-driven) compaction flow via handleAsync.
- *
- * Tests:
- * 1. Below-threshold TurnEnd → no compaction.
- * 2. Above-threshold TurnEnd (large fixture) → compaction runs, compaction custom entry appended.
- * 3. Two concurrent TurnEnds → only one compaction runs (lease serializes).
- */
-import { describe, expect, it, vi } from 'vitest';
-import { handleAsync } from '../../../src/context-compaction/handler-async.js';
-import { loadFixture } from '../../fixtures/load.js';
-
-// ---------------------------------------------------------------------------
-// Mock helpers
-// ---------------------------------------------------------------------------
-
-type MockTriggerReq = { function_id: string; payload: Record<string, unknown>; timeoutMs?: number };
-
-/**
- * Build an ISdk-compatible mock for the async flow.
- *
- * @param fixtureMessages - entries to return from session::messages
- * @param compactPayloads - collector array for compaction custom-entry appends
- * @param modelCatalog - what models::get should return (null = not found)
- */
-function buildAsyncMock(opts: {
-  fixtureMessages: Array<{ entry_id: string; message: unknown }>;
-  compactPayloads?: unknown[];
-  modelCatalog?: {
-    id?: string;
-    context_window: number;
-    max_output_tokens: number;
-  } | null;
-  stateStore?: Map<string, unknown>;
-}) {
-  const {
-    fixtureMessages,
-    compactPayloads = [],
-    modelCatalog = {
-      context_window: 200_000,
-      max_output_tokens: 4_096,
-    },
-    stateStore = new Map<string, unknown>(),
-  } = opts;
-
-  const doneSummaryMessage = {
-    role: 'assistant',
-    content: [{ type: 'text', text: 'Async compaction summary.' }],
-    stop_reason: 'end',
-    model: 'claude-haiku-4-5',
-    provider: 'anthropic',
-    timestamp: Date.now(),
-  };
-
-  let appendSeq = 0;
-  const trigger = vi.fn(async (req: MockTriggerReq) => {
-    const { function_id, payload } = req;
-
-    if (function_id === 'session::messages') {
-      return { messages: fixtureMessages };
-    }
-    if (function_id === 'session::append') {
-      if ((payload as { custom?: unknown }).custom) compactPayloads.push(payload);
-      return { entry_id: `appended-${++appendSeq}`, parent_id: null, timestamp: Date.now() };
-    }
-    if (function_id === 'session::update_message') {
-      return { updated: true, revision: 1 };
-    }
-    if (function_id === 'router::models::get') {
-      return modelCatalog ? { model: modelCatalog } : null;
-    }
-    if (function_id === 'state::get') {
-      const v = stateStore.get((payload as { key: string }).key);
-      return v !== undefined ? v : null;
-    }
-    if (function_id === 'state::set') {
-      const p = payload as { key: string; value: unknown };
-      if (p.value === null || p.value === undefined) {
-        stateStore.delete(p.key);
-      } else {
-        stateStore.set(p.key, p.value);
-      }
-      return { ok: true };
-    }
-    if (function_id === 'state::update') {
-      const p = payload as { key: string; ops: Array<{ type: string; value?: unknown }> };
-      const oldValue = stateStore.has(p.key) ? stateStore.get(p.key) : null;
-      let newValue: unknown = oldValue;
-      for (const op of p.ops ?? []) {
-        if (op.type === 'set') newValue = op.value;
-      }
-      if (newValue === null || newValue === undefined) {
-        stateStore.delete(p.key);
-      } else {
-        stateStore.set(p.key, newValue);
-      }
-      return { old_value: oldValue ?? null, new_value: newValue ?? null };
-    }
-    if (function_id === 'router::complete') {
-      return { message: doneSummaryMessage };
-    }
-
-    return undefined;
-  });
-
-  const iii = { trigger } as unknown as import('../../../src/runtime/iii.js').ISdk;
-
-  return { iii, trigger, compactPayloads };
-}
-
-/**
- * Build an on_turn_end queue payload that handleAsync can process.
- *
- * @param sessionId - the session identifier
- * @param totalTokens - total token usage to report (drives overflow check)
- * @param modelId - model identifier (default: claude-haiku-4-5)
- */
-function makeTurnEndFrame(sessionId: string, totalTokens: number, modelId = 'claude-haiku-4-5') {
-  return {
-    session_id: sessionId,
-    provider: 'anthropic',
-    model: modelId,
-    usage: { input: totalTokens, output: 0, cache_read: 0, cache_write: 0 },
-  };
-}
-
-// ---------------------------------------------------------------------------
-// Tests
-// ---------------------------------------------------------------------------
-
-const smallFixture = loadFixture('tiny');
-const largeFixture = loadFixture('large-with-media');
-
-describe('flow-async: below-threshold TurnEnd', () => {
-  it('does not append a compaction when token usage is below overflow threshold', async () => {
-    const compactPayloads: unknown[] = [];
-    const fixtureMessages = smallFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    const { iii, trigger } = buildAsyncMock({
-      fixtureMessages,
-      compactPayloads,
-      modelCatalog: {
-        context_window: 200_000,
-        max_output_tokens: 4_096,
-      },
-    });
-
-    // Use very small token count (well below the ~180k threshold for 200k context)
-    const frame = makeTurnEndFrame(smallFixture.session_id, 100);
-    await handleAsync(iii, frame);
-
-    // compact should NOT have been called
-    expect(compactPayloads).toHaveLength(0);
-
-    // Also, no session::append should appear in trigger calls at all
-    const appendCalls = trigger.mock.calls.filter(([req]) => req.function_id === 'session::append');
-    expect(appendCalls).toHaveLength(0);
-  });
-});
-
-describe('flow-async: above-threshold TurnEnd with large fixture', () => {
-  it('runs compaction and appends the compaction custom entry when tokens exceed overflow threshold', async () => {
-    const compactPayloads: unknown[] = [];
-    const fixtureMessages = largeFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    const { iii } = buildAsyncMock({
-      fixtureMessages,
-      compactPayloads,
-      modelCatalog: {
-        context_window: 200_000,
-        max_output_tokens: 4_096,
-      },
-    });
-
-    // Simulate overflow: input tokens at 185k (above 200k - 20k reserved = 180k threshold)
-    const frame = makeTurnEndFrame(largeFixture.session_id, 185_000);
-    await handleAsync(iii, frame);
-
-    // The compaction custom entry MUST have been appended at least once
-    expect(compactPayloads.length).toBeGreaterThan(0);
-
-    // Verify the compact payload shape
-    const cp = compactPayloads[0] as {
-      session_id: string;
-      custom: { custom_type: string; data: { summary: string } };
-    };
-    expect(cp.session_id).toBe(largeFixture.session_id);
-    expect(cp.custom.custom_type).toBe('compaction');
-    expect(typeof cp.custom.data.summary).toBe('string');
-    expect(cp.custom.data.summary.length).toBeGreaterThan(0);
-  });
-});
-
-describe('flow-async: concurrent TurnEnds serialized by lease', () => {
-  it('runs compaction only once when two TurnEnds fire concurrently', async () => {
-    const compactPayloads: unknown[] = [];
-    const fixtureMessages = largeFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    // Shared state store so both concurrent calls see the same lease state
-    const stateStore = new Map<string, unknown>();
-
-    const { iii } = buildAsyncMock({
-      fixtureMessages,
-      compactPayloads,
-      modelCatalog: {
-        context_window: 200_000,
-        max_output_tokens: 4_096,
-      },
-      stateStore,
-    });
-
-    // Fire both concurrently
-    const sessionId = `${largeFixture.session_id}-concurrent`;
-    const frame = makeTurnEndFrame(sessionId, 185_000);
-    await Promise.all([handleAsync(iii, frame), handleAsync(iii, frame)]);
-
-    // Only ONE compaction should have been appended despite two concurrent calls
-    expect(compactPayloads).toHaveLength(1);
-  });
-});
diff --git a/harness/tests/context-compaction/integration/flow-prune.test.ts b/harness/tests/context-compaction/integration/flow-prune.test.ts
deleted file mode 100644
index be9cd9a7..00000000
--- a/harness/tests/context-compaction/integration/flow-prune.test.ts
+++ /dev/null
@@ -1,114 +0,0 @@
-/**
- * Integration tests: prune flow using the large fixture.
- *
- * Tests:
- * 1. Large fixture with big tool output → pruned_tokens >= PRUNE_MIN_FREE (20_000).
- * 2. Pruned parts count matches session::update_message calls captured.
- */
-import { describe, expect, it, vi } from 'vitest';
-import { prune } from '../../../src/context-compaction/prune.js';
-import type { AgentMessage } from '../../../src/types/agent-message.js';
-import { loadFixture } from '../../fixtures/load.js';
-
-// ---------------------------------------------------------------------------
-// Mock helpers
-// ---------------------------------------------------------------------------
-
-const PRUNE_MIN_FREE = 20_000;
-
-type MockTriggerReq = { function_id: string; payload: Record<string, unknown>; timeoutMs?: number };
-
-function buildPruneMock(entries: Array<{ entry_id: string; message: AgentMessage }>) {
-  const updatePartCalls: Array<{ entry_id: string }> = [];
-
-  const trigger = vi.fn(async (req: MockTriggerReq) => {
-    const { function_id, payload } = req;
-
-    if (function_id === 'session::messages') {
-      return { messages: entries };
-    }
-    if (function_id === 'session::update_message') {
-      updatePartCalls.push({ entry_id: (payload as { entry_id: string }).entry_id });
-      return { updated: true, revision: 1 };
-    }
-
-    return null;
-  });
-
-  const iii = { trigger } as unknown as Parameters<typeof prune>[0];
-
-  return { iii, updatePartCalls };
-}
-
-// ---------------------------------------------------------------------------
-// Tests
-// ---------------------------------------------------------------------------
-
-const largeFixture = loadFixture('large-with-media');
-
-describe('flow-prune: large fixture with big tool output', () => {
-  it('prunes at least PRUNE_MIN_FREE tokens from the large fixture', async () => {
-    const entries = largeFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message as AgentMessage,
-    }));
-
-    const { iii } = buildPruneMock(entries);
-
-    const result = await prune(iii, largeFixture.session_id, {
-      protectTokens: 0,
-      minFree: PRUNE_MIN_FREE,
-      protectedTools: [],
-    });
-
-    // The 50KB tool output alone = ~12,500 tokens; with other tool outputs should exceed 20k
-    expect(result.pruned_tokens).toBeGreaterThanOrEqual(PRUNE_MIN_FREE);
-    expect(result.pruned_parts).toBeGreaterThan(0);
-  });
-
-  it('pruned_parts count matches the number of session::update_message calls', async () => {
-    const entries = largeFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message as AgentMessage,
-    }));
-
-    const { iii, updatePartCalls } = buildPruneMock(entries);
-
-    const result = await prune(iii, largeFixture.session_id, {
-      protectTokens: 0,
-      minFree: PRUNE_MIN_FREE,
-      protectedTools: [],
-    });
-
-    // The update_part calls must exactly match pruned_parts
-    expect(updatePartCalls).toHaveLength(result.pruned_parts);
-  });
-
-  it('does not prune protected tools', async () => {
-    const entries = largeFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message as AgentMessage,
-    }));
-
-    // Find the entry IDs of shell::run function results (which exist in the fixture)
-    const shellRunEntries = entries.filter(
-      (e) =>
-        e.message.role === 'function_result' &&
-        (e.message as { function_id?: string }).function_id === 'shell::run',
-    );
-
-    const { iii, updatePartCalls } = buildPruneMock(entries);
-
-    await prune(iii, largeFixture.session_id, {
-      protectTokens: 0,
-      minFree: PRUNE_MIN_FREE,
-      protectedTools: ['shell::run'],
-    });
-
-    // None of the protected shell::run entries should appear in update_part calls
-    const prunedIds = new Set(updatePartCalls.map((c) => c.entry_id));
-    for (const e of shellRunEntries) {
-      expect(prunedIds.has(e.entry_id)).toBe(false);
-    }
-  });
-});
diff --git a/harness/tests/context-compaction/integration/flow-sync.test.ts b/harness/tests/context-compaction/integration/flow-sync.test.ts
deleted file mode 100644
index 198714e7..00000000
--- a/harness/tests/context-compaction/integration/flow-sync.test.ts
+++ /dev/null
@@ -1,229 +0,0 @@
-/**
- * Integration tests: sync (pre-turn) compaction flow via handleSync.
- *
- * Tests:
- * 1. Success path: status === 'ok', no replay reinjection (the last user
- *    message stays on the path), a synthetic user append posts the
- *    continue nudge under the compaction entry.
- * 2. Summariser stream returns error → status === 'overflow'.
- * 3. Lease held → status === 'busy'.
- */
-import { describe, expect, it, vi } from 'vitest';
-import { payloadStoreKey, stateStoreKey } from '../../_helpers/stateStoreKey.js';
-import { handleSync } from '../../../src/context-compaction/handler-sync.js';
-import type { ISdk } from '../../../src/runtime/iii.js';
-import { loadFixture } from '../../fixtures/load.js';
-
-// ---------------------------------------------------------------------------
-// Mock helpers
-// ---------------------------------------------------------------------------
-
-type MockTriggerReq = { function_id: string; payload: Record<string, unknown>; timeoutMs?: number };
-
-const defaultModel = {
-  id: 'claude-haiku-4-5',
-  providerID: 'anthropic',
-  limit: { context: 200_000, input: 200_000, output: 4_096 },
-};
-
-/**
- * Build an ISdk-compatible mock for the sync flow.
- *
- * @param opts.fixtureMessages - entries returned by session::messages
- * @param opts.providerError - if true, router::complete rejects to simulate a summariser error
- * @param opts.stateStore - optional shared state store for lease simulation
- */
-function buildSyncMock(opts: {
-  fixtureMessages: Array<{ entry_id: string; message: unknown }>;
-  providerError?: boolean;
-  stateStore?: Map<string, unknown>;
-}) {
-  const { fixtureMessages, providerError = false, stateStore = new Map<string, unknown>() } = opts;
-
-  const doneSummaryMessage = {
-    role: 'assistant',
-    content: [{ type: 'text', text: 'Sync compaction summary text.' }],
-    stop_reason: 'end',
-    model: 'claude-haiku-4-5',
-    provider: 'anthropic',
-    timestamp: Date.now(),
-  };
-
-  const compactionAppends: unknown[] = [];
-  const messageAppends: unknown[] = [];
-  const updateMessageCalls: unknown[] = [];
-  let appendSeq = 0;
-
-  const trigger = vi.fn(async (req: MockTriggerReq) => {
-    const { function_id, payload } = req;
-
-    if (function_id === 'session::messages') {
-      return { messages: fixtureMessages };
-    }
-    if (function_id === 'session::append') {
-      if ((payload as { custom?: unknown }).custom) compactionAppends.push(payload);
-      else messageAppends.push(payload);
-      return { entry_id: `appended-${++appendSeq}`, parent_id: null, timestamp: Date.now() };
-    }
-    if (function_id === 'session::update_message') {
-      updateMessageCalls.push(payload);
-      return { updated: true, revision: 1 };
-    }
-    if (function_id === 'state::get') {
-      const v = stateStore.get(payloadStoreKey(payload as { scope?: string; key?: string }));
-      return v !== undefined ? v : null;
-    }
-    if (function_id === 'state::set') {
-      const p = payload as { key: string; value: unknown; scope?: string };
-      const storeKey = payloadStoreKey(p);
-      if (p.value === null || p.value === undefined) {
-        stateStore.delete(storeKey);
-      } else {
-        stateStore.set(storeKey, p.value);
-      }
-      return { ok: true };
-    }
-    if (function_id === 'state::update') {
-      const p = payload as {
-        key: string;
-        scope?: string;
-        ops: Array<{ type: string; value?: unknown }>;
-      };
-      const storeKey = payloadStoreKey(p);
-      const oldValue = stateStore.has(storeKey) ? stateStore.get(storeKey) : null;
-      let newValue: unknown = oldValue;
-      for (const op of p.ops ?? []) {
-        if (op.type === 'set') newValue = op.value;
-      }
-      if (newValue === null || newValue === undefined) {
-        stateStore.delete(storeKey);
-      } else {
-        stateStore.set(storeKey, newValue);
-      }
-      return { old_value: oldValue ?? null, new_value: newValue ?? null };
-    }
-    if (function_id === 'router::complete') {
-      if (providerError) {
-        throw new Error('summariser unavailable');
-      }
-      return { message: doneSummaryMessage };
-    }
-
-    return undefined;
-  });
-
-  const iii = { trigger } as unknown as ISdk;
-
-  return { iii, trigger, compactionAppends, messageAppends, updateMessageCalls };
-}
-
-// ---------------------------------------------------------------------------
-// Tests
-// ---------------------------------------------------------------------------
-
-const mediumFixture = loadFixture('medium-with-tools');
-
-describe('flow-sync: success path', () => {
-  it('returns ok and appends the continue nudge (no replay reinjection)', async () => {
-    const fixtureMessages = mediumFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    // Use the second-to-last user message as the replay target
-    const userEntries = mediumFixture.entries.filter((e) => {
-      const msg = e.message as { role: string };
-      return msg.role === 'user';
-    });
-    // Pick a user entry that is NOT the very first so extractReplayTarget finds a truncated head
-    const lastUserEntry = userEntries.at(-1);
-    expect(lastUserEntry).toBeDefined();
-
-    const { iii, compactionAppends, messageAppends } = buildSyncMock({ fixtureMessages });
-
-    const result = await handleSync(iii, {
-      session_id: mediumFixture.session_id,
-      projected_tokens: 150_000,
-      last_user_message_id: lastUserEntry.id,
-      model: defaultModel,
-    });
-
-    expect(result.status).toBe('ok');
-
-    // The compaction record lands as one custom entry.
-    expect(compactionAppends).toHaveLength(1);
-
-    // The last user message is already on the path as the compaction's
-    // parent; the ONLY message append is the synthetic continue nudge,
-    // chained under the compaction entry.
-    expect(messageAppends).toHaveLength(1);
-    const nudge = messageAppends[0] as {
-      session_id: string;
-      parent_id?: string;
-      origin?: Record<string, unknown>;
-      message: { role: string; content: Array<{ type: string; text?: string }> };
-    };
-    expect(nudge.session_id).toBe(mediumFixture.session_id);
-    expect(nudge.message.role).toBe('user');
-    expect(nudge.message.content[0]?.text).toContain('Continue');
-    expect(nudge.origin).toEqual({ synthetic: true });
-  });
-});
-
-describe('flow-sync: summariser error → overflow', () => {
-  it('returns overflow when the summariser stream fails', async () => {
-    const fixtureMessages = mediumFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    // providerError=true makes createChannel reject, causing summarizeAndAppend to return 'compact'
-    const { iii } = buildSyncMock({ fixtureMessages, providerError: true });
-
-    const result = await handleSync(iii, {
-      session_id: `${mediumFixture.session_id}-err`,
-      projected_tokens: 150_000,
-      last_user_message_id: 'nonexistent-last-msg',
-      model: defaultModel,
-    });
-
-    expect(result.status).toBe('overflow');
-  });
-});
-
-describe('flow-sync: lease held → busy', () => {
-  it('returns busy when another compaction is already holding the lease', async () => {
-    const fixtureMessages = mediumFixture.entries.map((e) => ({
-      entry_id: e.id,
-      message: e.message,
-    }));
-
-    // Pre-populate the state store with an active lease for this session
-    const sessionId = `${mediumFixture.session_id}-busy`;
-    const stateStore = new Map<string, unknown>();
-    const leaseStoreKey = stateStoreKey('compaction_lease', sessionId);
-    stateStore.set(leaseStoreKey, { nonce: 'held-by-another', ts: Date.now() - 1000 });
-
-    const { iii } = buildSyncMock({ fixtureMessages, stateStore });
-
-    // Set a very short busy timeout so the test completes quickly
-    const original = process.env.COMPACT_BUSY_TIMEOUT_MS;
-    process.env.COMPACT_BUSY_TIMEOUT_MS = '1';
-    try {
-      const result = await handleSync(iii, {
-        session_id: sessionId,
-        projected_tokens: 150_000,
-        last_user_message_id: 'any-msg',
-        model: defaultModel,
-      });
-
-      expect(result.status).toBe('busy');
-    } finally {
-      if (original === undefined) {
-        delete process.env.COMPACT_BUSY_TIMEOUT_MS;
-      } else {
-        process.env.COMPACT_BUSY_TIMEOUT_MS = original;
-      }
-    }
-  });
-});
diff --git a/harness/tests/context-compaction/lease.test.ts b/harness/tests/context-compaction/lease.test.ts
deleted file mode 100644
index f0fbf21a..00000000
--- a/harness/tests/context-compaction/lease.test.ts
+++ /dev/null
@@ -1,272 +0,0 @@
-import { describe, expect, it, vi } from 'vitest';
-import { payloadStoreKey, stateStoreKey } from '../_helpers/stateStoreKey.js';
-import {
-  LEASE_TTL_SECS,
-  mintLeaseNonce,
-  readLeaseTimestampSecs,
-  acquireLease,
-  acquireLeaseWithWait,
-} from '../../src/context-compaction/lease.js';
-
-describe('lease helpers', () => {
-  it('stateStoreKey namespaces compaction lease by session', () => {
-    const k = stateStoreKey('compaction_lease', 's9');
-    expect(k).toBe('compaction_lease/s9');
-  });
-
-  it('mintLeaseNonce produces unique values across rapid calls', () => {
-    const seen = new Set<string>();
-    for (let i = 0; i < 1000; i++) seen.add(mintLeaseNonce());
-    expect(seen.size).toBe(1000);
-  });
-
-  it('readLeaseTimestampSecs reads new {nonce, ts} shape', () => {
-    expect(readLeaseTimestampSecs({ nonce: 'a', ts: 1_700_000_000_000 })).toBe(1_700_000_000);
-  });
-
-  it('readLeaseTimestampSecs treats bare-int values as inactive', () => {
-    expect(readLeaseTimestampSecs(1_700_000_000)).toBe(0);
-  });
-
-  it('readLeaseTimestampSecs returns 0 for garbage', () => {
-    expect(readLeaseTimestampSecs(null)).toBe(0);
-    expect(readLeaseTimestampSecs('abc')).toBe(0);
-    expect(readLeaseTimestampSecs({ unrelated: true })).toBe(0);
-  });
-
-  it('LEASE_TTL_SECS exceeds the summariser timeout (120s)', () => {
-    expect(LEASE_TTL_SECS).toBeGreaterThan(120);
-  });
-});
-
-// Minimal in-memory ISdk stub for lease integration tests.
-// state::update is atomic by construction (Map ops are synchronous), which
-// matches what we need from iii-database in production.
-function makeStateIii() {
-  const store = new Map<string, unknown>();
-  const iii = {
-    trigger: vi.fn(async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-      const p = payload as Record<string, unknown>;
-      if (function_id === 'state::get') {
-        const v = store.get(payloadStoreKey(p as { scope?: string; key?: string }));
-        return v !== undefined ? v : null;
-      }
-      if (function_id === 'state::set') {
-        const v = p.value;
-        const key = payloadStoreKey(p as { scope?: string; key?: string });
-        if (v === null || v === undefined) {
-          store.delete(key);
-        } else {
-          store.set(key, v);
-        }
-        return { ok: true };
-      }
-      if (function_id === 'state::update') {
-        const key = payloadStoreKey(p as { scope?: string; key?: string });
-        const ops = (p.ops ?? []) as Array<{ type: string; value?: unknown }>;
-        const oldValue = store.has(key) ? store.get(key) : null;
-        let newValue: unknown = oldValue;
-        for (const op of ops) {
-          if (op.type === 'set') newValue = op.value;
-        }
-        if (newValue === null || newValue === undefined) {
-          store.delete(key);
-        } else {
-          store.set(key, newValue);
-        }
-        return { old_value: oldValue ?? null, new_value: newValue ?? null };
-      }
-      return null;
-    }),
-  };
-  return { iii: iii as unknown as Parameters<typeof acquireLease>[0], store };
-}
-
-describe('lease kinds', () => {
-  it('compaction and prune leases are independent for the same session', async () => {
-    const { iii } = makeStateIii();
-    const nonce1 = await acquireLease(iii, 'sid', 'compaction');
-    const nonce2 = await acquireLease(iii, 'sid', 'prune');
-    // Both should succeed — they use different state keys.
-    expect(nonce1).not.toBeNull();
-    expect(nonce2).not.toBeNull();
-    // And they must be distinct nonces.
-    expect(nonce1).not.toBe(nonce2);
-  });
-
-  it('stateStoreKey produces different keys for compaction vs prune', () => {
-    const k1 = stateStoreKey('compaction_lease', 'sess1');
-    const k2 = stateStoreKey('prune_lease', 'sess1');
-    expect(k1).toBe('compaction_lease/sess1');
-    expect(k2).toBe('prune_lease/sess1');
-    expect(k1).not.toBe(k2);
-  });
-});
-
-describe('acquireLeaseWithWait', () => {
-  it('returns null when lease never frees up within timeout', async () => {
-    const { iii } = makeStateIii();
-    // Pre-occupy the compaction lease so subsequent acquires fail.
-    const first = await acquireLease(iii, 'sid2', 'compaction');
-    expect(first).not.toBeNull();
-
-    // Now acquireLeaseWithWait should give up quickly (50 ms timeout).
-    const result = await acquireLeaseWithWait(iii, 'sid2', 'compaction', 50);
-    expect(result).toBeNull();
-  });
-});
-
-// Adds artificial latency to BOTH state::set and state::update, mimicking
-// a real state store (iii-database over IPC, remote Redis). The key
-// invariant is that state::update remains atomic — read+write inside a
-// single async function tick — which is what iii-database guarantees.
-function makeRacyStateIii(writeLatencyMs: number) {
-  const store = new Map<string, unknown>();
-  const iii = {
-    trigger: vi.fn(async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-      const p = payload as Record<string, unknown>;
-      if (function_id === 'state::get') {
-        const v = store.get(payloadStoreKey(p as { scope?: string; key?: string }));
-        return v !== undefined ? v : null;
-      }
-      if (function_id === 'state::set') {
-        await new Promise((r) => setTimeout(r, writeLatencyMs));
-        const v = p.value;
-        const key = payloadStoreKey(p as { scope?: string; key?: string });
-        if (v === null || v === undefined) {
-          store.delete(key);
-        } else {
-          store.set(key, v);
-        }
-        return { ok: true };
-      }
-      if (function_id === 'state::update') {
-        await new Promise((r) => setTimeout(r, writeLatencyMs));
-        const key = payloadStoreKey(p as { scope?: string; key?: string });
-        const ops = (p.ops ?? []) as Array<{ type: string; value?: unknown }>;
-        const oldValue = store.has(key) ? store.get(key) : null;
-        let newValue: unknown = oldValue;
-        for (const op of ops) {
-          if (op.type === 'set') newValue = op.value;
-        }
-        if (newValue === null || newValue === undefined) {
-          store.delete(key);
-        } else {
-          store.set(key, newValue);
-        }
-        return { old_value: oldValue ?? null, new_value: newValue ?? null };
-      }
-      return null;
-    }),
-  };
-  return { iii: iii as unknown as Parameters<typeof acquireLease>[0], store };
-}
-
-// Simulates a transient state-store outage: state::update fails (returns
-// null) while state::get still works. acquireLease must refuse to claim
-// the lease in this case — otherwise every concurrent caller becomes a
-// "winner" during the outage and the single-writer invariant breaks.
-function makeFailingUpdateIii() {
-  const store = new Map<string, unknown>();
-  const iii = {
-    trigger: vi.fn(async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-      const p = payload as Record<string, unknown>;
-      if (function_id === 'state::get') {
-        const v = store.get(payloadStoreKey(p as { scope?: string; key?: string }));
-        return v !== undefined ? v : null;
-      }
-      if (function_id === 'state::set') {
-        const v = p.value;
-        const key = payloadStoreKey(p as { scope?: string; key?: string });
-        if (v === null || v === undefined) {
-          store.delete(key);
-        } else {
-          store.set(key, v);
-        }
-        return { ok: true };
-      }
-      if (function_id === 'state::update') {
-        // Simulate the stateUpdate wrapper's error path: backend failed,
-        // wrapper swallowed and returned null.
-        return null;
-      }
-      return null;
-    }),
-  };
-  return { iii: iii as unknown as Parameters<typeof acquireLease>[0], store };
-}
-
-describe('acquireLease wire shape', () => {
-  it('sends state::update with path: "" (root) on the set op', async () => {
-    // Regression: the engine UpdateOp::Set requires a path field
-    // (FieldPath; "" = root). A previous version of the lease shipped
-    // ops without `path`, the engine failed to deserialize, the wrapper
-    // returned null, and every /compact bailed in production.
-    const captured: Array<{ type: string; path?: unknown; value?: unknown }> = [];
-    const iii = {
-      trigger: vi.fn(
-        async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-          const p = payload as Record<string, unknown>;
-          if (function_id === 'state::get') return null;
-          if (function_id === 'state::update') {
-            const ops = (p.ops ?? []) as Array<{
-              type: string;
-              path?: unknown;
-              value?: unknown;
-            }>;
-            for (const op of ops) captured.push(op);
-            return { old_value: null, new_value: { dummy: true } };
-          }
-          return null;
-        },
-      ),
-    };
-    const nonce = await acquireLease(
-      iii as unknown as Parameters<typeof acquireLease>[0],
-      'wireshape-sid',
-      'compaction',
-    );
-    expect(nonce).not.toBeNull();
-    expect(captured).toHaveLength(1);
-    expect(captured[0]?.type).toBe('set');
-    expect(captured[0]?.path).toBe('');
-    expect(captured[0]?.value).toMatchObject({ nonce, ts: expect.any(Number) });
-  });
-});
-
-describe('acquireLease state-store failures', () => {
-  it('returns null when state::update fails (does not falsely claim the lease)', async () => {
-    const { iii } = makeFailingUpdateIii();
-    const got = await acquireLease(iii, 'failsid', 'compaction');
-    expect(got).toBeNull();
-  });
-
-  it('still returns null for every concurrent caller during an outage', async () => {
-    const { iii } = makeFailingUpdateIii();
-    const results = await Promise.all(
-      Array.from({ length: 4 }, () => acquireLease(iii, 'failsid-many', 'compaction')),
-    );
-    expect(results.every((r) => r === null)).toBe(true);
-  });
-});
-
-describe('acquireLease concurrency', () => {
-  it('lets exactly one of two concurrent acquirers win even with write latency', async () => {
-    const { iii } = makeRacyStateIii(10);
-    const results = await Promise.all([
-      acquireLease(iii, 'race-sid', 'compaction'),
-      acquireLease(iii, 'race-sid', 'compaction'),
-    ]);
-    const winners = results.filter((r) => r !== null);
-    expect(winners.length).toBe(1);
-  });
-
-  it('lets exactly one of many concurrent acquirers win', async () => {
-    const { iii } = makeRacyStateIii(10);
-    const results = await Promise.all(
-      Array.from({ length: 8 }, () => acquireLease(iii, 'race-sid-multi', 'compaction')),
-    );
-    const winners = results.filter((r) => r !== null);
-    expect(winners.length).toBe(1);
-  });
-});
diff --git a/harness/tests/context-compaction/overflow.test.ts b/harness/tests/context-compaction/overflow.test.ts
deleted file mode 100644
index 89fa849c..00000000
--- a/harness/tests/context-compaction/overflow.test.ts
+++ /dev/null
@@ -1,81 +0,0 @@
-// tests/context-compaction/overflow.test.ts
-import { describe, expect, it } from 'vitest';
-import { usable, isOverflow, preserveRecentBudget } from '../../src/context-compaction/overflow.js';
-
-const MODEL_INPUT_LIMITED = {
-  id: 'm1',
-  limit: { context: 200_000, input: 180_000, output: 20_000 },
-};
-const MODEL_CONTEXT_ONLY = {
-  id: 'm2',
-  limit: { context: 100_000, input: 0, output: 8_000 },
-};
-
-describe('usable', () => {
-  it('returns input - reserved for input-limited models', () => {
-    expect(usable({ model: MODEL_INPUT_LIMITED, reserved: 20_000 })).toBe(160_000);
-  });
-  it('returns context - maxOutput for context-only models', () => {
-    expect(usable({ model: MODEL_CONTEXT_ONLY, reserved: 20_000 })).toBe(92_000);
-  });
-  it('returns 0 for a zero-context model', () => {
-    const m = { id: 'm3', limit: { context: 0, input: 0, output: 0 } };
-    expect(usable({ model: m, reserved: 1000 })).toBe(0);
-  });
-});
-
-describe('isOverflow', () => {
-  it('false when auto disabled', () => {
-    expect(
-      isOverflow({
-        tokens: { input: 200_000, output: 0, cache_read: 0, cache_write: 0 },
-        model: MODEL_INPUT_LIMITED,
-        reserved: 20_000,
-        autoDisabled: true,
-      }),
-    ).toBe(false);
-  });
-  it('false for zero-context model', () => {
-    const m = { id: 'z', limit: { context: 0, input: 0, output: 0 } };
-    expect(
-      isOverflow({
-        tokens: { input: 50_000 },
-        model: m,
-        reserved: 1000,
-      }),
-    ).toBe(false);
-  });
-  it('sums tokens.total when present', () => {
-    expect(
-      isOverflow({
-        tokens: { total: 161_000 },
-        model: MODEL_INPUT_LIMITED,
-        reserved: 20_000,
-      }),
-    ).toBe(true);
-  });
-  it('falls back to input + output + cache_read + cache_write', () => {
-    expect(
-      isOverflow({
-        tokens: { input: 100_000, output: 50_000, cache_read: 10_000, cache_write: 1_000 },
-        model: MODEL_INPUT_LIMITED,
-        reserved: 20_000,
-      }),
-    ).toBe(true);
-  });
-});
-
-describe('preserveRecentBudget', () => {
-  it('clamps to MIN_PRESERVE when usable * 0.25 too small', () => {
-    const tiny = { id: 't', limit: { context: 8000, input: 6000, output: 2000 } };
-    expect(preserveRecentBudget({ model: tiny, reserved: 1000 })).toBe(2_000);
-  });
-  it('clamps to MAX_PRESERVE when usable * 0.25 too large', () => {
-    expect(preserveRecentBudget({ model: MODEL_INPUT_LIMITED, reserved: 20_000 })).toBe(8_000);
-  });
-  it('honors override', () => {
-    expect(
-      preserveRecentBudget({ model: MODEL_INPUT_LIMITED, reserved: 20_000, override: 5_000 }),
-    ).toBe(5_000);
-  });
-});
diff --git a/harness/tests/context-compaction/prune.test.ts b/harness/tests/context-compaction/prune.test.ts
deleted file mode 100644
index 7bb67617..00000000
--- a/harness/tests/context-compaction/prune.test.ts
+++ /dev/null
@@ -1,95 +0,0 @@
-import { describe, expect, it, vi } from 'vitest';
-import { prune } from '../../src/context-compaction/prune.js';
-import type { AgentMessage } from '../../src/types/agent-message.js';
-
-function user(): AgentMessage {
-  return { role: 'user', content: [{ type: 'text', text: 'q' }], timestamp: 0 };
-}
-function asst(): AgentMessage {
-  return {
-    role: 'assistant',
-    content: [{ type: 'text', text: 'a' }],
-    stop_reason: 'end',
-    model: 'm',
-    provider: 'p',
-    timestamp: 0,
-  };
-}
-function tool(fn: string, output: string): AgentMessage {
-  return {
-    role: 'function_result',
-    function_call_id: 'fc',
-    function_id: fn,
-    content: [{ type: 'text', text: output }],
-    details: {},
-    is_error: false,
-    timestamp: 0,
-  };
-}
-
-function makeIii(entries: Array<{ entry_id: string; message: AgentMessage }>) {
-  const updates: Array<{ entry_id: string; details?: unknown }> = [];
-  const iii = {
-    trigger: vi.fn(async ({ function_id, payload }: { function_id: string; payload: unknown }) => {
-      if (function_id === 'session::messages') return { messages: entries };
-      if (function_id === 'session::update_message') {
-        const p = payload as { entry_id: string; details?: unknown };
-        updates.push({ entry_id: p.entry_id, details: p.details });
-        return { updated: true, revision: 1 };
-      }
-      return null;
-    }),
-  };
-  return { iii: iii as unknown as Parameters<typeof prune>[0], updates };
-}
-
-describe('prune', () => {
-  it('returns pruned: 0 when nothing exceeds protectTokens', async () => {
-    const entries = [
-      { entry_id: 'e1', message: user() },
-      { entry_id: 'e2', message: asst() },
-      { entry_id: 'e3', message: tool('shell::fs::ls', 'small') },
-      { entry_id: 'e4', message: user() },
-      { entry_id: 'e5', message: asst() },
-    ];
-    const { iii, updates } = makeIii(entries);
-    const res = await prune(iii, 'sid', {
-      protectTokens: 40_000,
-      minFree: 20_000,
-      protectedTools: [],
-    });
-    expect(res.pruned_parts).toBe(0);
-    expect(updates).toHaveLength(0);
-  });
-
-  it('prunes older tool outputs beyond protectTokens once minFree is hit', async () => {
-    const big = 'x'.repeat(200_000);
-    const entries: Array<{ entry_id: string; message: AgentMessage }> = [];
-    for (let i = 0; i < 5; i++) {
-      entries.push({ entry_id: `u${i}`, message: user() });
-      entries.push({ entry_id: `a${i}`, message: asst() });
-      entries.push({ entry_id: `t${i}`, message: tool('shell::fs::cat', big) });
-    }
-    const { iii, updates } = makeIii(entries);
-    const res = await prune(iii, 'sid', {
-      protectTokens: 40_000,
-      minFree: 20_000,
-      protectedTools: [],
-    });
-    expect(res.pruned_parts).toBeGreaterThan(0);
-    expect(updates.length).toBe(res.pruned_parts);
-  });
-
-  it('skips parts in protectedTools', async () => {
-    const big = 'x'.repeat(200_000);
-    const entries: Array<{ entry_id: string; message: AgentMessage }> = [];
-    for (let i = 0; i < 3; i++) {
-      entries.push({ entry_id: `u${i}`, message: user() });
-    }
-    entries.push({ entry_id: 't1', message: tool('skill::ask', big) });
-    entries.push({ entry_id: 't2', message: tool('shell::fs::cat', big) });
-    const { iii, updates } = makeIii(entries);
-    await prune(iii, 'sid', { protectTokens: 0, minFree: 1000, protectedTools: ['skill::ask'] });
-    expect(updates.some((u) => u.entry_id === 't1')).toBe(false);
-  });
-});
diff --git a/harness/tests/context-compaction/registration.test.ts b/harness/tests/context-compaction/registration.test.ts
index dd427efb..a0a82df6 100644
--- a/harness/tests/context-compaction/registration.test.ts
+++ b/harness/tests/context-compaction/registration.test.ts
@@ -3,7 +3,7 @@ import { register } from '../../src/context-compaction/register.js';
 import type { ISdk } from '../../src/runtime/iii.js';
 
 describe('context-compaction registration', () => {
-  it('registers on_turn_end as a function and no stream trigger (woken by a queue message)', async () => {
+  it('registers only compact_session (the thin /compact wrapper) and no triggers', async () => {
     const registerFunction = vi.fn();
     const registerTrigger = vi.fn();
     const iii = {
@@ -14,11 +14,14 @@ describe('context-compaction registration', () => {
 
     await register(iii);
 
-    // Compaction is woken by a turn-orchestrator queue message, not a stream.
+    // The async post-turn compactor and its wakes are retired — context-manager
+    // owns assemble/compact, and the harness compacts inline on the hot path.
     expect(registerTrigger).not.toHaveBeenCalled();
 
     const fns = registerFunction.mock.calls.map((c) => c[0] as string);
-    expect(fns).toContain('context-compaction::on_turn_end');
-    expect(fns).not.toContain('context-compaction::on_agent_event');
+    expect(fns).toEqual(['context-compaction::compact_session']);
+    expect(fns).not.toContain('context-compaction::on_turn_end');
+    expect(fns).not.toContain('context-compaction::compact_now');
+    expect(fns).not.toContain('context-compaction::prune_tool_outputs');
   });
 });
diff --git a/harness/tests/context-compaction/replay.test.ts b/harness/tests/context-compaction/replay.test.ts
deleted file mode 100644
index 0978cb01..00000000
--- a/harness/tests/context-compaction/replay.test.ts
+++ /dev/null
@@ -1,39 +0,0 @@
-import { describe, expect, it } from 'vitest';
-import { extractReplayTarget } from '../../src/context-compaction/replay.js';
-import type { AgentMessage } from '../../src/types/agent-message.js';
-
-const user = (text = 'q'): { entry_id: string; message: AgentMessage } => ({
-  entry_id: `u-${text}`,
-  message: { role: 'user', content: [{ type: 'text', text }], timestamp: 0 },
-});
-const asst = (text = 'a'): { entry_id: string; message: AgentMessage } => ({
-  entry_id: `a-${text}`,
-  message: {
-    role: 'assistant',
-    content: [{ type: 'text', text }],
-    stop_reason: 'end',
-    model: 'm',
-    provider: 'p',
-    timestamp: 0,
-  },
-});
-
-describe('extractReplayTarget', () => {
-  it('extracts the message matching lastUserMessageId', () => {
-    const entries = [user('q1'), asst('a1'), user('q2')];
-    const { replay, truncatedMessages } = extractReplayTarget(entries, 'u-q2');
-    expect(replay?.entry_id).toBe('u-q2');
-    expect(truncatedMessages.map((e) => e.entry_id)).toEqual(['u-q1', 'a-a1']);
-  });
-  it('returns undefined when target not found', () => {
-    const entries = [user('q1'), asst('a1')];
-    const { replay, truncatedMessages } = extractReplayTarget(entries, 'missing');
-    expect(replay).toBeUndefined();
-    expect(truncatedMessages).toEqual(entries);
-  });
-  it('returns undefined when target is not a user message', () => {
-    const entries = [user('q1'), asst('a1')];
-    const { replay } = extractReplayTarget(entries, 'a-a1');
-    expect(replay).toBeUndefined();
-  });
-});
diff --git a/harness/tests/context-compaction/selection.test.ts b/harness/tests/context-compaction/selection.test.ts
deleted file mode 100644
index c24d9bd7..00000000
--- a/harness/tests/context-compaction/selection.test.ts
+++ /dev/null
@@ -1,163 +0,0 @@
-import { describe, expect, it } from 'vitest';
-import {
-  completedCompactions,
-  select,
-  selectWithEntryIds,
-  splitTurn,
-  turns,
-} from '../../src/context-compaction/selection.js';
-import type { AgentMessage } from '../../src/types/agent-message.js';
-
-function user(text = ''): AgentMessage {
-  return { role: 'user', content: [{ type: 'text', text }], timestamp: 0 };
-}
-function asst(text = ''): AgentMessage {
-  return {
-    role: 'assistant',
-    content: [{ type: 'text', text }],
-    stop_reason: 'end',
-    model: 'm',
-    provider: 'p',
-    timestamp: 0,
-  };
-}
-
-describe('turns', () => {
-  it('groups messages into user→...→user turns', () => {
-    const msgs = [user('q1'), asst('a1'), asst('a1b'), user('q2'), asst('a2')];
-    const ts = turns(msgs);
-    expect(ts).toEqual([
-      { index: 0, start: 0, end: 3 },
-      { index: 3, start: 3, end: 5 },
-    ]);
-  });
-  it('returns empty array for no-user transcript', () => {
-    expect(turns([asst('x')])).toEqual([]);
-  });
-  it('produces non-overlapping contiguous ranges', () => {
-    const msgs = [user(), asst(), user(), asst(), user(), asst()];
-    const ts = turns(msgs);
-    expect(ts[0]?.end).toBe(ts[1]?.start);
-    expect(ts[1]?.end).toBe(ts[2]?.start);
-    expect(ts[2]?.end).toBe(msgs.length);
-  });
-});
-
-describe('splitTurn', () => {
-  const fakeEstimate = (msgs: AgentMessage[]) => msgs.length * 1000;
-  it('returns undefined when budget is non-positive', () => {
-    const msgs = [user(), asst(), asst()];
-    expect(
-      splitTurn({
-        messages: msgs,
-        turn: { index: 0, start: 0, end: 3 },
-        budget: 0,
-        estimate: fakeEstimate,
-      }),
-    ).toBeUndefined();
-  });
-  it('returns undefined for single-message turns', () => {
-    const msgs = [user()];
-    expect(
-      splitTurn({
-        messages: msgs,
-        turn: { index: 0, start: 0, end: 1 },
-        budget: 999_999,
-        estimate: fakeEstimate,
-      }),
-    ).toBeUndefined();
-  });
-  it('finds the first split point that fits the budget (walking forward)', () => {
-    const msgs = [user(), asst(), asst(), asst(), asst()];
-    const out = splitTurn({
-      messages: msgs,
-      turn: { index: 0, start: 0, end: 5 },
-      budget: 2_500,
-      estimate: fakeEstimate,
-    });
-    expect(out).toEqual({ start: 3, index: 3 });
-  });
-  it('returns undefined when no split fits', () => {
-    const msgs = [user(), asst(), asst()];
-    const out = splitTurn({
-      messages: msgs,
-      turn: { index: 0, start: 0, end: 3 },
-      budget: 500,
-      estimate: fakeEstimate,
-    });
-    expect(out).toBeUndefined();
-  });
-});
-
-describe('select', () => {
-  const sizeEach = 1000;
-  const estimate = (msgs: AgentMessage[]) => msgs.length * sizeEach;
-  it('keeps full tail when budget allows', () => {
-    const msgs = [user('a'), asst(), user('b'), asst()];
-    const r = select({ messages: msgs, budget: 99_999, tailTurns: 2, estimate });
-    expect(r.head).toEqual(msgs);
-    expect(r.tail_start_index).toBeUndefined();
-  });
-  it('walks backward, includes a turn that fits', () => {
-    const msgs = [user('a'), asst(), asst(), user('b'), asst()];
-    const r = select({ messages: msgs, budget: 2_000, tailTurns: 2, estimate });
-    expect(r.tail_start_index).toBe(3);
-    expect(r.head.length).toBe(3);
-  });
-  it('falls back to splitTurn when a single turn exceeds budget', () => {
-    const msgs = [user(), asst(), asst(), asst(), asst()];
-    const r = select({ messages: msgs, budget: 2_500, tailTurns: 2, estimate });
-    expect(r.tail_start_index).toBe(3);
-  });
-  it('returns full messages when tailTurns is 0', () => {
-    const msgs = [user(), asst()];
-    const r = select({ messages: msgs, budget: 999, tailTurns: 0, estimate });
-    expect(r.head).toEqual(msgs);
-    expect(r.tail_start_index).toBeUndefined();
-  });
-  it('returns full messages when no turns found', () => {
-    const r = select({ messages: [asst()], budget: 999, tailTurns: 2, estimate });
-    expect(r.tail_start_index).toBeUndefined();
-  });
-});
-
-describe('selectWithEntryIds', () => {
-  it('maps tail_start_index back to entry_id', () => {
-    const entries = [
-      { entry_id: 'e1', message: user() },
-      { entry_id: 'e2', message: asst() },
-      { entry_id: 'e3', message: user() },
-      { entry_id: 'e4', message: asst() },
-    ];
-    const r = selectWithEntryIds({
-      entries,
-      budget: 2_000,
-      tailTurns: 1,
-      estimate: (m) => m.length * 1000,
-    });
-    expect(r.tail_start_id).toBe('e3');
-    expect(r.head.length).toBe(2);
-  });
-});
-
-describe('completedCompactions', () => {
-  it('returns summaries in timestamp order', () => {
-    const entries = [
-      { id: 'c2', summary: 'second', tokens_before: 100, timestamp: 200 },
-      { id: 'c1', summary: 'first', tokens_before: 50, timestamp: 100 },
-    ];
-    expect(completedCompactions(entries)).toEqual([
-      { summary: 'first', userIndex: -1, assistantIndex: -1 },
-      { summary: 'second', userIndex: -1, assistantIndex: -1 },
-    ]);
-  });
-  it('skips entries with empty summary', () => {
-    const entries = [
-      { id: 'c1', summary: '', tokens_before: 0, timestamp: 100 },
-      { id: 'c2', summary: 'ok', tokens_before: 0, timestamp: 200 },
-    ];
-    expect(completedCompactions(entries)).toEqual([
-      { summary: 'ok', userIndex: -1, assistantIndex: -1 },
-    ]);
-  });
-});
diff --git a/harness/tests/context-compaction/strip-media.test.ts b/harness/tests/context-compaction/strip-media.test.ts
deleted file mode 100644
index 6d16fbce..00000000
--- a/harness/tests/context-compaction/strip-media.test.ts
+++ /dev/null
@@ -1,77 +0,0 @@
-import { describe, expect, it } from 'vitest';
-import { stripMedia } from '../../src/context-compaction/strip-media.js';
-import type { AgentMessage } from '../../src/types/agent-message.js';
-
-function user(): AgentMessage {
-  return {
-    role: 'user',
-    content: [
-      { type: 'text', text: 'hi' },
-      // ContentBlock allows image blocks; use a structural shape.
-      { type: 'image' } as unknown as { type: 'text'; text: string },
-    ],
-    timestamp: 0,
-  };
-}
-
-function bigToolResult(): AgentMessage {
-  return {
-    role: 'function_result',
-    function_call_id: 'fc1',
-    function_id: 'shell::fs::cat',
-    content: [{ type: 'text', text: 'x'.repeat(10_000) }],
-    details: {},
-    is_error: false,
-    timestamp: 0,
-  };
-}
-
-describe('stripMedia', () => {
-  it('replaces image content blocks with text placeholders', () => {
-    const out = stripMedia([user()], { toolOutputMaxChars: 2000 });
-    expect(out[0]?.content).toEqual([
-      { type: 'text', text: 'hi' },
-      { type: 'text', text: '[image stripped]' },
-    ]);
-  });
-  it('truncates oversized tool outputs', () => {
-    const out = stripMedia([bigToolResult()], { toolOutputMaxChars: 100 });
-    const r = out[0];
-    if (r?.role !== 'function_result') throw new Error('role drift');
-    const block = r.content[0] as { type: 'text'; text: string };
-    expect(block.text.length).toBeLessThanOrEqual(100 + 32);
-    expect(block.text.endsWith('[truncated]')).toBe(true);
-  });
-  it('leaves short tool outputs untouched', () => {
-    const m: AgentMessage = {
-      role: 'function_result',
-      function_call_id: 'fc',
-      function_id: 'x',
-      content: [{ type: 'text', text: 'short' }],
-      details: {},
-      is_error: false,
-      timestamp: 0,
-    };
-    const out = stripMedia([m], { toolOutputMaxChars: 2000 });
-    expect(out[0]?.content[0]).toEqual({ type: 'text', text: 'short' });
-  });
-  it('strips images inside function_result content', () => {
-    const m: AgentMessage = {
-      role: 'function_result',
-      function_call_id: 'fc',
-      function_id: 'screenshot',
-      content: [
-        { type: 'text', text: 'capture' },
-        { type: 'image' } as unknown as { type: 'text'; text: string },
-      ],
-      details: {},
-      is_error: false,
-      timestamp: 0,
-    };
-    const out = stripMedia([m], { toolOutputMaxChars: 2000 });
-    expect(out[0]?.content).toEqual([
-      { type: 'text', text: 'capture' },
-      { type: 'text', text: '[image stripped]' },
-    ]);
-  });
-});
diff --git a/harness/tests/context-compaction/summarize.test.ts b/harness/tests/context-compaction/summarize.test.ts
deleted file mode 100644
index cf3c8cb9..00000000
--- a/harness/tests/context-compaction/summarize.test.ts
+++ /dev/null
@@ -1,447 +0,0 @@
-import { describe, expect, it, vi } from 'vitest';
-import {
-  estimateTokenCount,
-  renderUserPrompt,
-  summarizeAndAppend,
-} from '../../src/context-compaction/summarize.js';
-import type { AgentMessage, AssistantMessage } from '../../src/types/agent-message.js';
-import type { ContentBlock } from '../../src/types/content.js';
-
-// ---------------------------------------------------------------------------
-// Helpers
-// ---------------------------------------------------------------------------
-
-const userMsg = (t: string): AgentMessage => ({
-  role: 'user',
-  content: [{ type: 'text', text: t }],
-  timestamp: 0,
-});
-
-const assistantMsg = (t: string): AgentMessage => ({
-  role: 'assistant',
-  content: [{ type: 'text', text: t }],
-  stop_reason: 'end',
-  model: 'm',
-  provider: 'p',
-  timestamp: 0,
-});
-
-const withId = (entry_id: string, msg: AgentMessage) => ({ entry_id, message: msg });
-
-// Builds entries large enough that selectWithEntryIds finds a non-empty head
-// (needs > tailTurns=2 user turns to split off some turns into head).
-const makeEntries = () => [
-  withId('id-0', userMsg('user 0')),
-  withId('id-1', assistantMsg('assistant 1')),
-  withId('id-2', userMsg('user 2')),
-  withId('id-3', assistantMsg('assistant 3')),
-  withId('id-4', userMsg('user 4')),
-  withId('id-5', assistantMsg('assistant 5')),
-  withId('id-6', userMsg('user 6')),
-  withId('id-7', assistantMsg('assistant 7')),
-];
-
-// Model stub passed to summarizeAndAppend
-const testModel = {
-  providerID: 'anthropic',
-  modelID: 'claude-haiku-4-5',
-  modelLimit: { context: 200_000, input: 200_000, output: 4_096 },
-};
-
-// ---------------------------------------------------------------------------
-// ISdk mock factory
-//
-// The summariser now runs through one `router::complete` call — no channel.
-// `onComplete` receives the payload and returns the final AssistantMessage
-// (or throws to fail the call, feeding the retry logic).
-// ---------------------------------------------------------------------------
-
-function doneMessage(summary = 'the summary'): AssistantMessage {
-  return {
-    role: 'assistant',
-    content: [{ type: 'text', text: summary }],
-    stop_reason: 'end',
-    model: 'm',
-    provider: 'p',
-    timestamp: 0,
-  };
-}
-
-type MockTriggerReq = { function_id: string; payload: Record<string, unknown> };
-
-/**
- * Build a minimal ISdk mock.
- *
- * @param onComplete - Called with the `router::complete` payload; returns the
- *   AssistantMessage wrapped as `{ message }`, or throws to fail the call.
- * @param extraHandlers - per function_id response overrides.
- */
-function buildMock(
-  onComplete: (payload: Record<string, unknown>) => AssistantMessage = () => doneMessage(),
-  extraHandlers: Record<string, (payload: Record<string, unknown>) => unknown> = {},
-) {
-  const trigger = vi.fn(async (req: MockTriggerReq) => {
-    const { function_id, payload } = req;
-
-    // Check extra handlers first
-    if (extraHandlers[function_id]) {
-      return extraHandlers[function_id](payload);
-    }
-
-    // Default responses for session-manager ops
-    if (function_id === 'session::messages') {
-      return {
-        messages: makeEntries().map((e) => ({ entry_id: e.entry_id, message: e.message })),
-      };
-    }
-    if (function_id === 'session::append') {
-      return { entry_id: 'appended-1', parent_id: null, timestamp: 0 };
-    }
-    // state::set / state::get used by stampLastCompaction / acquireLease
-    if (function_id === 'state::set' || function_id === 'state::get') {
-      return undefined;
-    }
-
-    if (function_id === 'router::complete') {
-      return { message: onComplete(payload), provider: payload.provider, model: payload.model };
-    }
-
-    return undefined;
-  });
-
-  const iii = { trigger } as unknown as import('iii-sdk').ISdk;
-
-  return { iii, trigger };
-}
-
-// ---------------------------------------------------------------------------
-// renderUserPrompt
-// ---------------------------------------------------------------------------
-
-describe('renderUserPrompt', () => {
-  it('includes role markers and text content', () => {
-    const msgs: AgentMessage[] = [userMsg('hello'), assistantMsg('hi')];
-    const out = renderUserPrompt(msgs);
-    expect(out).toContain('[user]');
-    expect(out).toContain('[assistant]');
-    expect(out).toContain('hello');
-    expect(out).toContain('hi');
-    expect(out).toContain('<conversation>');
-    expect(out).toContain('</conversation>');
-  });
-
-  it('renders function_call blocks with verbatim ids', () => {
-    const msg: AgentMessage = {
-      role: 'assistant',
-      content: [
-        {
-          type: 'function_call',
-          id: 'tc1',
-          function_id: 'shell::run',
-          arguments: { command: 'ls' },
-        },
-      ],
-      stop_reason: 'function_call',
-      model: 'm',
-      provider: 'p',
-      timestamp: 0,
-    };
-    const out = renderUserPrompt([msg]);
-    expect(out).toContain('[tool_call] shell::run');
-    expect(out).toContain('"command"');
-  });
-
-  it('skips non-text non-function_call blocks silently', () => {
-    const msg: AgentMessage = {
-      role: 'user',
-      content: [{ type: 'image', media_type: 'image/png', data: 'abc' } as unknown as ContentBlock],
-      timestamp: 0,
-    };
-    const out = renderUserPrompt([msg]);
-    expect(out).toContain('[user]');
-    expect(out).not.toContain('image/png');
-  });
-});
-
-// ---------------------------------------------------------------------------
-// estimateTokenCount
-// ---------------------------------------------------------------------------
-
-describe('estimateTokenCount', () => {
-  it('returns positive for non-empty messages', () => {
-    expect(estimateTokenCount([userMsg('hello')])).toBeGreaterThan(0);
-  });
-
-  it('scales with message size', () => {
-    expect(estimateTokenCount([userMsg('x'.repeat(4000))])).toBeGreaterThan(
-      estimateTokenCount([userMsg('x')]),
-    );
-  });
-
-  it('returns 0 for empty array', () => {
-    expect(estimateTokenCount([])).toBe(0);
-  });
-});
-
-// ---------------------------------------------------------------------------
-// summarizeAndAppend — async mode
-// ---------------------------------------------------------------------------
-
-describe('summarizeAndAppend async mode', () => {
-  it('loads messages, calls summariser, appends compaction, returns ok', async () => {
-    const compactPayloads: unknown[] = [];
-
-    const { iii, trigger } = buildMock(() => doneMessage('my summary'), {
-      'session::messages': () => ({
-        messages: makeEntries().map((e) => ({ entry_id: e.entry_id, message: e.message })),
-      }),
-      'session::append': (p) => {
-        compactPayloads.push(p);
-        return { entry_id: 'comp-new', parent_id: null, timestamp: 0 };
-      },
-    });
-
-    const result = await summarizeAndAppend(iii, 'sess-1', { mode: 'async' }, testModel);
-
-    // Returns ok shape
-    expect(result).not.toBe('compact');
-    expect(result).not.toBe('empty');
-    const ok = result as { tail_start_id: string | null; tokens_before: number };
-    expect(ok.tokens_before).toBeGreaterThan(0);
-
-    // The summariser ran through router::complete
-    const completeCall = trigger.mock.calls.find(([req]) => req.function_id === 'router::complete');
-    expect(completeCall).toBeDefined();
-
-    // Compaction custom entry was appended with session_id and non-empty summary
-    expect(compactPayloads).toHaveLength(1);
-    const cp = compactPayloads[0] as {
-      session_id: string;
-      custom: { custom_type: string; data: { summary: string } };
-    };
-    expect(cp.session_id).toBe('sess-1');
-    expect(cp.custom.custom_type).toBe('compaction');
-    expect(typeof cp.custom.data.summary).toBe('string');
-    expect(cp.custom.data.summary.length).toBeGreaterThan(0);
-  });
-
-  it('returns "empty" when summariser produces no text content', async () => {
-    const { iii } = buildMock(() => ({ ...doneMessage(), content: [] }));
-
-    const result = await summarizeAndAppend(iii, 'sess-empty', { mode: 'async' }, testModel);
-    expect(result).toBe('empty');
-  });
-
-  it('returns "compact" when router::complete rejects', async () => {
-    const { iii } = buildMock(() => {
-      throw new Error('router unreachable');
-    });
-
-    const result = await summarizeAndAppend(iii, 'sess-fail', { mode: 'async' }, testModel);
-    expect(typeof result === 'object' && 'kind' in result ? result.kind : result).toBe('compact');
-    if (typeof result === 'object' && 'kind' in result && result.kind === 'compact') {
-      expect(result.reason).toContain('router unreachable');
-    }
-  });
-
-  it('retries router::complete once and succeeds when the first attempt fails', async () => {
-    // Transient provider failures (429, network blip, 5xx) should not
-    // surface to /compact when a single retry would succeed.
-    let attempts = 0;
-    const { iii } = buildMock(() => {
-      attempts += 1;
-      if (attempts === 1) {
-        throw new Error('transient provider failure');
-      }
-      return doneMessage('recovered summary');
-    });
-
-    const result = await summarizeAndAppend(iii, 'sess-retry', { mode: 'async' }, testModel);
-    expect(attempts).toBe(2);
-    expect(typeof result === 'object' && 'kind' in result ? result.kind : result).toBe('ok');
-  }, 10_000);
-
-  it('returns "compact" when both stream attempts fail', async () => {
-    // Generic permanent failure (no auth/4xx markers) — still retries once
-    // then surfaces the failure.
-    let attempts = 0;
-    const { iii } = buildMock(() => {
-      attempts += 1;
-      throw new Error('permanent provider failure');
-    });
-
-    const result = await summarizeAndAppend(iii, 'sess-fail-twice', { mode: 'async' }, testModel);
-    expect(attempts).toBe(2);
-    expect(typeof result === 'object' && 'kind' in result ? result.kind : result).toBe('compact');
-    if (typeof result === 'object' && 'kind' in result && result.kind === 'compact') {
-      expect(result.reason).toContain('permanent provider failure');
-    }
-  }, 10_000);
-
-  it('skips retry on non-retryable errors (401/auth/malformed)', async () => {
-    // Auth and 4xx errors won't fix on retry. Skip the retry so the
-    // compaction lease releases ~1s sooner.
-    let attempts = 0;
-    const { iii } = buildMock(() => {
-      attempts += 1;
-      throw new Error('401 unauthorized: invalid_api_key');
-    });
-
-    const result = await summarizeAndAppend(iii, 'sess-auth-fail', { mode: 'async' }, testModel);
-    expect(attempts).toBe(1); // no retry
-    expect(typeof result === 'object' && 'kind' in result ? result.kind : result).toBe('compact');
-    if (typeof result === 'object' && 'kind' in result && result.kind === 'compact') {
-      expect(result.reason).toContain('401');
-    }
-  });
-
-  it('returns "compact" when the provider stream emits an error terminal event', async () => {
-    // Regression: a not_found_error from Anthropic (e.g. wrong provider/model
-    // combo) used to be silently stored as the compaction summary text,
-    // showing "COMPACTED · N TOKENS" in the UI with the error JSON as
-    // the summary. The error terminal must surface as a failure instead.
-    const { iii } = buildMock(() => ({
-      role: 'assistant',
-      content: [
-        {
-          type: 'text',
-          text: '{"type":"error","error":{"type":"not_found_error","message":"model: gpt-5-mini"}}',
-        },
-      ],
-      stop_reason: 'error',
-      error_message: 'model: gpt-5-mini',
-      model: 'gpt-5-mini',
-      provider: 'anthropic',
-      timestamp: 0,
-    }));
-
-    const result = await summarizeAndAppend(
-      iii,
-      'sess-provider-error',
-      { mode: 'async' },
-      testModel,
-    );
-    expect(typeof result === 'object' && 'kind' in result ? result.kind : result).toBe('compact');
-    if (typeof result === 'object' && 'kind' in result && result.kind === 'compact') {
-      expect(result.reason.toLowerCase()).toMatch(/not.found|gpt-5-mini/);
-    }
-  });
-
-  it('pins the session provider/model on the router::complete call', async () => {
-    // /compact uses the session's own provider/model, pinned explicitly so
-    // the router executes on exactly the provider the session streams on.
-    const { iii, trigger } = buildMock();
-
-    const openAiModel = {
-      providerID: 'openai',
-      modelID: 'gpt-5-mini',
-      modelLimit: { context: 200_000, input: 200_000, output: 4_096 },
-    };
-    await summarizeAndAppend(iii, 'sess-openai', { mode: 'async' }, openAiModel);
-    const call = trigger.mock.calls.find(([req]) => req.function_id === 'router::complete');
-    expect(call?.[0].payload).toEqual(
-      expect.objectContaining({ provider: 'openai', model: 'gpt-5-mini' }),
-    );
-  });
-
-  it('pins kimi sessions to the kimi provider, not anthropic', async () => {
-    const { iii, trigger } = buildMock();
-
-    const kimiModel = {
-      providerID: 'kimi',
-      modelID: 'kimi-k2.5',
-      modelLimit: { context: 256_000, input: 256_000, output: 16_384 },
-    };
-    await summarizeAndAppend(iii, 'sess-kimi', { mode: 'async' }, kimiModel);
-    const call = trigger.mock.calls.find(([req]) => req.function_id === 'router::complete');
-    expect(call?.[0].payload).toEqual(
-      expect.objectContaining({ provider: 'kimi', model: 'kimi-k2.5' }),
-    );
-  });
-});
-
-// ---------------------------------------------------------------------------
-// summarizeAndAppend — anchored prompt (prior compaction present)
-// ---------------------------------------------------------------------------
-
-describe('summarizeAndAppend anchored prompt', () => {
-  it('includes <previous-summary> in system prompt when a prior compaction exists', async () => {
-    const PRIOR_SUMMARY = 'Prior compaction summary text here.';
-    const capturedPrompts: string[] = [];
-
-    const { iii } = buildMock(
-      (payload) => {
-        capturedPrompts.push(payload.system_prompt as string);
-        return doneMessage('updated summary');
-      },
-      {
-        'session::messages': (p) => ({
-          messages: [
-            ...(p.include_custom
-              ? [
-                  {
-                    entry_id: 'comp-1',
-                    custom: {
-                      custom_type: 'compaction',
-                      data: {
-                        summary: PRIOR_SUMMARY,
-                        tokens_before: 1000,
-                        tail_start_id: null,
-                        details: {},
-                        timestamp: Date.now() - 60_000,
-                      },
-                    },
-                  },
-                ]
-              : []),
-            ...makeEntries().map((e) => ({ entry_id: e.entry_id, message: e.message })),
-          ],
-        }),
-      },
-    );
-
-    await summarizeAndAppend(iii, 'sess-anchored', { mode: 'async' }, testModel);
-
-    expect(capturedPrompts).toHaveLength(1);
-    expect(capturedPrompts[0]).toContain('<previous-summary>');
-    expect(capturedPrompts[0]).toContain(PRIOR_SUMMARY);
-    expect(capturedPrompts[0]).toContain('</previous-summary>');
-  });
-});
-
-// ---------------------------------------------------------------------------
-// summarizeAndAppend — sync mode (truncatedEntries supplied by caller)
-// ---------------------------------------------------------------------------
-
-describe('summarizeAndAppend sync mode', () => {
-  it('uses caller-supplied truncatedEntries and skips the transcript reload', async () => {
-    const messageReads: Array<Record<string, unknown>> = [];
-    const compactPayloads: unknown[] = [];
-
-    const { iii } = buildMock(() => doneMessage('sync summary'), {
-      'session::messages': (p) => {
-        messageReads.push(p);
-        return { messages: [] };
-      },
-      'session::append': (p) => {
-        compactPayloads.push(p);
-        return { entry_id: 'comp-new', parent_id: null, timestamp: 0 };
-      },
-    });
-
-    const entries = makeEntries();
-    const result = await summarizeAndAppend(
-      iii,
-      'sess-sync',
-      { mode: 'sync', truncatedEntries: entries },
-      testModel,
-    );
-
-    // Only the prior-compaction lookup (include_custom) reads the path; the
-    // message window itself comes from the caller-supplied truncated set.
-    expect(messageReads.filter((p) => p.include_custom !== true)).toHaveLength(0);
-    expect(result).not.toBe('compact');
-    expect(result).not.toBe('empty');
-    expect(compactPayloads).toHaveLength(1);
-  });
-});
diff --git a/harness/tests/context-compaction/template.test.ts b/harness/tests/context-compaction/template.test.ts
deleted file mode 100644
index 2c668ecf..00000000
--- a/harness/tests/context-compaction/template.test.ts
+++ /dev/null
@@ -1,42 +0,0 @@
-import { describe, expect, it } from 'vitest';
-import { SUMMARY_TEMPLATE, buildPrompt } from '../../src/context-compaction/template.js';
-
-describe('SUMMARY_TEMPLATE', () => {
-  it('contains the harness-specific Tool Calls Made section', () => {
-    expect(SUMMARY_TEMPLATE).toContain('## Tool Calls Made');
-  });
-  it('contains all required sections', () => {
-    for (const h of [
-      '## Goal',
-      '## Constraints & Preferences',
-      '## Progress',
-      '### Done',
-      '### In Progress',
-      '### Blocked',
-      '## Key Decisions',
-      '## Next Steps',
-      '## Critical Context',
-      '## Relevant Files',
-    ]) {
-      expect(SUMMARY_TEMPLATE).toContain(h);
-    }
-  });
-});
-
-describe('buildPrompt', () => {
-  it('omits previous-summary block on first compaction', () => {
-    const p = buildPrompt({ previousSummary: undefined, context: [] });
-    expect(p).not.toContain('<previous-summary>');
-    expect(p).toContain('Create a new anchored summary');
-  });
-  it('injects previous-summary block when present', () => {
-    const p = buildPrompt({ previousSummary: 'prior text', context: [] });
-    expect(p).toContain('<previous-summary>');
-    expect(p).toContain('prior text');
-    expect(p).toContain('Update the anchored summary');
-  });
-  it('appends extra context blocks', () => {
-    const p = buildPrompt({ previousSummary: undefined, context: ['extra-note'] });
-    expect(p).toContain('extra-note');
-  });
-});
diff --git a/harness/tests/runtime/compaction.test.ts b/harness/tests/runtime/compaction.test.ts
new file mode 100644
index 00000000..de6794b6
--- /dev/null
+++ b/harness/tests/runtime/compaction.test.ts
@@ -0,0 +1,249 @@
+/**
+ * Unit tests for the compaction round-trip helpers (src/runtime/compaction.ts):
+ * window loading + summary anchoring, the best-effort assemble client with its
+ * raw fallback, and additive round-trip persistence.
+ */
+
+import { describe, expect, it, vi } from 'vitest';
+import {
+  assembleContext,
+  loadAssembleWindow,
+  persistCompactionRoundTrip,
+} from '../../src/runtime/compaction.js';
+import type { ISdk } from '../../src/runtime/iii.js';
+import type { AgentMessage } from '../../src/types/agent-message.js';
+import type { MessageWithEntryId } from '../../src/runtime/session.js';
+import { FakeContextManager, FakeSessionManager } from '../_helpers/fakeSessionManager.js';
+
+function buildIii(): { iii: ISdk; sessions: FakeSessionManager; context: FakeContextManager } {
+  const sessions = new FakeSessionManager();
+  const context = new FakeContextManager();
+  const iii = {
+    trigger: vi.fn(async (req: { function_id: string; payload?: unknown }) => {
+      const s = await sessions.handle(req.function_id, req.payload);
+      if (s !== undefined) return s;
+      const c = await context.handle(req.function_id, req.payload);
+      if (c !== undefined) return c;
+      return null;
+    }),
+    registerFunction: vi.fn(),
+    registerTrigger: vi.fn(),
+  } as unknown as ISdk;
+  return { iii, sessions, context };
+}
+
+const user = (text: string): AgentMessage => ({
+  role: 'user',
+  content: [{ type: 'text', text }],
+  timestamp: 1,
+});
+const asst = (text: string): AgentMessage => ({
+  role: 'assistant',
+  content: text ? [{ type: 'text', text }] : [],
+  stop_reason: 'end',
+  model: 'm',
+  provider: 'p',
+  timestamp: 2,
+});
+const win = (entry_id: string, message: AgentMessage): MessageWithEntryId => ({
+  entry_id,
+  message,
+});
+
+describe('loadAssembleWindow', () => {
+  it('returns the full window and no summary when there is no compaction', async () => {
+    const { iii, sessions } = buildIii();
+    sessions.seedMessages('s1', [user('a'), asst('b')], ['u1', 'a1']);
+
+    const { window, previousSummary } = await loadAssembleWindow(iii, 's1');
+
+    expect(window.map((w) => w.entry_id)).toEqual(['u1', 'a1']);
+    expect(previousSummary).toBeUndefined();
+  });
+
+  it('slices from tail_start_entry_id and surfaces the anchored summary', async () => {
+    const { iii, sessions } = buildIii();
+    sessions.seedMessages(
+      's1',
+      [user('a'), asst('b'), user('c'), asst('d')],
+      ['u1', 'a1', 'u2', 'a2'],
+    );
+    sessions.seedCompaction('s1', { summary: 'PRIOR', tail_start_entry_id: 'u2' });
+
+    const { window, previousSummary } = await loadAssembleWindow(iii, 's1');
+
+    expect(window.map((w) => w.entry_id)).toEqual(['u2', 'a2']);
+    expect(previousSummary).toBe('PRIOR');
+  });
+
+  it('reads the legacy tail_start_id key for pre-migration sessions', async () => {
+    const { iii, sessions } = buildIii();
+    sessions.seedMessages('s1', [user('a'), asst('b'), user('c')], ['u1', 'a1', 'u2']);
+    sessions.seedCompaction('s1', {
+      summary: 'PRIOR',
+      tail_start_entry_id: 'u2',
+      legacyTailField: true,
+    });
+
+    const { window, previousSummary } = await loadAssembleWindow(iii, 's1');
+
+    expect(window.map((w) => w.entry_id)).toEqual(['u2']);
+    expect(previousSummary).toBe('PRIOR');
+  });
+
+  it('keeps only messages after the compaction entry when the tail is null', async () => {
+    const { iii, sessions } = buildIii();
+    sessions.seedMessages('s1', [user('a'), asst('b')], ['u1', 'a1']);
+    sessions.seedCompaction('s1', { summary: 'PRIOR', tail_start_entry_id: null });
+    sessions.seedMessages('s1', [user('c')], ['u2']);
+
+    const { window } = await loadAssembleWindow(iii, 's1');
+
+    expect(window.map((w) => w.entry_id)).toEqual(['u2']);
+  });
+
+  it('drops excluded entries and empty assistant placeholders', async () => {
+    const { iii, sessions } = buildIii();
+    sessions.seedMessages('s1', [user('a'), asst(''), asst('b')], ['u1', 'placeholder', 'a1']);
+
+    const { window } = await loadAssembleWindow(iii, 's1', { excludeEntryIds: ['a1'] });
+
+    // 'a1' excluded, the empty-content assistant ('placeholder') filtered out.
+    expect(window.map((w) => w.entry_id)).toEqual(['u1']);
+  });
+});
+
+describe('assembleContext', () => {
+  it('sends inline limits + lease_key + previous_summary and returns the worker response', async () => {
+    const { iii, context } = buildIii();
+    context.assemble = () => ({
+      system_prompt: 'BASE\n\n# Conversation summary\nX',
+      messages: [user('hi')],
+      applied: {
+        pruned: false,
+        pruned_tokens: 0,
+        compacted: true,
+        summary: 'X',
+        tail_start_index: 0,
+        tokens_before: 9,
+      },
+    });
+
+    const res = await assembleContext(iii, {
+      session_id: 's1',
+      window: [win('u1', user('hi'))],
+      baseSystemPrompt: 'BASE',
+      modelId: 'm',
+      provider: 'p',
+      modelMeta: { id: 'm', provider: 'p', context_window: 1000, max_output_tokens: 100 },
+      thinkingLevel: 'high',
+      previousSummary: 'PRIOR',
+    });
+
+    expect(res.system_prompt).toContain('# Conversation summary');
+    expect(res.applied.compacted).toBe(true);
+
+    const req = context.callsTo('context::assemble')[0];
+    expect(req?.options).toMatchObject({
+      lease_key: 's1',
+      previous_summary: 'PRIOR',
+      thinking_level: 'high',
+    });
+    expect(req?.model).toMatchObject({
+      id: 'm',
+      provider: 'p',
+      limits: { context_window: 1000, max_output_tokens: 100 },
+    });
+  });
+
+  it('falls back to the raw window when context::assemble fails', async () => {
+    const { iii, context } = buildIii();
+    context.assemble = () => {
+      throw new Error('worker unavailable');
+    };
+    const window = [win('u1', user('hi'))];
+
+    const res = await assembleContext(iii, {
+      session_id: 's1',
+      window,
+      baseSystemPrompt: 'BASE',
+      modelId: 'm',
+    });
+
+    expect(res.system_prompt).toBe('BASE');
+    expect(res.messages).toEqual([window[0]?.message]);
+    expect(res.applied).toEqual({ pruned: false, pruned_tokens: 0, compacted: false });
+  });
+});
+
+describe('persistCompactionRoundTrip', () => {
+  it('writes an additive compaction entry mapping tail_start_index to an entry id', async () => {
+    const { iii, sessions } = buildIii();
+    const window = [win('u1', user('a')), win('a1', asst('b')), win('u2', user('c'))];
+
+    const entryId = await persistCompactionRoundTrip(
+      iii,
+      's1',
+      {
+        pruned: false,
+        pruned_tokens: 0,
+        compacted: true,
+        summary: 'S',
+        tail_start_index: 2,
+        tokens_before: 50,
+      },
+      window,
+    );
+
+    expect(typeof entryId).toBe('string');
+    const data = (
+      sessions
+        .callsTo('session::append')
+        .find((p) => (p.custom as { custom_type?: string })?.custom_type === 'compaction')
+        ?.custom as { data: Record<string, unknown> }
+    ).data;
+    expect(data.tail_start_entry_id).toBe('u2');
+    expect(data.summary).toBe('S');
+    expect(data.tokens_before).toBe(50);
+  });
+
+  it('writes a null tail boundary when everything was summarised', async () => {
+    const { iii, sessions } = buildIii();
+    const window = [win('u1', user('a'))];
+
+    await persistCompactionRoundTrip(
+      iii,
+      's1',
+      {
+        pruned: false,
+        pruned_tokens: 0,
+        compacted: true,
+        summary: 'S',
+        tail_start_index: null,
+        tokens_before: 1,
+      },
+      window,
+    );
+
+    const data = (
+      sessions
+        .callsTo('session::append')
+        .find((p) => (p.custom as { custom_type?: string })?.custom_type === 'compaction')
+        ?.custom as { data: Record<string, unknown> }
+    ).data;
+    expect(data.tail_start_entry_id).toBeNull();
+  });
+
+  it('is a no-op (returns null, writes nothing) when not compacted', async () => {
+    const { iii, sessions } = buildIii();
+    const entryId = await persistCompactionRoundTrip(
+      iii,
+      's1',
+      { pruned: false, pruned_tokens: 0, compacted: false },
+      [win('u1', user('a'))],
+    );
+
+    expect(entryId).toBeNull();
+    expect(sessions.callsTo('session::append')).toHaveLength(0);
+  });
+});
diff --git a/harness/tests/turn-orchestrator/_helpers/mockTurnStore.ts b/harness/tests/turn-orchestrator/_helpers/mockTurnStore.ts
index 9058bcdb..129385d6 100644
--- a/harness/tests/turn-orchestrator/_helpers/mockTurnStore.ts
+++ b/harness/tests/turn-orchestrator/_helpers/mockTurnStore.ts
@@ -24,7 +24,6 @@ export function mockTurnStore(overrides: Partial<TurnStore> = {}): MockTurnStore
     saveRecord: vi.fn(async () => {}),
     writeRecord: vi.fn(async () => {}),
     ensureSession: vi.fn(async () => {}),
-    loadMessages: vi.fn(async () => []),
     appendMessages: vi.fn(async () => {}),
     loadRunRequest: vi.fn(async () => defaultRunRequest),
     saveRunRequest: vi.fn(async () => {}),
diff --git a/harness/tests/turn-orchestrator/assistant-streaming.test.ts b/harness/tests/turn-orchestrator/assistant-streaming.test.ts
index 3b0b6022..72e86a3c 100644
--- a/harness/tests/turn-orchestrator/assistant-streaming.test.ts
+++ b/harness/tests/turn-orchestrator/assistant-streaming.test.ts
@@ -32,7 +32,6 @@ function stubStreamingPorts(
   overrides: Partial<AssistantStreamingPorts> = {},
 ): AssistantStreamingPorts {
   return {
-    loadMessages: vi.fn(async () => []),
     appendMessages: vi.fn(async () => {}),
     checkpoint: vi.fn(async () => {}),
     emitTurnEnd: vi.fn(async () => {}),
@@ -46,7 +45,13 @@ function stubStreamingPorts(
       system_prompt: 'sys',
       function_schemas: [{ name: 'agent_trigger', description: 'd', parameters: {} }],
     })),
-    runPreflight: vi.fn(async () => 'ok' as const),
+    loadAssembleWindow: vi.fn(async () => ({ window: [] })),
+    assembleContext: vi.fn(async (params) => ({
+      system_prompt: params.baseSystemPrompt,
+      messages: params.window.map((w) => w.message),
+      applied: { pruned: false, pruned_tokens: 0, compacted: false },
+    })),
+    persistCompaction: vi.fn(async () => {}),
     streamTurn: vi.fn(async () => ({ final: null, error: null })),
     emitMessageComplete: vi.fn(async () => {}),
     appendAssistantPlaceholder: vi.fn(async () => {}),
@@ -82,24 +87,69 @@ describe('turnAssistantEntryId', () => {
 });
 
 describe('prepareStreamContext', () => {
-  it("reloads messages when preflight compacts, excluding this turn's entry", async () => {
-    const loadMessages = vi
-      .fn()
-      .mockResolvedValueOnce([{ role: 'user', content: [], timestamp: 1 }])
-      .mockResolvedValueOnce([{ role: 'user', content: [], timestamp: 2 }]);
-    const ports = stubStreamingPorts({
-      loadMessages,
-      runPreflight: vi.fn(async () => 'compacted'),
-    });
+  it("assembles the window into system_prompt + messages, excluding this turn's entry", async () => {
+    const window = [
+      { entry_id: 'u1', message: { role: 'user' as const, content: [], timestamp: 1 } },
+    ];
+    const loadAssembleWindow = vi.fn(async () => ({ window, previousSummary: 'PRIOR' }));
+    const assembleContext = vi.fn(async () => ({
+      system_prompt: 'sys\n\n# Conversation summary\nPRIOR',
+      messages: [{ role: 'user' as const, content: [], timestamp: 1 }],
+      applied: { pruned: false, pruned_tokens: 0, compacted: false },
+    }));
+    const persistCompaction = vi.fn(async () => {});
+    const ports = stubStreamingPorts({ loadAssembleWindow, assembleContext, persistCompaction });
     const rec = newRecord('s1');
     rec.state = 'assistant_streaming';
 
     const ctx = await prepareStreamContext(ports, rec, 'asst-entry-1');
 
-    expect(loadMessages).toHaveBeenCalledTimes(2);
-    expect(loadMessages).toHaveBeenCalledWith('s1', { excludeEntryIds: ['asst-entry-1'] });
-    expect(ctx.messages).toEqual([{ role: 'user', content: [], timestamp: 2 }]);
+    expect(loadAssembleWindow).toHaveBeenCalledWith('s1', { excludeEntryIds: ['asst-entry-1'] });
+    // The anchored summary is passed back into the assemble request.
+    expect(assembleContext.mock.calls[0]?.[0]?.previousSummary).toBe('PRIOR');
+    // The returned (merged) system prompt + budgeted messages drive router::chat.
+    expect(ctx.system_prompt).toContain('# Conversation summary');
+    expect(ctx.messages).toEqual([{ role: 'user', content: [], timestamp: 1 }]);
     expect(ctx.tools[0]?.name).toBe('agent_trigger');
+    // No fresh compaction this turn → nothing persisted.
+    expect(persistCompaction).not.toHaveBeenCalled();
+  });
+
+  it('persists the compaction round trip when assemble compacts', async () => {
+    const window = [
+      { entry_id: 'u1', message: { role: 'user' as const, content: [], timestamp: 1 } },
+      {
+        entry_id: 'a1',
+        message: {
+          role: 'assistant' as const,
+          content: [],
+          stop_reason: 'end' as const,
+          model: 'm',
+          provider: 'p',
+          timestamp: 2,
+        },
+      },
+    ];
+    const applied = {
+      pruned: false,
+      pruned_tokens: 0,
+      compacted: true,
+      summary: 'S',
+      tail_start_index: 1,
+      tokens_before: 100,
+    };
+    const persistCompaction = vi.fn(async () => {});
+    const ports = stubStreamingPorts({
+      loadAssembleWindow: vi.fn(async () => ({ window })),
+      assembleContext: vi.fn(async () => ({ system_prompt: 'sys', messages: [], applied })),
+      persistCompaction,
+    });
+    const rec = newRecord('s1');
+    rec.state = 'assistant_streaming';
+
+    await prepareStreamContext(ports, rec, 'asst-entry-1');
+
+    expect(persistCompaction).toHaveBeenCalledWith('s1', applied, window);
   });
 });
 
diff --git a/harness/tests/turn-orchestrator/assistant.test.ts b/harness/tests/turn-orchestrator/assistant.test.ts
index 612dbb7b..43da6c44 100644
--- a/harness/tests/turn-orchestrator/assistant.test.ts
+++ b/harness/tests/turn-orchestrator/assistant.test.ts
@@ -3,7 +3,6 @@ import type { ISdk } from '../../src/runtime/iii.js';
 import type { AssistantMessage } from '../../src/types/agent-message.js';
 import { installMockTurnStore } from './_helpers/mockTurnStore.js';
 import { FakeSessionManager } from '../_helpers/fakeSessionManager.js';
-import * as preflightModule from '../../src/turn-orchestrator/preflight.js';
 import { type TurnStateRecord, newRecord } from '../../src/turn-orchestrator/state.js';
 import { handleFinishing } from '../../src/turn-orchestrator/finishing.js';
 import { handleStreaming } from '../../src/turn-orchestrator/assistant-streaming/process.js';
@@ -91,7 +90,6 @@ function mockStreamingStore(overrides: Parameters<typeof installMockTurnStore>[0
       system_prompt: '',
       function_schemas: [],
     })),
-    loadMessages: vi.fn(async () => []),
     ...overrides,
   });
 }
@@ -105,7 +103,6 @@ describe('handleStreaming turn start', () => {
       },
     });
     mockStreamingStore();
-    vi.spyOn(preflightModule, 'runPreflight').mockResolvedValue('ok');
 
     await handleStreaming(iii, rec);
 
@@ -125,7 +122,6 @@ describe('handleStreaming', () => {
       },
     });
     mockStreamingStore();
-    vi.spyOn(preflightModule, 'runPreflight').mockResolvedValue('ok');
 
     await handleStreaming(iii, rec);
 
@@ -149,7 +145,6 @@ describe('handleStreaming', () => {
     const { iii, calls, sessions } = fakeIiiWithDone(finalMsg);
 
     mockStreamingStore();
-    vi.spyOn(preflightModule, 'runPreflight').mockResolvedValue('ok');
 
     await handleStreaming(iii, rec);
 
@@ -173,7 +168,6 @@ describe('handleStreaming', () => {
     const { iii } = fakeIiiWithDone(finalMsg);
 
     mockStreamingStore();
-    vi.spyOn(preflightModule, 'runPreflight').mockResolvedValue('ok');
 
     await handleStreaming(iii, rec);
 
@@ -209,7 +203,6 @@ describe('handleStreaming', () => {
     });
 
     mockStreamingStore();
-    vi.spyOn(preflightModule, 'runPreflight').mockResolvedValue('ok');
 
     await handleStreaming(iii, rec);
 
@@ -223,7 +216,6 @@ describe('handleStreaming', () => {
     const { iii, sessions } = fakeIiiWithDone(finalMsg);
 
     mockStreamingStore();
-    vi.spyOn(preflightModule, 'runPreflight').mockResolvedValue('ok');
 
     await handleStreaming(iii, rec);
 
@@ -253,7 +245,6 @@ describe('handleStreaming', () => {
     const { iii, sessions } = fakeIiiWithDone(finalMsg);
 
     mockStreamingStore();
-    vi.spyOn(preflightModule, 'runPreflight').mockResolvedValue('ok');
 
     await handleStreaming(iii, rec);
     expect(sessions.messageEntries('s1')).toHaveLength(1);
diff --git a/harness/tests/turn-orchestrator/context-view.test.ts b/harness/tests/turn-orchestrator/context-view.test.ts
deleted file mode 100644
index 7175f95d..00000000
--- a/harness/tests/turn-orchestrator/context-view.test.ts
+++ /dev/null
@@ -1,86 +0,0 @@
-import { describe, expect, it } from 'vitest';
-import { buildSummaryMessage } from '../../src/context-compaction/flat-state.js';
-import { buildContextView } from '../../src/turn-orchestrator/state-runtime/context-view.js';
-import type { MessageWithEntryId } from '../../src/runtime/session.js';
-import type { AgentMessage } from '../../src/types/agent-message.js';
-
-function user(text: string): AgentMessage {
-  return { role: 'user', content: [{ type: 'text', text }], timestamp: 0 };
-}
-
-function asst(text: string): AgentMessage {
-  return {
-    role: 'assistant',
-    content: [{ type: 'text', text }],
-    stop_reason: 'end',
-    error_message: null,
-    error_kind: null,
-    usage: null,
-    model: 'm',
-    provider: 'p',
-    timestamp: 0,
-  };
-}
-
-function entry(id: string, message: AgentMessage): MessageWithEntryId {
-  return { entry_id: id, message };
-}
-
-/** buildSummaryMessage stamps Date.now(); the view under test calls it again,
- *  so an exact-timestamp expectation races the millisecond boundary. */
-function summary(text: string) {
-  return { ...buildSummaryMessage(text), timestamp: expect.any(Number) };
-}
-
-describe('buildContextView', () => {
-  it('returns raw path when there is no compaction', () => {
-    const messages = [entry('a', user('one')), entry('b', asst('two'))];
-    expect(buildContextView(messages, [])).toEqual([user('one'), asst('two')]);
-  });
-
-  it('drops empty assistant placeholders from the provider window', () => {
-    const placeholder: AgentMessage = { ...asst(''), content: [] };
-    const messages = [entry('a', user('one')), entry('b', placeholder)];
-    expect(buildContextView(messages, [])).toEqual([user('one')]);
-  });
-
-  it('reconstructs summary + tail from tail_start_id', () => {
-    const messages = [
-      entry('head', user('old')),
-      entry('tail1', asst('keep')),
-      entry('last', user('in flight')),
-    ];
-    const compactions = [{ summary: 'condensed', tail_start_id: 'tail1', timestamp: 100 }];
-
-    expect(buildContextView(messages, compactions)).toEqual([
-      summary('condensed'),
-      asst('keep'),
-      user('in flight'),
-    ]);
-  });
-
-  it('uses the latest compaction when several exist', () => {
-    const messages = [
-      entry('h', user('old')),
-      entry('t1', asst('early tail')),
-      entry('t2', user('recent')),
-    ];
-    const compactions = [
-      { summary: 'first', tail_start_id: 'h', timestamp: 10 },
-      { summary: 'latest', tail_start_id: 't2', timestamp: 20 },
-    ];
-
-    expect(buildContextView(messages, compactions)).toEqual([summary('latest'), user('recent')]);
-  });
-
-  it('keeps the whole tail when tail_start_id is absent from the path', () => {
-    const messages = [entry('a', user('one')), entry('b', asst('two'))];
-    const compactions = [{ summary: 's', tail_start_id: 'gone', timestamp: 1 }];
-
-    expect(buildContextView(messages, compactions)).toEqual([
-      summary('s'),
-      user('one'),
-      asst('two'),
-    ]);
-  });
-});
diff --git a/harness/tests/turn-orchestrator/events.test.ts b/harness/tests/turn-orchestrator/events.test.ts
index 62feb8f2..b3764425 100644
--- a/harness/tests/turn-orchestrator/events.test.ts
+++ b/harness/tests/turn-orchestrator/events.test.ts
@@ -35,7 +35,7 @@ describe('emit (agent event producer)', () => {
     expect(streamSets(calls).map((c) => c.payload.stream_name)).toEqual(['agent::events']);
   });
 
-  it('enqueues a compaction wake on turn_end instead of mirroring to a stream', async () => {
+  it('writes turn_end only to agent::events with no out-of-band compaction wake', async () => {
     const { iii, calls } = buildSdk();
     const event = {
       type: 'turn_end',
@@ -53,15 +53,10 @@ describe('emit (agent event producer)', () => {
     // The event is written only to agent::events — no dedicated turn_end stream.
     expect(streamSets(calls).map((c) => c.payload.stream_name)).toEqual(['agent::events']);
 
-    // A typed compaction wake is enqueued to the out-of-band compactor.
-    const wake = calls.find((c) => c.function_id === 'context-compaction::on_turn_end');
-    expect(wake).toBeDefined();
-    expect(wake?.payload).toMatchObject({
-      session_id: SID,
-      usage: { input: 100, output: 5 },
-      provider: 'anthropic',
-      model: 'claude-haiku-4-5',
-    });
+    // Async post-turn compaction is retired: the harness compacts inline on the
+    // hot path (context::assemble), so turn_end no longer enqueues a wake.
+    expect(calls.some((c) => c.function_id === 'context-compaction::on_turn_end')).toBe(false);
+    expect(calls.some((c) => c.function_id.startsWith('context-compaction::'))).toBe(false);
   });
 });
 
diff --git a/harness/tests/turn-orchestrator/finish.test.ts b/harness/tests/turn-orchestrator/finish.test.ts
index 371104de..b9727eae 100644
--- a/harness/tests/turn-orchestrator/finish.test.ts
+++ b/harness/tests/turn-orchestrator/finish.test.ts
@@ -6,7 +6,7 @@ import { installMockTurnStore } from './_helpers/mockTurnStore.js';
 
 describe('TurnStatePorts.finishSession', () => {
   it('emits agent_end as a signal (no transcript reload) and sets state to stopped', async () => {
-    const store = installMockTurnStore();
+    installMockTurnStore();
     const emitted: Array<{ type: string; messages?: unknown }> = [];
     const iii = {
       trigger: vi.fn(async (req: { function_id: string; payload: unknown }) => {
@@ -27,6 +27,5 @@ describe('TurnStatePorts.finishSession', () => {
     // agent_end is a turn-end signal; no consumer reads .messages, so the
     // session is no longer reloaded just to populate it.
     expect(agentEnd?.messages).toEqual([]);
-    expect(store.loadMessages).not.toHaveBeenCalled();
   });
 });
diff --git a/harness/tests/turn-orchestrator/function-awaiting-approval.test.ts b/harness/tests/turn-orchestrator/function-awaiting-approval.test.ts
index 4e7a0653..dd0db90d 100644
--- a/harness/tests/turn-orchestrator/function-awaiting-approval.test.ts
+++ b/harness/tests/turn-orchestrator/function-awaiting-approval.test.ts
@@ -115,7 +115,6 @@ describe('handleAwaitingApproval', () => {
     );
 
     installMockTurnStore({
-      loadMessages: vi.fn(async () => []),
       appendMessages: vi.fn(async () => {}),
     });
     const emitSpy = vi.spyOn(events, 'emit').mockResolvedValue(undefined);
diff --git a/harness/tests/turn-orchestrator/function-execute.test.ts b/harness/tests/turn-orchestrator/function-execute.test.ts
index 0f6a30fe..40a4fb41 100644
--- a/harness/tests/turn-orchestrator/function-execute.test.ts
+++ b/harness/tests/turn-orchestrator/function-execute.test.ts
@@ -49,7 +49,6 @@ function stubPorts(overrides: Partial<FunctionExecutePorts> = {}): FunctionExecu
       content: [{ type: 'text' as const, text: 'ok' }],
       details: {},
     })),
-    loadMessages: vi.fn(async () => []),
     appendMessages: vi.fn(async () => {}),
     emitTurnEnd: vi.fn(async () => {}),
     finishSession: vi.fn(async (rec) => {
@@ -184,8 +183,7 @@ describe('finalizeBatch', () => {
   it('appends results in one idempotent batch (no transcript reload for dedup)', async () => {
     const fc = { id: 'fc-1', function_id: 'shell::run', arguments: {} };
     const appendMessages = vi.fn(async () => {});
-    const loadMessages = vi.fn(async () => []);
-    const ports = stubPorts({ loadMessages, appendMessages });
+    const ports = stubPorts({ appendMessages });
     const rec = newRecord('s1');
     enterFunctionExecute(rec, makeAssistant([fc]));
     rec.state = 'function_execute';
@@ -203,7 +201,6 @@ describe('finalizeBatch', () => {
     };
     await finalizeBatch(ports, rec);
 
-    expect(loadMessages).not.toHaveBeenCalled();
     expect(appendMessages).toHaveBeenCalledTimes(1);
     expect(appendMessages).toHaveBeenCalledWith(
       's1',
diff --git a/harness/tests/turn-orchestrator/functions.test.ts b/harness/tests/turn-orchestrator/functions.test.ts
index 2ab8a90b..3344a3bb 100644
--- a/harness/tests/turn-orchestrator/functions.test.ts
+++ b/harness/tests/turn-orchestrator/functions.test.ts
@@ -19,7 +19,6 @@ afterEach(() => {
 
 function mockFinalizePersistence(): void {
   installMockTurnStore({
-    loadMessages: vi.fn(async () => []),
     appendMessages: vi.fn(async () => {}),
   });
 }
@@ -379,7 +378,6 @@ describe('handleExecute new flow', () => {
       asst,
     );
     installMockTurnStore({
-      loadMessages: vi.fn(async () => []),
       appendMessages: vi.fn(async () => {}),
     });
     const emitSpy = vi.spyOn(events, 'emit').mockResolvedValue(undefined);
diff --git a/harness/tests/turn-orchestrator/preflight.test.ts b/harness/tests/turn-orchestrator/preflight.test.ts
deleted file mode 100644
index 2a09833b..00000000
--- a/harness/tests/turn-orchestrator/preflight.test.ts
+++ /dev/null
@@ -1,176 +0,0 @@
-/**
- * Smoke tests for runPreflight.
- *
- * The integration (pre-flight wired into handleStreaming) is covered by the
- * phase-11 flow-sync integration tests. These unit tests validate the
- * preflight helper's branching logic in isolation.
- */
-
-import { describe, expect, it } from 'vitest';
-import type { ISdk } from '../../src/runtime/iii.js';
-import { CompactionBusyError, ContextOverflowError } from '../../src/turn-orchestrator/errors.js';
-import { runPreflight } from '../../src/turn-orchestrator/preflight.js';
-
-type TriggerCall = { function_id: string; payload: unknown };
-
-function makeIii(overrides?: {
-  modelsGetResult?: unknown;
-  sessionTreeResult?: unknown;
-  compactNowResult?: unknown;
-}): { iii: ISdk; calls: TriggerCall[] } {
-  const calls: TriggerCall[] = [];
-
-  const iii = {
-    trigger: async <_T, R>(req: { function_id: string; payload: unknown }): Promise<R> => {
-      calls.push({ function_id: req.function_id, payload: req.payload });
-
-      if (req.function_id === 'router::models::get') {
-        const m = overrides?.modelsGetResult ?? null;
-        return (m ? { model: m } : null) as R;
-      }
-      if (req.function_id === 'session::messages') {
-        return (overrides?.sessionTreeResult ?? { messages: [] }) as R;
-      }
-      if (req.function_id === 'context-compaction::compact_now') {
-        return (overrides?.compactNowResult ?? { status: 'ok' }) as R;
-      }
-      return null as R;
-    },
-  } as unknown as ISdk;
-
-  return { iii, calls };
-}
-
-const smallMessage = { role: 'user' as const, content: [] as never[], timestamp: 0 };
-
-describe('runPreflight', () => {
-  it('returns ok without calling compact_now when under the limit', async () => {
-    const { iii, calls } = makeIii({
-      modelsGetResult: { context_window: 200_000, max_output_tokens: 8_096 },
-    });
-
-    const result = await runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3');
-
-    expect(result).toBe('ok');
-    expect(calls.some((c) => c.function_id === 'context-compaction::compact_now')).toBe(false);
-  });
-
-  it('returns ok and skips compact_now when router::models::get returns null', async () => {
-    const { iii, calls } = makeIii({ modelsGetResult: null });
-
-    const result = await runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3');
-
-    expect(result).toBe('ok');
-    expect(calls.some((c) => c.function_id === 'context-compaction::compact_now')).toBe(false);
-  });
-
-  it('calls compact_now and returns compacted when projected >= usable', async () => {
-    const { iii, calls } = makeIii({
-      modelsGetResult: { context_window: 10, max_output_tokens: 0 },
-      compactNowResult: { status: 'ok' },
-    });
-
-    const result = await runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3');
-
-    expect(result).toBe('compacted');
-    expect(calls.some((c) => c.function_id === 'context-compaction::compact_now')).toBe(true);
-  });
-
-  it('throws ContextOverflowError when compact_now returns overflow', async () => {
-    const { iii } = makeIii({
-      modelsGetResult: { context_window: 10, max_output_tokens: 0 },
-      compactNowResult: { status: 'overflow' },
-    });
-
-    await expect(
-      runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3'),
-    ).rejects.toBeInstanceOf(ContextOverflowError);
-  });
-
-  it('throws CompactionBusyError when compact_now returns busy', async () => {
-    const { iii } = makeIii({
-      modelsGetResult: { context_window: 10, max_output_tokens: 0 },
-      compactNowResult: { status: 'busy' },
-    });
-
-    await expect(
-      runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3'),
-    ).rejects.toBeInstanceOf(CompactionBusyError);
-  });
-
-  it("returns ok without claiming compacted when compact_now reports 'empty'", async () => {
-    const { iii } = makeIii({
-      modelsGetResult: { context_window: 10, max_output_tokens: 0 },
-      compactNowResult: { status: 'empty' },
-    });
-
-    const result = await runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3');
-
-    expect(result).toBe('ok');
-  });
-
-  it('returns ok on an unknown compact_now status instead of claiming compacted', async () => {
-    const { iii } = makeIii({
-      modelsGetResult: { context_window: 10, max_output_tokens: 0 },
-      compactNowResult: { status: 'something-new' },
-    });
-
-    const result = await runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3');
-
-    expect(result).toBe('ok');
-  });
-
-  it('uses the pre-resolved model and skips router::models::get', async () => {
-    const { iii, calls } = makeIii({
-      modelsGetResult: { context_window: 1, max_output_tokens: 0 },
-    });
-    const modelMeta = {
-      id: 'claude-3',
-      provider: 'anthropic',
-      api: 'anthropic-messages',
-      display_name: 'Claude 3',
-      context_window: 200_000,
-      max_output_tokens: 8_096,
-    };
-
-    const result = await runPreflight(
-      iii,
-      'session-1',
-      [smallMessage],
-      'anthropic',
-      'claude-3',
-      modelMeta,
-    );
-
-    expect(result).toBe('ok');
-    expect(calls.some((c) => c.function_id === 'router::models::get')).toBe(false);
-    expect(calls.some((c) => c.function_id === 'context-compaction::compact_now')).toBe(false);
-  });
-
-  it('fetches router::models::get when no pre-resolved model is threaded', async () => {
-    const { iii, calls } = makeIii({
-      modelsGetResult: { context_window: 200_000, max_output_tokens: 8_096 },
-    });
-
-    const result = await runPreflight(iii, 'session-1', [smallMessage], 'anthropic', 'claude-3');
-
-    expect(result).toBe('ok');
-    expect(calls.some((c) => c.function_id === 'router::models::get')).toBe(true);
-  });
-
-  it('passes session_id and model info to compact_now', async () => {
-    const { iii, calls } = makeIii({
-      modelsGetResult: { context_window: 10, max_output_tokens: 0 },
-      compactNowResult: { status: 'ok' },
-    });
-
-    await runPreflight(iii, 'my-session', [smallMessage], 'openai', 'gpt-4o');
-
-    const compactCall = calls.find((c) => c.function_id === 'context-compaction::compact_now');
-    expect(compactCall).toBeDefined();
-    const p = compactCall?.payload as Record<string, unknown>;
-    expect(p.session_id).toBe('my-session');
-    expect((p.model as Record<string, unknown>).id).toBe('gpt-4o');
-    expect((p.model as Record<string, unknown>).providerID).toBe('openai');
-  });
-});
diff --git a/harness/tests/turn-orchestrator/run-transition.test.ts b/harness/tests/turn-orchestrator/run-transition.test.ts
index f609b303..44a468bc 100644
--- a/harness/tests/turn-orchestrator/run-transition.test.ts
+++ b/harness/tests/turn-orchestrator/run-transition.test.ts
@@ -1,6 +1,6 @@
 import { afterEach, describe, expect, it, vi } from 'vitest';
 import type { ISdk } from '../../src/runtime/iii.js';
-import { CompactionBusyError, TransientError } from '../../src/turn-orchestrator/errors.js';
+import { TransientError } from '../../src/turn-orchestrator/errors.js';
 import { TURN_STATE_SCOPE } from '../../src/turn-orchestrator/state.js';
 import { runTransition } from '../../src/turn-orchestrator/run-transition.js';
 import {
@@ -156,45 +156,21 @@ describe('runTransition error model', () => {
     expect(ended).toBe(true);
   });
 
-  it('re-throws TransientError so the queue retries', async () => {
-    const { iii } = fakeIii({ ...base });
-    await expect(
-      runTransition(
-        iii,
-        'function_execute',
-        async () => {
-          throw new TransientError('retry me');
-        },
-        { session_id: 's1' },
-      ),
-    ).rejects.toThrow('retry me');
-  });
-
-  it('CompactionBusyError IS a TransientError — the subclass relation is the fix', () => {
-    // A one-line revert of `extends TransientError` back to `extends Error`
-    // resurrects the terminal-failure bug; pin the contract at the source.
-    const err = new CompactionBusyError('compaction already in progress');
-    expect(err).toBeInstanceOf(TransientError);
-    expect(err).toBeInstanceOf(Error);
-    expect(err.name).toBe('CompactionBusyError');
-  });
-
-  it('re-throws CompactionBusyError (transient) instead of failing the session', async () => {
-    // Regression: a busy compaction lease (async post-turn summarize of a
-    // large session) used to route the turn to terminal `failed` with
-    // "response failed: from assistant_streaming: compaction already in
-    // progress". Busy is lease-TTL-bounded — the queue must retry instead.
+  it('re-throws TransientError so the queue retries instead of failing the session', async () => {
+    // A transient handler error (e.g. a serialize-lease contender) must be
+    // re-thrown for the durable queue to retry — never routed to terminal
+    // `failed`.
     const { iii, writes } = fakeIii({ ...base, state: 'assistant_streaming' });
     await expect(
       runTransition(
         iii,
         'assistant_streaming',
         async () => {
-          throw new CompactionBusyError('compaction already in progress');
+          throw new TransientError('retry me');
         },
         { session_id: 's1' },
       ),
-    ).rejects.toThrow('compaction already in progress');
+    ).rejects.toThrow('retry me');
     // The session record must NOT be routed to failed.
     const failedWrite = writes.find(
       (w) =>
diff --git a/harness/tests/turn-orchestrator/store.test.ts b/harness/tests/turn-orchestrator/store.test.ts
index 00b3b2a6..3490f36b 100644
--- a/harness/tests/turn-orchestrator/store.test.ts
+++ b/harness/tests/turn-orchestrator/store.test.ts
@@ -126,13 +126,6 @@ describe('session-manager call reduction', () => {
     return { iii, calls };
   }
 
-  it('loadMessages reads the window in one paginated read without re-ensuring', async () => {
-    const { iii, calls } = recordingIii();
-    await createTurnStore(iii).loadMessages('sess-a');
-    expect(calls.filter((c) => c === 'session::messages')).toHaveLength(1);
-    expect(calls).not.toContain('session::ensure');
-  });
-
   it('appendMessages writes without re-ensuring the session', async () => {
     const { iii, calls } = recordingIii();
     const msg = {
diff --git a/iii-permissions.yaml b/iii-permissions.yaml
index 79fc1abd..bf646d3e 100644
--- a/iii-permissions.yaml
+++ b/iii-permissions.yaml
@@ -63,6 +63,12 @@ rules:
   - '!session::update_message'
   - '!session::fork'
   - '!session::set_active_leaf'
+  # context-manager: assemble/compact are pure transforms (nothing to leak)
+  # but each can spend a summariser LLM call, so deny them to in-run agents in
+  # this cost-sensitive deployment (context-manager.md § Agent exposure).
+  # count_tokens/prune are always safe and stay at the needs_approval default.
+  - '!context::assemble'
+  - '!context::compact'
   # Internal lifecycle hooks — agents must never force a config reload directly.
   # The handlers already re-fetch from configuration::get (ignoring the event
   # payload), so a direct call cannot inject config. These denies are
diff --git a/tech-specs/2026-06-agentic/context-manager.md b/tech-specs/2026-06-agentic/context-manager.md
index c7b29a32..ce7eb27c 100644
--- a/tech-specs/2026-06-agentic/context-manager.md
+++ b/tech-specs/2026-06-agentic/context-manager.md
@@ -22,7 +22,7 @@ callers don't summarise the same logical session concurrently) under its own iii
 
 - A non-chat feature that needs "summarise these messages to fit model X" calls `context::compact`
   directly.
-- A cost-sensitive caller calls `context::count_tokens` before deciding which model to use, with no
+- A cost-sensitive caller calls `context::count-tokens` before deciding which model to use, with no
   agent loop involved.
 - A different harness implementation reuses `context::assemble` as its pre-flight step and persists
   results into its own store.
@@ -85,7 +85,7 @@ Whatever pruning or compaction does, the returned context must still be accepted
 - `context::compact` — Summarise older history into a single compaction summary and return the
   preserved tail. Transient: the caller uses the result; the session keeps its full transcript.
 - `context::prune` — Strip/truncate verbose function outputs without summarising. A cheaper first pass.
-- `context::count_tokens` — Estimate token usage for a set of messages (+ optional invocation schema /
+- `context::count-tokens` — Estimate token usage for a set of messages (+ optional invocation schema /
   system) vs a model.
 
 ## Triggers
@@ -110,7 +110,7 @@ iii.registerFunction("context::on_message_added", async (evt) => {
   });
   // Measure / pre-warm; no persistence. The harness still calls context::assemble on the hot path.
   await iii.trigger({
-    function_id: "context::count_tokens",
+    function_id: "context::count-tokens",
     payload: { messages: messages.map((m) => m.message), model: { id: "<model>" } },
   });
 });
@@ -303,7 +303,7 @@ type PruneResponse = {
 };
 ```
 
-### `context::count_tokens`
+### `context::count-tokens`
 
 Estimate token usage for a set of messages, optionally including the invocation schema (typically the
 single `agent_trigger` entry) and a system prompt.
@@ -351,7 +351,7 @@ type CountTokensResponse = {
 
 All functions are pure transforms over caller-supplied messages — nothing secret to leak (see
 [README § Security model](README.md#security-model)). The only caveat is cost: `context::assemble`
-and `context::compact` can trigger a summariser LLM call. `context::count_tokens` and
+and `context::compact` can trigger a summariser LLM call. `context::count-tokens` and
 `context::prune` are safe; deny the other two in cost-sensitive deployments.
 
 ## Boundaries
diff --git a/tech-specs/2026-06-agentic/presentation/src/content/workers.ts b/tech-specs/2026-06-agentic/presentation/src/content/workers.ts
index e5605360..df9e9899 100644
--- a/tech-specs/2026-06-agentic/presentation/src/content/workers.ts
+++ b/tech-specs/2026-06-agentic/presentation/src/content/workers.ts
@@ -90,7 +90,7 @@ export const WORKERS: Record<string, WorkerInfo> = {
       { id: 'context::assemble', desc: 'system prompt + budgeted messages: count, prune, compact, fit.' },
       { id: 'context::compact', desc: 'summarise older history into one compaction summary.' },
       { id: 'context::prune', desc: 'strip verbose function output — the cheap first pass.' },
-      { id: 'context::count_tokens', desc: 'estimate usage before committing to a model.' },
+      { id: 'context::count-tokens', desc: 'estimate usage before committing to a model.' },
     ],
     notes: [
       'one summarisation per overflow, amortised — the harness persists the summary, so it never re-pays per turn.',