- Skip gate for files under 1,500 bytes — timeline (~370 tokens) costs
more than just reading small files directly
- Deduplicate observations by memory_session_id (one per session)
- Rank by specificity: files_modified > files_read, fewer tagged files > many
- Fetch 40 candidates, dedup/score down to 15 for display
- Reduce default by-file query limit from 30 to 15
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The per-prompt Chroma vector search injection on UserPromptSubmit adds latency
and context noise. Disable by default while we iterate on a more precise
file-context approach. Users can still opt in via CLAUDE_MEM_SEMANTIC_INJECT=true.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
sort -V is a GNU extension not available on macOS/BSD sort. Replaced
with sed 's/^v//' | sort -t. -k1,1n -k2,2n -k3,3n which strips the
'v' prefix, sorts numerically by major.minor.patch, then re-prepends
'v' for the final path. Works on both GNU and BSD sort.
Prepend a PATH export to every hook command that invokes node, so that
/bin/sh can locate the binary when Node.js is managed by nvm or
installed via Homebrew. Falls back gracefully when nvm is not present.
Closes#1598
Observer prompt now explicitly requires XML observation blocks or empty
responses — prose explanations like "Skipping" are discarded. ResponseProcessor
logs a warning when non-XML content is received. Recording focus expanded to
include concrete debugging findings.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The parser correctly strips observation types from concepts arrays when the
LLM ignores the prompt instruction. This is routine data normalization, not
an error — downgrade to debug to reduce log noise.
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>
System reminders (CLAUDE.md contents, deferred tool lists) were being
stored in memory observations. Add system-reminder to the tag stripping
pipeline alongside <private> and <system_instruction>, and extract the
duplicated regex into a shared SYSTEM_REMINDER_REGEX constant.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Instead of using shell:true with spawn(), use cmd.exe as the command
with /c flag to properly execute bun.cmd on Windows.
Without this, spawn() with shell:true fails because cmd.exe doesn't
know how to handle the bun shell script directly.
Fixes: Stop hook "Failed to start Bun: spawn bun ENOENT"
Windows npm installs both bun (shell script) and bun.cmd (batch file).
When spawning bun, cmd.exe cannot execute the shell script directly.
This change makes findBun() return the full path to bun.cmd on Windows.
Fixes: Stop hook "spawn bun ENOENT" on Windows
Windows npm installs bun as a shell script (C:\Users\...\AppData\Roaming\npm\bun),
not a native executable. Without shell:true, spawn() fails with ENOENT
when trying to execute it.
Fixes Stop hook failure: "Failed to start Bun: spawn bun ENOENT"
The fallback path for CLAUDE_PLUGIN_ROOT was pointing to the old
marketplaces install location which no longer exists. Hooks now first
try to find the latest versioned cache directory
(~/.claude/plugins/cache/thedotmack/claude-mem/<version>/) using ls -dt,
with the marketplaces path kept as a final fallback.
This mirrors the self-resolution pattern already used in bun-runner.js
(resolve(__bun_runner_dirname, '..')) but at the shell level, so node
can find bun-runner.js in the first place.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Update POST_SPAWN_WAIT test assertion from 5000 to 15000 to match
the constant change in hook-constants.ts
- Remove redundant readPidFile() from aggressiveStartupCleanup() —
start() writes the new PID before this runs, so it always returns
process.pid (already protected)
- Add waitForReadiness() to the reused-worker path in
ensureWorkerStarted() to prevent concurrent hooks from racing
past a cold-starting worker's initialization guard
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three independent fixes for worker daemon instability:
1. Remove version mismatch auto-restart from ensureWorkerStarted() (#1435).
The marketplace bundle ships with __DEFAULT_PACKAGE_VERSION__ unbaked,
causing BUILT_IN_VERSION to fall back to "development". This creates a
100% reproducible mismatch on every hook call, killing a healthy worker
and often failing to restart. Same pattern across #566, #665, #667,
#669, #689, #1124, #1145 (8+ releases).
2. Add process.ppid and PID-file PID to aggressiveStartupCleanup()
exclusions (#1426). Without this, a newly spawned daemon SIGKILLs
the hook process that spawned it and any already-running worker
the PID file points to.
3. Increase POST_SPAWN_WAIT from 5s to 15s (#1423). The 5s timeout was
sized for Linux (<1s startup) but macOS ARM64 cold starts take 6-8s
with Chroma enabled.
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>
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>
Address CodeRabbit review: add a final health check after the retry
loop so genuine worker startup failures surface as hook errors instead
of being silently masked.
Defense-in-depth for #1505. When the 'start' subcommand forks a daemon,
the parent bun process may be killed by signal (exit > 128). If the
close handler fires, treat this as success since the daemon started fine.
Note: the primary fix is in hooks.json since the SIGKILL often kills
the entire process group before this handler fires.
The worker-start hook's `start` subcommand forks a daemon then SIGKILLs
its own process group, killing bun-runner.js before it can report exit 0.
Since all SessionStart hooks run in parallel, the context hook also fails
because the worker isn't listening yet.
Fix:
- worker-start: continue after the SIGKILL via `;`, poll the worker
health endpoint until ready, then output valid JSON (exit 0)
- context: wait for worker health before attempting to fetch context
Fixes#1505
Tighten platform source persistence so legacy callers cannot silently relabel existing sessions, repair migration 24 when schema_versions drifts from the real schema, and polish the follow-up UI/error-handler review nits.
- only backfill platform_source when it is blank and raise on explicit source conflicts for an existing session
- make migration 24 verify both the sdk_sessions column and its index before treating it as applied
- expose platform_source from the functional session getters and add regression tests for source preservation and schema drift recovery
- add the required APPROVED OVERRIDE annotation for centralized HTTP error translation
- keep mobile source pills on a single horizontal row
Persist platform_source across session creation, transcript ingestion, API query paths, and viewer state so Claude and Codex data can coexist without bleeding into each other.
- add platform-source normalization helpers and persist platform_source in sdk_sessions via migration 24 with backfill and indexing
- thread platformSource through CLI hooks, transcript processing, context generation, pagination, search routes, SSE payloads, and session management
- expose source-aware project catalogs, viewer tabs, context preview selectors, and source badges for observations, prompts, and summaries
- start the transcript watcher from the worker for transcript-based clients and preserve platform source during Codex ingestion
- auto-start the worker from the MCP server for MCP-only clients and tighten stdio-driven cleanup during shutdown
- keep createSDKSession backward compatible with existing custom-title callers while allowing explicit platform source forwarding
GeminiAgent sends the full conversation history with every API call,
causing quadratic token growth per session. A 100-observation session
sends ~30M cumulative input tokens. This ports the proven truncateHistory()
sliding window from OpenRouterAgent to GeminiAgent.
- Add CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES (default: 20) and
CLAUDE_MEM_GEMINI_MAX_TOKENS (default: 100000) settings
- Add truncateHistory() to GeminiAgent using shared estimateTokens()
- Always preserve at least the newest message to avoid empty API requests
- Add settings validation in SettingsRoutes (1-100 messages, 1K-1M tokens)
- Add regression tests for truncation and oversized single-prompt edge case
Co-Authored-By: Claude Opus 4.6 (1M context) <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>
Add a PreToolUse gate that blocks file reads on first attempt when rich
observation history exists, presenting the timeline as feedback. Claude
then decides: use get_observations() (skip read, save tokens) or re-read
(allowed on second attempt).
- FileReadGate: in-memory session-scoped gate with 4h TTL
- POST /api/file-context/gate endpoint in worker
- stderrMessage plumbing in hook-command for exit code 2
- file-context handler uses gate to block/allow reads
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>
- Use getProjectContext(cwd).allProjects for project scoping (same as SessionStart)
- Convert absolute file_path to relative using cwd (observations store relative paths)
- API accepts comma-separated projects param with IN() SQL filter
- Remove basename matching — use full relative path to avoid cross-file collisions
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The handler was passing input.cwd (full absolute path) as the project
filter, but observations store short project names ('san-diego', not
'/Users/.../san-diego'). This caused zero results for every query.
Removing the filter entirely is better: cross-project observations
about the same file are useful for duplicate prevention.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When Claude reads a file, the PreToolUse hook queries for existing
observations about that file and injects the timeline into context
via additionalContext + permissionDecision: allow. This prevents
duplicate observations and saves tokens through active rediscovery.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Alex Newman