feat: add memory system for persistent context across conversations#14
Open
feat: add memory system for persistent context across conversations#14
Conversation
- Add per-provider model options, defaults, and slug aliases - Add provider-aware model normalization/resolution helpers - Preserve Codex-only constants/functions for backward compatibility - Extend tests to cover Claude aliases and provider-specific fallback behavior
- introduce `ClaudeCodeAdapter` service and live layer wiring - map runtime/session/request failures into provider adapter error types - add coverage for validation, session-not-found mapping, lifecycle forwarding, and event passthrough
Add provider-service routing coverage for explicit claudeCode sessions. Co-authored-by: codex <codex@users.noreply.github.com>
Add restart semantics when requested provider changes and cover with tests. Co-authored-by: codex <codex@users.noreply.github.com>
Add provider-aware model options and include provider in turn-start dispatch. Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
Co-authored-by: codex <codex@users.noreply.github.com>
- add `@anthropic-ai/claude-agent-sdk` dependency for `apps/server` - replace placeholder Claude adapter with live session/query/event handling - add comprehensive adapter tests for runtime events, approvals, resume, rollback, and model overrides
- Replace manual prompt async iterator with `Stream.fromQueue(...).toAsyncIterable` - Use `Ref` for shared session context in tool approval callbacks - Simplify timestamp/ID generation and close sessions via queue shutdown
- Stop synthesizing `resume` from generated thread IDs - Persist resume session ID from query messages when available - Validate resume/sessionId values as UUIDs before reuse - Add tests for valid UUID resume passthrough and no synthesized resume
Co-authored-by: codex <codex@users.noreply.github.com>
- Do not pass `resumeCursor` when restarting a session after changing providers - Treat synthetic Claude thread IDs (`claude-thread-*`) as unscoped during runtime ingestion - Add tests covering provider-switch restart behavior and Claude turn lifecycle acceptance
Co-authored-by: codex <codex@users.noreply.github.com>
- Preserve active turn state when `session.started`/`thread.started` arrive mid-turn - Emit `message.delta` from assistant text when stream deltas are missing, then complete the message - Add Claude native SDK NDJSON observability logging and wire its log path in server layers - Expand ingestion/adapter tests to cover mid-turn lifecycle and delta fallback behavior
- Default Claude sessions to bypass permissions when approval policy is `never`, while preserving explicit `permissionMode` - Hide/send reasoning effort only for providers that support it in ChatView - Add coverage for Claude permission-mode derivation and precedence, and update runtime event model docs
- add per-session `sessionSequence` on provider runtime events and persist activity `sequence` - migrate `projection_thread_activities` with nullable `sequence` column + index - sort server/web activity projections by sequence fallback to timestamp/id - allow provider sessions/turns to start before a real threadId is emitted
- define `CursorAdapter` service contract and Cursor stream-json schema types - add `CursorCliStreamEvent` decoding tests for system/thinking/tool/result/retry events - add implementation plan for Cursor provider integration Co-authored-by: codex <codex@users.noreply.github.com>
- Accept `cursor` as a first-class provider in orchestration, persistence, and session directory flows - Update model/provider inference and normalization to handle Cursor model aliases - Revamp chat provider/model picker with Cursor-specific trait controls and add coverage in tests
- add `scripts/cursor-acp-probe.mjs` to run ACP protocol probing scenarios - update `package.json` scripts for probe execution - include generated probe summaries/transcripts under `.tmp/acp-probe/`
) * feat: improve review popover UX with inline expand, sorting, and visual hierarchy - Inline morph: clicking a row expands it in-place (no duplicate detail section) using CSS grid height animation for smooth reveal of action buttons - Sort by status (in_review first, then pending) and PR number descending for stable, triage-optimized ordering - Add sticky group headers ("In progress" / "Awaiting review") with backdrop blur - De-emphasize repo name, promote PR title as primary visual element - Fix nested button a11y: outer element is now div[role=button] with keyboard handlers - Stabilize useMemo dependency with module-level empty array constant * feat: improve notification panel mobile UX - Larger panel: full-width on mobile, w-96 on desktop - More height: 70dvh on mobile, 500px on desktop - Block sidebar interaction behind open panel via z-indexed portal backdrop - Raise popover positioner above Sheet (z-52) on mobile so it clears the sidebar drawer - Close panel and mobile sidebar when navigating to a review - Add positionerClassName prop to PopoverPopup for z-index overrides * fix: suppress composer auto-focus on touch devices Prevents the virtual keyboard from opening automatically when navigating to a thread on mobile. focusComposer now bails early on pointer:coarse devices, covering all call sites (navigation, mode switches, model selection, drag-drop, terminal close). * feat: persist and expose PR body and labels on review requests Fetches body and labels from the GitHub CLI alongside existing PR fields, stores them in two new columns (pr_body, pr_labels) added via migration 018, and surfaces them through the ReviewRequest contract so the notification panel can render PR description and label chips in the expanded item view.
* feat: add review approve/request-changes actions in chat header Add ReviewActionsControl component with a collapsible dropdown menu (matching Git/Jira action patterns) that appears when the current thread has an active PR review. Supports submitting APPROVE or REQUEST_CHANGES via the GitHub API, then auto-dismisses the review request locally. * perf: reduce GitHub API rate-limit consumption - Cache findLatestPr results per (cwd, branch) with 60s TTL, avoiding 1-3 gh API calls every git status poll - Cache repository clone URLs for 5 minutes (essentially static data fetched up to 4x per PR worktree setup) - Cache PR head SHA for 2 minutes during review comment publishing - Increase git status polling from 15s to 30s (stale 5s to 15s) - Increase git branches polling from 60s to 120s (stale 15s to 30s) - Increase review request polling from 60s to 120s (stale 30s to 60s) - Invalidate PR cache on PR creation to ensure fresh data after mutations * perf: further reduce GitHub API rate-limit consumption - Extract createTtlCache to @t3tools/shared/cache (dedup GitManager + wsServer) - Cache getDefaultBranch results for 15 minutes (essentially static) - Server-side 60s cache for reviewRequestList (collapses multi-tab polls) - Batch review comments into single POST /reviews call (N→1 API calls) - Add refetchIntervalInBackground: false to all polled queries (stops polling when tab is in background) - Add refetchOnWindowFocus to reviewRequestListQueryOptions - Add stdin passthrough to GitHubCli.execute for batch API payloads
- Swipe-to-close: crisp 1px dark edge with tight depth shadow instead of blurry theme-color border; sidebar content live-translates under finger - Swipe-to-open: edge indicator now uses sidebar background color, follows finger width directly, fades to fully opaque quickly - Pull-to-reveal: bell icon indicator with smooth fade/scale feedback; PullToReveal component wired into ThreadSidebar - NotificationBell: controlled open/onOpenChange props so pull gesture can drive it; mobile backdrop at z-51 blocks sidebar while panel is open; popover z-52 ensures it renders above the backdrop
Remove key={threadId} from ChatView so the DOM stays mounted across
thread switches (no unmount/remount blank frame). Add explicit resets
for all remaining state that isn't keyed by threadId: responding request
IDs, pending user input state, attachment preview handoffs, in-flight
send guard, and composer select lock.
On thread change, scroll the messages container to bottom synchronously
in useLayoutEffect (before paint) to avoid showing the previous thread's
scroll offset for a frame. Follow with a 120ms opacity fade-in so the
content change feels smooth rather than an abrupt swap.
- Add approved/changes_requested statuses to ReviewRequestStatus - Show review outcome in chat header (approved, changes requested) - Add "Done" tab in notification panel with proper outcome labels - Fix "Go to Review" link for completed reviews with existing threads - Unlink deleted threads from review requests on each poll - Fix race condition: await linkThread before navigating to review thread - Only auto-dismiss pending reviews (preserve completed review history) - Add listReviewedPrs GitHub CLI method for future use
- Add "Restart" button next to "Done" in notification panel for in_review items - Extract buildReviewPrompt and normalizePrReference to packages/shared/prReview - Add implementation plan for future server-driven review flow
…eads The query used NOT IN (alive threads), treating threads that don't exist yet in projection_threads the same as deleted threads. This caused a race where linkThread sets the threadId, but the next poll wipes it because the client-side draft thread hasn't materialized on the server yet. Changed to IN (explicitly deleted threads) so only threads that were created and then deleted get unlinked. Also reset status back to pending when unlinking so the UI stays consistent.
The server swallowed all gh api errors during comment publishing and
returned { published: 0 } as a success response. The client never
checked the count, so the UI always showed "Comment published to GitHub"
even when nothing was posted.
Now the server captures the actual gh error message and includes it in
the response. The client checks published count and throws with the real
error, which the existing error toast handler displays.
- ProviderRuntimeIngestion: use Effect.succeed(null) instead of Effect.succeed(undefined) to satisfy the effectSucceedWithVoid rule while preserving the truthiness check downstream - ClaudeCodeAdapter: wrap JSON.parse in Effect.sync callback to move try/catch out of the generator scope (tryCatchInEffectGen rule) - wsServer: use runProcess with allowNonZeroExit for batch comment publish so the GitHub API response body (stdout) is captured on 422 errors — previously only stderr was shown, losing the actual error detail. Also add server-side logging for publish attempts.
Check which files are in the PR diff before attempting to publish. If a comment targets a file outside the diff, return immediately with a clear error message instead of hitting GitHub's 422 rejection. Also adds file context hint to any remaining 422 errors.
The GitHub PR review API only accepts comments on files that are part of the diff. Previously, the publish button appeared on all comments, leading to confusing 422 errors for files not in the PR. Now DiffPanel passes the set of branch diff file paths through to useDiffAnnotations, which only attaches the onPublish callback (and thus the button) to comments whose file is in the diff. Also keeps the server-side pre-flight check as a safety net.
- Review prompt: remove /find-skills reference, condense tool docs to just mention review_comment tool. Keep skill check as Important. - Timeline: normalize MCP tool labels from "Mcp__server__tool_name" to "Tool name" for readability (e.g. "Review comment" instead of "Mcp__review-comments__review_comment").
Introduces a full-stack memory feature inspired by Claude Code's memory system. Memories can be scoped per-project or global, categorized (preference, pattern, decision, fact, convention), and created manually or (in a future phase) extracted automatically from conversations. Key components: - Effect Schema contracts for Memory entity, WS inputs/results - SQLite migration with FTS5 full-text search + sync triggers - MemoryRepository service/layer following the ReviewComment pattern - 7 WebSocket RPC methods (list, search, create, update, archive, delete, getForThread) - React Query hooks with cache invalidation and optimistic patterns - MemoryPanel UI with search, category filters, and CRUD actions - MemoryBadge in ChatHeader showing memory count per thread - MemoryCreateDialog for manual memory creation - Sidebar integration with sheet-based memory panel Retrieval uses hybrid FTS5 keyword matching + relevance/recency scoring. Auto-extraction from completed turns is designed but deferred to a follow-up phase.
github-actions
bot
added
size:XXL
vouch:trusted
Add batch memory extraction that analyzes conversation threads using Claude Haiku and stores extracted knowledge as project-scoped memories and daily summaries. Triggered manually via Extract button in MemoryPanel. Key changes: - Add MemoryExtraction service + layer with LLM-powered extraction - Extract shared runAgentQuery utility from CodexTextGeneration - Add daily scope + date field to memory schema (migration 020) - Add optional source field to MemoryCreateInput for auto vs manual - Add memory.extract WS endpoint, React Query mutation, and UI dialog - Dedup via Jaccard word-similarity against existing project memories - Two LLM passes: per-project knowledge extraction + global daily summary
Restructure the memory system from project/global/daily to project/thread/daily scopes with autonomous extraction: - Replace "global" scope with "thread" for per-thread summaries - Add MemoryReactor that auto-generates thread summaries on turn.processing.quiesced (debounced 30s per thread) - Skip re-summarization when thread has no new activity - Periodic project + daily extraction after 10 thread summaries - Improve extraction prompt quality bar (max 5, actionable only) - Add thread_id column + unique index (migration 021) - Add findThreadSummary + upsertThreadSummary to MemoryRepository - Extract shared threadTranscript module - Update UI: thread scope icon/labels, replace global references
…maries - MemoryBadge: use Button size="xs" to align with other header controls - MemoryPanel: remove Extract and Add buttons (extraction is autonomous) - Daily summaries: delete existing entries for the date before inserting new ones, preventing duplicate daily memories on re-extraction - Add deleteDailyByDate to MemoryRepository
Load existing daily summaries for today and pass them into the LLM prompt so new extractions build on earlier context instead of replacing it. The LLM produces a comprehensive merged summary, then the old entries are deleted and replaced with the combined result. - Add listDailyByDate to MemoryRepository - Pass existing summaries to buildDailySummaryPrompt - Prompt instructs LLM to incorporate earlier summaries
42tg
force-pushed
the
main
branch
7 times, most recently
from
54574d8 to
18c5f27
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds a full-stack memory feature inspired by Claude Code's recently shipped memory system. Memories persist context across conversations — things like coding preferences, architectural decisions, project conventions, and facts that the agent should remember.
Memoryentity, 5 categories (preference/pattern/decision/fact/convention), project + global scopes, manual + auto sources, and 7 WS input/result schemasMemoryRepositoryservice + layer following theReviewCommentRepositorypattern; 7 WebSocket RPC handlersMemoryPanelwith search + category filters + CRUD;MemoryCreateDialog;MemoryBadgein ChatHeader; Sidebar sheet integrationDesign decisions
sanitizeFts5Query()wraps tokens in double-quotes(project_id, archived_at, updated_at)Research references
Studied implementations from Claude Code (file-based CLAUDE.md), OpenClaw (hybrid vector+BM25), ClawMem (hooks-based), Copilot (citation-verified), and Cursor (phase-based memory bank) to inform the design.
Files changed (18 files, +1,631 lines)
Contracts (
packages/contracts/src/)memory.ts— Domain schemas: MemoryId, Memory entity, WS inputs/resultsws.ts— 7 memory WS methods + request body union entriesipc.ts— Memory namespace on NativeApi interfaceindex.ts— Barrel exportServer (
apps/server/src/)persistence/Migrations/019_Memories.ts— SQLite table + FTS5 virtual table + triggers + indexespersistence/Migrations.ts— Migration registrationpersistence/Services/MemoryRepository.ts— Service interfacepersistence/Layers/MemoryRepository.ts— Full implementation with FTS5 searchserverLayers.ts— Layer registrationwsServer.ts— 7 WS route handlersWeb (
apps/web/src/)lib/memoryReactQuery.ts— Query keys, query options, mutation optionscomponents/MemoryPanel.tsx— Main panel with search, filters, memory cardscomponents/MemoryCreateDialog.tsx— Dialog form for manual memory creationcomponents/MemoryBadge.tsx— Badge showing memory count in ChatHeadercomponents/Sidebar.tsx— Memory sheet integrationcomponents/chat/ChatHeader.tsx— MemoryBadge integrationcomponents/ChatView.tsx— Pass projectId to ChatHeaderwsNativeApi.ts— Memory namespace transport bindingsTest plan
bun typecheck,bun fmt,bun lintall pass (✅ confirmed)