claude-mem/Auto Run Docs/PR-Triage/PR-Triage-08.md

Phase 08: Session Management & Message Processing

These PRs fix various session lifecycle issues — orphaned messages, provider recovery, infinite loops, and stateless provider support.

High Priority (Data Loss / Cost Risk)

• Review PR #934 (fix: terminate session on prompt-too-long instead of retrying indefinitely by @jayvenn21). File: src/services/worker/SDKAgent.ts. Currently, prompt-too-long errors cause infinite retry loops. Steps: (1) gh pr checkout 934 (2) Review — should detect the specific error and terminate the session cleanly (3) Verify no data loss on termination (pending observations should still be saved) (4) Run npm run build (5) If clean: gh pr merge 934 --rebase --delete-branch

  • MERGED (2026-02-06): Clean +5 line fix. Detects "Prompt is too long" in SDK response text and throws a fatal error, which propagates up to startSessionProcessor catch handler for clean session termination. No data loss — previously processed observations are already saved before this point. Build passes.

• Review PR #693 (fix: prevent infinite restart loop that causes runaway API costs by @ajbmachon). Files: src/services/worker-types.ts, GeminiAgent.ts, OpenRouterAgent.ts, SessionManager.ts, SessionRoutes.ts. Steps: (1) gh pr checkout 693 (2) Review restart limiting logic — should have max restart count or cooldown (3) Verify it applies to all providers (Claude, Gemini, OpenRouter) (4) Run npm run build (5) If clean: gh pr merge 693 --rebase --delete-branch

  • MERGED (2026-02-06): Applied directly to main (branch was 3+ weeks old). Adds consecutiveRestarts counter to ActiveSession, limits crash-recovery restarts to 3 with exponential backoff (1s, 2s, 4s), and adds defensive memorySessionId checks in GeminiAgent and OpenRouterAgent before expensive LLM calls. Build passes. Addresses $402+ runaway cost scenario reported by the author.

Session Processing

• Review PR #940 (fix: Backfill project field on SDK session creation by @miclip). Files: src/services/sqlite/SessionStore.ts, src/services/sqlite/sessions/create.ts. Steps: (1) gh pr checkout 940 (2) Review — should populate the project field when creating a session so observations are properly scoped (3) Small, focused change (4) Run npm run build (5) If clean: gh pr merge 940 --rebase --delete-branch

  • MERGED (2026-02-06): Clean +18/-2 line fix. Adds a conditional UPDATE after INSERT OR IGNORE to backfill the project field when a later hook provides it — only updates when existing value is NULL or empty. Fixes race condition where PostToolUse hook creates the session before UserPromptSubmit sets the project. Both functional (sessions/create.ts) and class (SessionStore.ts) versions updated identically. Build passes.

• Review PR #937 (fix(worker): gracefully process orphaned pending messages after session termination by @jayvenn21). Files: src/services/sqlite/PendingMessageStore.ts, src/services/worker-service.ts, src/services/worker/SessionManager.ts. Steps: (1) gh pr checkout 937 (2) Review — orphaned messages should be processed or discarded cleanly, not stuck forever (3) Run npm run build (4) If clean: gh pr merge 937 --rebase --delete-branch

  • MERGED (2026-02-06): Clean +125/-3 line fix addressing issue #936 (51+ orphaned messages/day). Adds isSessionTerminatedError() to detect SDK resume failures from closed terminals. On failure, falls back to Gemini/OpenRouter agents if available to drain the queue. If no fallback, markAllSessionMessagesAbandoned() marks pending/processing messages as failed and removeSessionImmediate() cleans up the session without deadlocking on the generator promise. Build passes. Rebased cleanly onto main.

• Review PR #899 (fix: resolve message processing failures in multi-session scenarios by @hahaschool). Files: 7 files including SessionStore, SDKAgent, ResponseProcessor, SessionRoutes. Steps: (1) gh pr checkout 899 (2) This is a broader fix — review carefully for scope creep (3) Check that multi-session message routing is correct (messages go to the right session) (4) Run npm run build (5) If focused and correct: gh pr merge 899 --rebase --delete-branch. If too broad, request scope reduction.

  • CLOSED (2026-02-06): Too broad — conflicts in 3 files (worker-service.ts, worker-types.ts, SDKAgent.ts) due to overlap with recently merged PRs #693, #937, and #940. Crash recovery logic already addressed by #693 (restart limits with exponential backoff) and #937 (orphaned message fallback + session termination detection). Requested author re-submit as focused PRs for the genuinely valuable parts: (1) AbortController stale-check at generator start (~10 lines, real bug), (2) FK constraint protection in SDKAgent/ResponseProcessor (prevents FK violation when SDK generates new session ID after restart), (3) queueDepth logging fix using pendingStore.getPendingCount() instead of deprecated session.pendingMessages.length.

• Review PR #627 (fix: Reset AbortController before starting generator by @TranslateMe). File: src/services/worker/http/routes/SessionRoutes.ts. Steps: (1) gh pr checkout 627 (2) Old PR (Dec 27) — check if still applicable after v8.5.2 memory leak fix (3) If the abort controller reset is still needed: rebase and merge. If already handled: close.

  • MERGED (2026-02-06): Clean +10 line fix still applicable. Adds stale-AbortController check at the start of startGeneratorWithProvider() — if signal.aborted is already true, creates a fresh AbortController before proceeding. This prevents infinite "Generator aborted" loops where: (1) session aborts, setting signal.aborted=true, (2) generatorPromise is set to null, (3) new observations trigger ensureGeneratorRunning, (4) new generator immediately sees stale abort signal and exits, creating an infinite loop. The crash recovery path (merged in PR #693) already resets the controller for non-abort exits, but this fix covers the abort case. Rebased cleanly onto main, build passes.

• Review PR #741 (fix: Provider-aware recovery and stale session cleanup by @licutis). File: src/services/worker-service.ts. Steps: (1) gh pr checkout 741 (2) Review provider-aware recovery logic — should handle Gemini/OpenRouter differently from Claude SDK (3) Run npm run build (4) If clean: gh pr merge 741 --rebase --delete-branch

  • MERGED (2026-02-06): Applied directly to main (branch was 3 weeks old with conflicts from PRs #693, #937, #627). Two clean changes: (1) Adds getActiveAgent() to WorkerService matching SessionRoutes logic — startup recovery now uses the correct provider (OpenRouter/Gemini/SDK) instead of hardcoding SDKAgent. (2) Stale session cleanup in processPendingQueues() — marks sessions stuck 'active' for 6+ hours as failed along with their pending messages before processing orphaned queues. Conflicts resolved by combining PR #741's dynamic agent selection with PR #937's session-terminated fallback logic. Build passes.

Stateless Provider Support

These fix Gemini/OpenRouter providers that don't have SDK sessions.

• Review PR #910 (fix: complete stateless provider support with enhanced validation by @Scheevel). WARNING: This PR modifies ~90 files, most of which are CLAUDE.md files that should NOT have been included. Steps: (1) gh pr checkout 910 (2) Identify the actual source changes (look at .ts files only, ignore CLAUDE.md files) (3) Key files: src/services/sqlite/SessionStore.ts, src/services/worker/GeminiAgent.ts, src/services/worker/OpenRouterAgent.ts, src/services/worker/agents/ResponseProcessor.ts (4) If the core logic is good but CLAUDE.md pollution is unacceptable, request changes to remove all CLAUDE.md files from the PR. Or cherry-pick just the source changes.

  • CLOSED (2026-02-06): Stale branch with critical regressions against 5 recently merged PRs. Reverts: consecutiveRestarts (PR #693 — infinite restart prevention), removeSessionImmediate() (PR #937 — orphaned message handling), getCredential() (Issue #733 — centralized env), CLAUDE_MEM_FOLDER_CLAUDEMD_ENABLED check (PR #913), and createIterator() onIdleTimeout. Also bundles 3 unrelated features: observation deduplication (+285 lines Levenshtein in ResponseProcessor), summary soft-delete (hidden column + PATCH API), and deletion of ~80 CLAUDE.md files + compiled plugin artifacts. Core synthetic memorySessionId generation is valuable — requested re-submission as focused PRs: (1) synthetic ID gen for stateless providers, (2) observation deduplication, (3) summary soft-delete.

• Evaluate PR #615 (fix: generate memorySessionId for stateless providers by @JiehoonKwak). Files: 6 files. Steps: (1) gh pr checkout 615 (2) Check if #910 supersedes this (both fix stateless provider session IDs) (3) If #910 is more complete, close #615: gh pr close 615 --comment "Superseded by PR #910 which provides more complete stateless provider support. Thank you!" (4) If #615 has unique value, rebase and merge first.