- architecture-overview: add 'text' language to all fenced code blocks (MD040)
- architecture-overview: split 'Stop' lifecycle into 'Summary' + 'SessionEnd'
to match canonical 5-hook naming
- production-guide: reference PR numbers for proposed settings not yet upstream
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Architecture overview covers the 4-layer system design, hook lifecycle,
data flow, and key patterns (CLAIM-CONFIRM, circuit-breaker, graceful
degradation, deduplication, dual session IDs).
Production guide provides recommended settings, health monitoring
metrics and thresholds, quick health check commands, multi-machine
sync setup, growth expectations, common issues with solutions, and
log analysis tips.
Based on 23 days of production usage with 3,400+ observations
across two physical servers and 8 projects.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Critical:
- migrations: change version 8 → 25 to avoid collision with
MigrationRunner.addObservationHierarchicalFields (uses version 8)
- SessionRoutes: remove duplicate imports that prevent compilation
Major:
- SessionRoutes: call applyTierRouting() before every generator spawn
(stale-recovery and crash-recovery paths were missing it)
- applyTierRouting: clear session.modelOverride at top before re-evaluating
to prevent stale tier from persisting across spawns
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tier Routing:
- Inspect pending queue before starting generator
- Summarize messages → CLAUDE_MEM_TIER_SUMMARY_MODEL (e.g., Opus)
- All simple tools (Read, Glob, Grep, LS) → CLAUDE_MEM_TIER_SIMPLE_MODEL (Haiku)
- Mixed/complex → default model (no override)
- session.modelOverride in ActiveSession, used by SDKAgent.getModelId()
- peekPendingTypes() in PendingMessageStore for non-claiming inspection
- Configurable via CLAUDE_MEM_TIER_ROUTING_ENABLED (default: true)
Feedback Collection (schema only):
- New observation_feedback table via MigrationRunner (schema version 24)
- Tracks signal_type (semantic_inject_hit, search_accessed, etc.)
- Indexes on observation_id and signal_type
- Foundation for future Thompson Sampling optimization
Production data (24h tier routing test):
- 36 Haiku observations in 4 min, quality indistinguishable from Sonnet
- Estimated ~52% cost reduction on SDK Agent usage
- 835 → 6,695 feedback signals collected over 13 days
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Follow-up to PR #1568: fix stale doc comment that still said GET, and add
limit parameter validation (default 5, clamped to 1-20 range).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: semantic context injection via Chroma on every UserPromptSubmit
On each prompt, queries ChromaDB for the top-N most relevant past
observations and injects them as additionalContext. Replaces the
recency-based "last N observations" approach with relevance-based
semantic search.
Changes:
- session-init.ts: After session init, query /api/context/semantic
with user's prompt text. If results found, return as
hookSpecificOutput with hookEventName 'UserPromptSubmit'.
- SearchRoutes.ts: New GET /api/context/semantic endpoint that queries
SearchManager with format='json' and formats results as markdown.
- SettingsDefaultsManager.ts: New settings CLAUDE_MEM_SEMANTIC_INJECT
(default: true) and CLAUDE_MEM_SEMANTIC_INJECT_LIMIT (default: 5).
Key behaviors:
- Fires on every UserPromptSubmit (not just SessionStart)
- Minimum prompt length: 20 chars (skips "ok", "yes", etc.)
- Skips media-only prompts
- Graceful degradation: if worker/Chroma unavailable, no injection
- Survives /clear: re-injects on next prompt (not session-bound)
- Uses workerHttpRequest (v10.6.3 API, not raw fetch)
Production data (23 days, 3,400+ observations):
- Before: 8 most recent observations (often irrelevant to current topic)
- After: 5 most relevant observations (semantic match)
- Token cost: ~1800 → ~800-1200 per injection
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address CodeRabbit review on PR #1568
- session-init: don't skip semantic injection when contextInjected=true
(only skip agent re-init, semantic lookup must run every prompt)
- session-init: normalize SEMANTIC_INJECT toggle via String().toLowerCase()
- semantic endpoint: change from GET to POST to avoid URL-length limits
and prompt exposure in access logs. Handler accepts both body and query
for backwards compatibility.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Alessandro Costa <alessandro@claudio.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: resolve 3 upstream bugs in summarize, ChromaSync, and HealthMonitor
1. summarize.ts: Skip summary when transcript has no assistant message.
Prevents error loop where empty transcripts cause repeated failed
summarize attempts (~30 errors/day observed in production).
2. ChromaSync.ts: Fallback to chroma_update_documents when add fails
with "IDs already exist". Handles partial writes after MCP timeout
without waiting for next backfill cycle.
3. HealthMonitor.ts: Replace HTTP-based isPortInUse with atomic socket
bind on Unix. Eliminates TOCTOU race when two sessions start
simultaneously (HTTP check is non-atomic — both see "port free"
before either completes listen()). Updated tests accordingly.
All three bugs are pre-existing in v10.5.5. Confirmed via log analysis
of 543K lines over 17 days of production usage across two servers.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* chore: add CONTRIB_NOTES.md to gitignore
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address CodeRabbit review on PR #1566
- HealthMonitor: add APPROVED OVERRIDE annotation for Win32 HTTP fallback
- ChromaSync: replace chroma_update_documents with delete+add for proper
upsert (update only modifies existing IDs, silently ignores missing ones)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Alessandro Costa <alessandro@claudio.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bidirectional sync of observations and session summaries between
machines via SSH/SCP. Exports to JSON, transfers, imports with
deduplication by (created_at, title).
Commands:
claude-mem-sync push <remote-host> # local → remote
claude-mem-sync pull <remote-host> # remote → local
claude-mem-sync sync <remote-host> # bidirectional
claude-mem-sync status <remote-host> # compare counts
Features:
- Deduplication prevents duplicates on repeated runs
- Configurable paths via CLAUDE_MEM_DB / CLAUDE_MEM_REMOTE_DB
- Automatic temp file cleanup
- Requires only Python 3 + SSH on both machines
Tested syncing 3,400+ observations between two physical servers.
After sync, a session on the remote server used the transferred
memory to deliver a real feature PR — proving productive
cross-machine workflows.
Co-authored-by: Alessandro Costa <alessandro@claudio.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: drain orphaned pending messages on session completion (SIGTERM)
When deleteSession() aborts the SDK agent via SIGTERM, pending messages
in the queue are never processed. Without drain, they remain in
'pending' status forever — no future generator picks them up because
the session is already completed.
Adds markAllSessionMessagesAbandoned() call after deleteSession() in
completeByDbId(). This reuses the existing PendingMessageStore method
already used by worker-service.ts terminateSession().
Production evidence: 15 orphaned summarize messages found across
completed sessions (ages 3h to 3 days) before this fix. After fix:
0 orphaned messages over 23 days of operation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: document best-effort drain limitation per CodeRabbit review #1567
Add comment noting the rare race condition when generators outlive the
30s SIGTERM timeout. Practical risk is negligible (0 orphans over 23 days).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Alessandro Costa <alessandro@claudio.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
readGeminiSettings() throws on corrupt JSON since ae6915b, but
checkGeminiCliHooksStatus() called it without catching — violating
its "returns 0 always" contract.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- cpSync now does rmSync before copy to avoid stale file merges
- setupIDEs() returns failed IDE list; install reports partial success
- runSmartInstall() returns boolean status instead of void
- Worker port in next-steps URL reads CLAUDE_MEM_WORKER_PORT env var
- Goose YAML regex stops at column-0 keys (prevents eating sibling sections)
- AGENTS.md uninstall removes header-only stub files
- findBunPath() validated before use in WindsurfHooksInstaller
- Cursor marked unsupported in ide-detection until installer is wired
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add shebang banner to NPX CLI esbuild config so npx claude-mem works
- Remove manual backslash pre-escaping in WindsurfHooksInstaller (JSON.stringify handles it)
- Scope cache deletion to claude-mem only, not entire vendor namespace
- Use getWorkerPort() in OpenCodeInstaller instead of hard-coded 37777
- Throw on corrupt JSON in readJsonSafe/readGeminiSettings/Windsurf to prevent data loss
- Fix Cursor install stub to warn instead of silently succeeding
- Fix Gemini uninstall to remove individual hooks within groups, not whole groups
- Update tests for new corrupt-file-throws behavior
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
SessionEnd has a 1.5s hardcoded cap from Claude Code (CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS),
making it unsuitable for waiting on async work. Previously, the Stop hook would fire-and-forget
the summarize request, then SessionEnd would immediately call deleteSession — aborting the SDK
agent mid-summary.
Now the Stop hook (120s timeout, no cap) owns the full lifecycle:
1. Queue summarize request
2. Poll new GET /api/sessions/status endpoint until queue drains
3. Call /api/sessions/complete after summary finishes
SessionEnd is now a true fire-and-forget fallback (process.exit(0) immediately).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Resolve merge conflicts in adapter index, gemini-cli adapter, and rebuilt CJS artifacts.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes CursorHooksInstaller ESM compatibility, updates install command
with improved path resolution, and refreshes built plugin artifacts.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The MCP server (#!/usr/bin/env node) and context generator run under
Node.js, where import.meta.url throws SyntaxError in CJS mode. Only
the worker-service needs the banner since it runs under Bun.
CJS files under Node.js already have __dirname/__filename natively.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add sessionId to summarize.ts warning log for easier triage
- Add APPROVED OVERRIDE annotation to Windows spawn catch block
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. Fix esbuild inlining build-machine __dirname as string literal — use
CJS-compatible runtime banner with require("node:url").fileURLToPath
across worker-service, mcp-server, and context-generator builds.
2. Fix isMainModule check missing .cjs extension and Windows backslash
path normalization.
3. Wrap extractLastMessage in try-catch to prevent infinite Stop hook
feedback loop on malformed transcripts (exit 0 instead of exit 2).
4. Replace heavy SessionEnd hook (Node→Bun→1.7MB CJS→HTTP) with
lightweight inline node -e one-liner (~200ms vs >1s).
5. Add 7 Gemini/OpenRouter error patterns to unrecoverablePatterns
circuit breaker to prevent 77K+ retry loops on expired API keys.
6. Preserve CLAUDE_CODE_OAUTH_TOKEN and CLAUDE_CODE_GIT_BASH_PATH in
sanitizeEnv instead of stripping them with the CLAUDE_CODE_ prefix.
7. Use PowerShell -EncodedCommand for spawnDaemon to fix path quoting
when Windows usernames contain spaces.
Closes#1515, #1495, #1475, #1465, #1500, #1513, #1512, #1450, #1460,
#1486, #1449, #1481, #1451, #1480, #1453, #1445
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Removes installer/ directory (16 files) — fully replaced by src/npx-cli/.
Updates install.sh and installer.js to redirect to npx claude-mem.
Adds npx claude-mem as primary install method in docs and README.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds esbuild steps for npx-cli (57KB, Node.js ESM) and openclaw plugin
(12KB). Creates .npmignore to exclude node_modules and Bun binary from
npm package, reducing pack size from 146MB to 2MB.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replaces the old git-clone installer with a direct npm package copy workflow.
Supports 13 IDE auto-detection targets, runtime delegation to Bun worker,
and pure Node.js install path (no Bun required for installation).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: stop spinner from spinning forever due to orphaned DB messages
The activity spinner never stopped because isAnySessionProcessing() queried
ALL pending/processing messages in the database, including orphaned messages
from dead sessions that no generator would ever process.
Root cause: isAnySessionProcessing() used hasAnyPendingWork() which is a
global DB scan. Changed it to use getTotalQueueDepth() which only checks
sessions in the active in-memory Map.
Additional fixes:
- Add terminateSession() to enforce restart-or-terminate invariant
- Fix 3 zombie paths in .finally() handler that left sessions alive
- Clean up idle sessions from memory on successful completion
- Remove redundant bare isProcessing:true broadcast
- Replace inline require() with proper accessor
- Add 8 regression tests for session termination invariant
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address review findings — idle-timeout race, double broadcast, query amplification
- Move pendingCount check before idle-timeout termination to prevent
abandoning fresh messages that arrive between idle abort and .finally()
- Move broadcastProcessingStatus() inside restart branch only — the else
branch already broadcasts via removeSessionImmediate callback
- Compute queueDepth once in broadcastProcessingStatus() and derive
isProcessing from it, eliminating redundant double iteration
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: strip <system_instruction> tags before database storage
Extends the existing tag-stripping mechanism (used for <private> and
<claude-mem-context>) to also filter Conductor-injected system instructions,
preventing them from being persisted in the observation database.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: also strip <system-instruction> (hyphen variant) before DB storage
Conductor uses both <system_instruction> and <system-instruction> tag
formats. This adds the hyphen variant to the same stripping mechanism.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(openclaw): inject context via system prompt instead of overwriting MEMORY.md
The OpenClaw plugin was overwriting each agent's MEMORY.md with a large
auto-generated observation dump (~12-15KB) on every before_agent_start
and tool_result_persist event. This conflicts with OpenClaw's design
where MEMORY.md is agent-curated long-term memory.
Migrate context injection from file-based (writeFile MEMORY.md) to
OpenClaw's native before_prompt_build hook, which returns context via
appendSystemContext. This keeps MEMORY.md under agent control while
still providing cross-session observation context to the LLM.
Changes:
- Add before_prompt_build hook that returns { appendSystemContext }
- Remove writeFile/MEMORY.md sync from before_agent_start
- Remove MEMORY.md sync from tool_result_persist (observations still recorded)
- Add 60s TTL cache to avoid re-fetching context on every LLM turn
- Add syncMemoryFileExclude config for per-agent opt-out
- Remove dead workspaceDirsBySessionKey tracking map
- Rewrite test suite to verify prompt injection instead of file writes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(ui): align settings defaults with backend and use nullish coalescing
The web UI had two issues causing settings inflation:
1. DEFAULT_SETTINGS in the UI used FULL_COUNT='5' and all token columns
'true', while SettingsDefaultsManager (backend) uses FULL_COUNT='0'
and token columns 'false'. Opening the settings modal and saving
without changes would silently inflate the context.
2. useSettings used || for fallback, which treats '0' and 'false' as
falsy — even when the backend correctly returns these values, the UI
would replace them with inflated defaults. Changed to ?? (nullish
coalescing) so only null/undefined trigger the fallback.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs(openclaw): update integration docs for system prompt injection
Reflect the migration from MEMORY.md file writes to before_prompt_build
hook-based context injection:
- Update architecture diagram and overview to show new hook flow
- Replace "MEMORY.md Live Sync" section with "System Prompt Context Injection"
- Update event lifecycle steps (before_agent_start, tool_result_persist)
- Add before_prompt_build step with TTL cache description
- Document new syncMemoryFileExclude config parameter
- Update session tracking to reflect removed workspaceDirsBySessionKey
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs: fix terminology and update SKILL.md for system prompt injection
Replace "prompt injection" with "context injection" in docs to avoid
confusion with the OWASP security term. Update openclaw/SKILL.md to
reflect the new before_prompt_build hook and remove stale MEMORY.md
references.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Alex Newman <thedotmack@gmail.com>
Alessandro Costa