@context-chef/tanstack-ai@0.2.1

Patch Changes

• 2e13c66 Thanks @MyPrototypeWhat! - Bump @context-chef/core peer to pick up the new media-aware compression strategy (attachments are now stripped to [image] / [document] text placeholders before reaching the compression model). No source changes in this package — multimodal compression behavior is driven entirely by core.

• Updated dependencies [2e13c66]:

  • @context-chef/core@3.2.0

@context-chef/core@3.2.0

Minor Changes

• 2e13c66 Thanks @MyPrototypeWhat! - ### Compression now strips media attachments to text placeholders

  Janitor.executeCompression() no longer ships binary attachment data through the compression call. Each attachment in the messages being compressed is replaced inline with a [image] / [image: photo.png] / [document] / [document: report.pdf] text marker before the compressionModel is invoked. The summarizer sees that media existed at this point in the conversation without being asked to process raw base64.

  • Modeled on Claude Code's stripImagesFromMessages strategy
  • Avoids prompt-too-long failures on the compression call itself when histories contain many images
  • Empty mediaType produces [attachment] instead of misleading [document]
  • toKeep (the recent messages preserved verbatim) is untouched — its attachments still reach the main model through the target adapter

  Removed Prompts.MEDIA_DESCRIPTION_INSTRUCTION

  The constant is gone from the exported Prompts object. It was previously appended to the compression prompt when attachments were detected, asking the compression model to "describe the visual content." In practice this never worked — compressionModel is a (Message[]) => Promise<string> function with no adapter pipeline, so the binary data on Message.attachments was never actually forwarded to the LLM. The new placeholder-based strategy supersedes it.

  If you imported Prompts.MEDIA_DESCRIPTION_INSTRUCTION directly, remove the reference — the behavior it described was already a no-op.

@context-chef/ai-sdk-middleware@1.1.3

Patch Changes

• 2e13c66 Thanks @MyPrototypeWhat! - fromAISDK() now maps AI SDK FilePart (type 'file') on user and assistant messages to IR attachments, so multimodal turns participate in the new core compression placeholder logic ([image] / [document] markers in the compression payload).

  Attachment.data in the middleware path is a presence/metadata signal only — Janitor reads m.attachments?.length for placeholder injection but never the binary itself. The actual Uint8Array / URL / string payload round-trips losslessly through _userContent / _assistantContent, which toAISDK() hands back to the underlying AI SDK provider verbatim. No re-encoding, no data loss.

• Updated dependencies [2e13c66]:

  • @context-chef/core@3.2.0

@context-chef/tanstack-ai@0.2.0

Minor Changes

• 246175c Thanks @MyPrototypeWhat! - ### New Package: @context-chef/tanstack-ai

  TanStack AI ChatMiddleware powered by context-chef. Drop in as a single middleware to get transparent history compression, tool result truncation, and token budget management.

  Features

  • History Compression — automatically compresses older messages when conversation exceeds the token budget, with optional LLM-based summarization via a cheap adapter
  • Tool Result Truncation — large tool outputs are truncated while preserving head and tail, with optional VFS storage for full content retrieval
  • Token Budget Tracking — extracts promptTokens from onUsage callbacks and feeds it back to the compression engine automatically
  • Compact (Mechanical Pruning) — zero-LLM-cost removal of tool call/result pairs and empty messages with configurable retention modes
  • Dynamic State Injection — injects runtime state as XML into the last user message or system prompt on every call
  • Transform Context Hook — custom post-processing for RAG injection or prompt manipulation

  Adapter

  • fromTanStackAI() / toTanStackAI() — lossless round-trip converters between TanStack AI ModelMessage[] and context-chef Message[] IR
  • Preserves multimodal ContentPart[] content and providerMetadata on tool calls through round-trip via _originalContent / _originalToolCalls pass-through fields

Patch Changes

• Updated dependencies [246175c]:
  • @context-chef/core@3.1.1

@context-chef/core@3.1.1

Patch Changes

• 246175c Thanks @MyPrototypeWhat! - Change license from ISC to MIT

@context-chef/core@3.1.0

Minor Changes

• d6169e4 Thanks @MyPrototypeWhat! - ### Multimodal Attachment Support

  • Added Attachment interface and Message.attachments field to IR for provider-neutral media representation
  • Janitor detects attachments during compression and augments the prompt with MEDIA_DESCRIPTION_INSTRUCTION to guide the compression model toward describing image/media content in summaries
  • Output adapters (compile()) now convert attachments to provider-specific formats:
    • OpenAI: image_url / file content parts
    • Anthropic: image / document content blocks
    • Gemini: inlineData / fileData parts

  Input Adapters (Provider → IR)

  • Added fromOpenAI(), fromAnthropic(), fromGemini() to convert provider-native messages to ContextChef IR
  • Returns { system, history } — automatically separates system messages from conversation history
  • Multimodal content (images, files, documents) automatically mapped to IR attachments
  • New types: HistoryMessage, ParsedMessages

@context-chef/ai-sdk-middleware@1.1.2

Patch Changes

• 246175c Thanks @MyPrototypeWhat! - Change license from ISC to MIT

• Updated dependencies [246175c]:

  • @context-chef/core@3.1.1

@context-chef/ai-sdk-middleware@1.1.1

Patch Changes

• Updated dependencies [d6169e4]:
  • @context-chef/core@3.1.0

@context-chef/ai-sdk-middleware@1.1.0

Minor Changes

• dceea52 Thanks @MyPrototypeWhat! - Replace compact implementation with AI SDK's pruneMessages

  Breaking change to CompactConfig:

  Before:

  compact: {
    clear: ["thinking", { target: "tool-result", keepRecent: 5 }];
  }

  After:

  compact: { reasoning: 'all', toolCalls: 'before-last-message' }

  • CompactConfig.clear replaced with reasoning, toolCalls, and emptyMessages fields, matching pruneMessages parameters
  • Compact now runs before IR conversion (on raw AI SDK messages) instead of after
  • Removed TOOL_RESULT_CLEARED_INSTRUCTION system prompt injection — pruneMessages removes chunks entirely rather than replacing with placeholders
  • Per-tool pruning support via toolCalls array form: [{ type: 'before-last-message', tools: ['search'] }]

@context-chef/core@3.0.3

Patch Changes

• 25b2b98 Thanks @MyPrototypeWhat! - Janitor compression pipeline improvements and internal type cleanup.

  New Prompts.formatCompactSummary(raw) utility — strips <analysis> scratchpad blocks and extracts content from <summary> tags, falling back to cleaned raw text when no tags are present. executeCompression() now pipes compressionModel output through this cleaner before wrapping with getCompactSummaryWrapper, preventing XML scaffolding from leaking into the next context window. Before this change, the default prompt asked the model to wrap its output in <summary></summary> but nothing stripped the tags — they silently leaked into the continuation context.

  Upgraded CONTEXT_COMPACTION_INSTRUCTION — now uses a two-phase <analysis> + <summary> + <example> pattern (inspired by Claude Code's compact prompt) for measurably better summary quality. The 5 output sections remain domain-agnostic (Task Overview / Current State / Important Discoveries / Next Steps / Context to Preserve) so the prompt works for support, research, shopping, coding, or any other conversational agent — no coding-specific language introduced.

  New JanitorConfig.customCompressionInstructions?: string — additional focused instructions appended to the default prompt as an "Additional Instructions:" section. Additive (not replacement) so the default scaffolding that enforces the <analysis>/<summary> parsing contract is always preserved. Users who need radically different compression behavior can still provide their own compressionModel entirely.

  new ContextChef({
    janitor: {
      compressionModel,
      customCompressionInstructions:
        "Focus on customer sentiment, unresolved issues, and preserve ticket IDs verbatim.",
    },
  });

  Compression circuit breaker — after three consecutive compressionModel failures, subsequent compress() calls return history unchanged instead of retrying. This prevents sessions from hammering a broken compression endpoint on every turn (e.g. expired API key, rate limit lockout). The counter resets on successful compression, explicit janitor.reset(), or chef.clearHistory(). The consecutiveFailures field is part of JanitorSnapshot and preserved by chef.snapshot() / chef.restore(). restoreState() uses ?? 0 for defensive backward compatibility with snapshots serialized by older versions.

  Removed Prompts.DEEP_CONVERSATION_SUMMARIZATION — this export was unreferenced internal dead code with an inconsistent <history_summary> contract that diverged from the default prompt's <summary> contract. External code that imported it (unlikely, as it was never documented) should migrate to CONTEXT_COMPACTION_INSTRUCTION, which now covers the same detailed-summary use case via the upgraded scaffolding.

  Internal type cleanup — replaced scattered as type assertions with generics, type guards, and typed helpers across core source and test files. From 40+ cast sites, only two unavoidable assertions remain, both documented:

  • Assembler.orderKeysDeterministically<T> — single boundary assertion to express the "same shape, reordered keys" transformation which TypeScript cannot model at the type level. Function is now generic, so call sites no longer need their own casts.
  • TypedEventEmitter.emit — necessary widening to call stored EventHandler<never> with a concrete payload. Storage now uses contravariance (Set<EventHandler<never>>), so on() and off() no longer need casts; only the call site in emit() retains one, guarded by a runtime invariant established in on().

  Additional source cleanups: Pruner uses a new isRecord() type guard; VFSMemoryStore uses typed variable coercion instead of JSON.parse(...) as T; adapter implementations (anthropicAdapter, openAIAdapter, geminiAdapter) declare SDK types explicitly instead of trailing as SDKType on object literals.