- Updated `context-inject.sh` to refresh context before prompt submission and ensure worker is running.
- Added functionality to register and unregister projects for automatic context updates in `worker-service.ts`.
- Implemented methods to read and write the cursor project registry, allowing for better management of installed hooks.
- Integrated context updates into the `GeminiAgent`, `OpenRouterAgent`, and `SDKAgent` to ensure the latest context is available during sessions.
This update improves the integration of Claude-Mem with Cursor, ensuring that context is consistently updated and accessible across sessions.
- Introduced functionality for installing, uninstalling, and checking the status of Cursor hooks.
- Added a new command structure for managing hooks with detailed usage instructions.
- Implemented a method to locate the cursor-hooks directory across different environments.
- Updated build-hooks script to inform users about the location of Cursor hooks.
This enhancement streamlines the integration of Claude-Mem with Cursor, improving user experience and accessibility of hooks.
When the plugin updates but the worker was already running on the old version,
hooks would fail with 400 errors because the new hook scripts tried to call
APIs that don't exist in the old worker.
Changes:
- /api/version now returns BUILT_IN_VERSION (compiled at build time) instead
of reading from disk at runtime
- worker-service start command now checks for version mismatch and
auto-restarts if the running worker version differs from plugin version
- Downgraded hook version mismatch warning to debug logging (now handled
by auto-restart)
Fixes#484🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit simplifies worker startup coordination and addresses Windows-specific issues:
**Lock Removal**:
- Removed entire file-based locking system (~100 lines)
- Replaced with health-check-first approach
- Port binding provides natural mutual exclusion - multiple spawns fail cleanly
**Windows Stability**:
- Removed all AbortSignal.timeout() calls to reduce Bun libuv assertion errors
- Added 500ms shutdown delays on Windows to prevent zombie ports
- Worker service has its own timeouts, so client-side timeouts are redundant
**Package.json Updates**:
- Updated worker scripts to use worker-service.cjs directly
- Removed references to deleted worker-cli.js and worker-wrapper.cjs
**Key Changes**:
- src/services/worker-service.ts: Lock removal, shutdown delays, simplified start logic
- src/hooks/*.ts: Removed AbortSignal.timeout from all HTTP requests
- src/shared/worker-utils.ts: Removed AbortSignal.timeout from health checks
- package.json: Updated worker:* scripts
Resolves startup hangs, reduces assertion errors, and prevents zombie port issues.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Addresses review feedback to register signal handlers earlier in the
lifecycle. Previously handlers were registered in start() after HTTP
server initialization, leaving a vulnerability window.
Benefits:
- Signal handlers now active immediately after WorkerService construction
- Protects against signals received during initialization (DB setup,
HTTP server binding, background initialization)
- Prevents orphaned chroma-mcp processes even if killed during startup
- shutdown() method is defensive and safe to call at any stage
This closes the gap where a SIGTERM/SIGINT during initializeBackground()
could leave chroma-mcp subprocess running without cleanup.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Addresses code review feedback from PR #489:
1. Moved PID file cleanup into shutdown() method to ensure it's always
cleaned up regardless of how shutdown is triggered
2. Removed duplicate signal handlers in main() function that were
redundant with the handlers in start() method
This eliminates the race condition where both sets of handlers could
trigger, and ensures consistent PID file cleanup behavior.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
- Add BEGIN/COMMIT/ROLLBACK to session ID column rename migration
- Update claude-desktop.mdx to reflect streamlined MCP tools
- Remove obsolete skill zip download instructions
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Removed unused tool endpoints and schemas to streamline the codebase.
- Consolidated tool descriptions to emphasize a 3-layer workflow for memory search.
- Updated input schemas and handler functions for clarity and efficiency.
- Enhanced documentation within tool descriptions for better user guidance.
- Updated SQL queries in cleanup-duplicates.ts and context-generator.ts to use memory_session_id.
- Modified interfaces in context-generator.ts to reflect the new session ID naming.
- Implemented a repair migration in SessionStore.ts to rename columns in existing tables.
- Adjusted FormattingService.ts and SDKAgent.ts to utilize memory_session_id for session handling.
- Ensured SearchManager.ts retrieves summaries and observations using the updated memory_session_id.
This commit fixes the session ID confusion identified in PR #475:
PROBLEM:
- Using contentSessionId (user's Claude Code session) for SDK resume was wrong
- Memory agent conversation should persist across the entire user session
- Each SDK call was starting fresh, losing memory agent continuity
SOLUTION:
1. Semantic Renaming (clarity):
- claudeSessionId → contentSessionId (user's observed session)
- sdkSessionId → memorySessionId (memory agent's session for resume)
- Database migration 17 renames columns accordingly
2. Memory Session ID Capture:
- SDKAgent captures session_id from first SDK message
- Persists to database via updateMemorySessionId()
- SessionManager loads memorySessionId on session init
3. Resume Logic Fixed:
- Only resume if memorySessionId captured from previous interaction
- Enables memory agent continuity across user prompts
Files changed: 33 (types, database, agents, hooks, routes)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Implement auto-recovery of orphaned queues on startup in WorkerService.
- Introduce startSessionWithAutoRestart method for continuous processing of pending work.
- Modify SDKAgent to prevent session deletion during processing to avoid race conditions.
- Update SessionManager to allow continued processing after yielding summaries.
- Add logic in SessionRoutes to mark processing messages as failed upon generator errors.
- Create detailed documentation for the queue system logic, including recovery mechanisms and potential issues.
- Implement retry logic for acquiring file lock with a maximum of 3 attempts.
- Improve error handling for ENOENT errors by ensuring the directory exists before retrying.
- Update context injection handler to delegate to SearchRoutes, reducing code duplication and preventing "headers already sent" errors.
- Add checks for headersSent in error responses to avoid sending multiple responses.
- Log warnings when the port does not free up after shutdown, and handle forced shutdown scenarios more gracefully.
- Modified the logger to check if the data is an instance of Error.
- If it is an Error, the logger now formats the output to include the message and stack trace in debug mode, or just the message otherwise.
- Retained the existing behavior for other object types in debug mode.
- Removed the linger timeout mechanism to streamline the waiting process for new messages.
- Updated the message handling logic to use a single event listener for new messages.
- Improved abort handling by ensuring the session exits cleanly when aborted.
- Changed the order of operations in session initialization to first create/get the SDK session with the original prompt.
- Introduced a new step to clean the prompt of privacy tags after determining the session ID.
- Updated logging to reflect the new flow and ensure clarity on session creation and prompt number calculation.
The bugfix/session-continuity branch introduced a regression that broke
the privacy fix from PR #463 (commit 63fd158). Privacy tags must be
stripped BEFORE creating the session, not after.
CORRECT order:
1. Strip privacy tags
2. Create session with cleaned prompt
3. Get prompt number
BROKEN order (what was on main):
1. Create session with RAW prompt (stores private content!)
2. Get prompt number
3. Strip privacy tags (too late)
This commit restores the correct order from commit 63fd158.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
- Replaced console.log and console.error statements with logger.info and logger.error in new-hook.ts, SDKAgent.ts, SessionManager.ts, and SessionRoutes.ts for consistent logging.
- Introduced log file creation and management in logger.ts, ensuring logs are saved to a file with a date-based naming convention.
- Enhanced error handling in logger to prevent crashes if log file operations fail.
Added comprehensive diagnostic logging to trace session ID and prompt number flow through the entire system. This is Phase 1 of the session continuity regression fix.
Changes:
- Added logging in src/hooks/new-hook.ts (4 log points)
- Added logging in src/services/worker/http/routes/SessionRoutes.ts (4 log points)
- Added logging in src/services/worker/SessionManager.ts (4 log points)
- Added logging in src/services/worker/SDKAgent.ts (2 log points)
The logging will help identify where the session ID propagation breaks and whether prompt numbers are being calculated correctly.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
- Updated SDKAgent to include session.claudeSessionId in the options for resuming sessions.
- Added comprehensive logging across multiple files to trace session ID and prompt number flow, aiding in diagnosing session continuity issues.
- Introduced a detailed plan for addressing session continuity regression, outlining phases for logging, testing, and implementing fixes.
- Added functions for acquiring and releasing locks using a lock file.
- Implemented cleanup of stale locks from crashed processes.
- Modified 'start', 'stop', and 'restart' commands to use the locking mechanism.
- Ensured proper handling of concurrent operations and improved logging.
- Added logging for empty responses in GeminiAgent and OpenRouterAgent to track potential session context issues.
- Refactored settings file path usage in OpenRouterAgent to use a constant for better maintainability.
- Improved error handling in SessionRoutes to log generator failures with detailed context.
- Implemented JSON parsing error handling in SettingsRoutes to manage corrupted settings files gracefully.
- Added validation for CLAUDE_MEM_OPENROUTER_MAX_CONTEXT_MESSAGES, CLAUDE_MEM_OPENROUTER_MAX_TOKENS, and CLAUDE_MEM_OPENROUTER_SITE_URL in SettingsRoutes to ensure valid configuration.
Settings persistence was broken because 7 setting keys were missing from
the settingKeys array in SettingsRoutes.ts handleUpdateSettings():
- CLAUDE_MEM_GEMINI_RATE_LIMITING_ENABLED
- CLAUDE_MEM_OPENROUTER_API_KEY
- CLAUDE_MEM_OPENROUTER_MODEL
- CLAUDE_MEM_OPENROUTER_SITE_URL
- CLAUDE_MEM_OPENROUTER_APP_NAME
- CLAUDE_MEM_OPENROUTER_MAX_CONTEXT_MESSAGES
- CLAUDE_MEM_OPENROUTER_MAX_TOKENS
Phase 1/5 of PR #448 fix plan.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add PID validation to restart case (matches start case)
- Wrap forceKillProcess() in try/catch for graceful shutdown
- Wrap getChildProcesses() in try/catch for Windows failures
- Add logging to readPidFile(), removePidFile(), httpShutdown()
Fixes critical issues found in PR #458 review.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The waitForHealth function now checks /api/readiness which returns 503
until background initialization completes, rather than just checking if
the port is in use. This ensures callers wait for full worker readiness.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Wrap taskkill call in try/catch so one process failing to kill doesn't
abort cleanup of remaining processes.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace execAsync kill command with individual process.kill calls wrapped
in try/catch to gracefully handle processes that have already exited.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements critical safeguards for OpenRouter integration:
- Context truncation: sliding window keeps max 20 messages (configurable)
- Token limits: hard cap at 100k estimated tokens per request
- Cost tracking: logs token usage and estimated cost per API call
- High-usage warnings: alerts when single request exceeds 50k tokens
New settings:
- CLAUDE_MEM_OPENROUTER_MAX_CONTEXT_MESSAGES (default: 20)
- CLAUDE_MEM_OPENROUTER_MAX_TOKENS (default: 100000)
Prevents exponential context growth that caused $300 runaway charges
in initial testing. Context now automatically truncates to most recent
messages within token budget, with detailed logging for monitoring.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Added support for OpenRouter as an alternative LLM provider with new settings for API key, model selection, and app metadata configuration.
Enhanced context generation with improved settings management and updated worker service APIs.
Includes UI updates for context settings and new observation type configurations.
- Capture original timestamps for messages processed from the backlog to ensure accurate logging.
- Update processGeminiResponse to accept an original timestamp parameter, allowing for correct observation storage times.
- Modify observation and summary processing to utilize the original timestamp when available, improving data integrity.
- Removed billingEnabled setting and replaced it with rateLimitingEnabled in GeminiAgent.
- Updated enforceRateLimitForModel function to skip rate limiting based on rateLimitingEnabled.
- Adjusted getGeminiConfig to retrieve rateLimitingEnabled from settings.
- Changed settings management to reflect the new rate limiting logic in SettingsDefaultsManager and UI components.
- Updated ContextSettingsModal to toggle rate limiting instead of billing.
- Ensured default settings reflect the new rate limiting behavior for free tier users.
- Added new Gemini model types: 'gemini-2.5-pro', 'gemini-2.0-flash', and 'gemini-2.0-flash-lite'.
- Updated RPM limits for existing and new models.
- Enhanced model selection logic to validate configured model against available options, with fallback to a default model.
- Changed Gemini model types to 'gemini-2.5-flash-lite', 'gemini-2.5-flash', and 'gemini-3-flash'.
- Introduced RPM limits for free tier models with a maximum of 10 RPM for 'gemini-2.5-flash-lite' and 5 RPM for the others.
- Added rate limiting enforcement in the GeminiAgent class, which waits based on the model's RPM limit.
- Updated getGeminiConfig to include billingEnabled setting, allowing users to skip rate limiting if billing is enabled.
- Modified ContextSettingsModal to reflect new model options and added a toggle for enabling billing.
- Updated default settings to use the new model and billing configuration.
Alex Newman