- Updated docker-compose.yml to include PLAYWRIGHT_BROWSERS_PATH and MCP_PUBLIC_URL environment variables.
- Modified k8s-manifests-complete.yaml to add Playwright and MCP configurations in the ConfigMap and deployment spec.
- Adjusted resource limits in k8s manifests for improved performance during crawling.
- Updated Dockerfiles to install Playwright browsers in accessible locations for appuser.
- Added HTTP health check endpoint in mcp_server.py for better monitoring.
- Enhanced MCP API to utilize MCP_PUBLIC_URL for generating client configuration.
- Created MCP_PUBLIC_URL_GUIDE.md for detailed configuration instructions.
- Documented changes and recommendations in K8S_COMPLETE_ADJUSTMENTS.md.
Fixed the health endpoint implementation using the correct FastMCP API:
- Use @mcp.custom_route('/health', methods=['GET']) instead of mcp.create_app()
- Returns 200 when healthy, 503 when starting/degraded
- Provides full health status (api_service, agents_service, uptime)
The previous attempt used mcp.create_app() which doesn't exist in FastMCP 1.12.2.
FastMCP's correct API for custom HTTP endpoints is @mcp.custom_route().
Fixes frontend getting 404 errors on GET /health
The MCP only had health_check as an MCP tool, not as an HTTP endpoint.
Frontend was getting 404 when trying GET /health.
Added FastAPI route /health that returns:
- 200 when healthy with full health status
- 503 when starting/degraded with status message
Fixes frontend MCP status dashboard showing 'unhealthy'
Resolves#420 - Tasks being duplicated instead of updated
Changes:
1. Fixed update_task function signature to use individual optional parameters
- Changed from TypedDict to explicit parameters (title, status, etc.)
- Consistent with update_project and update_document patterns
- Builds update_fields dict internally from provided parameters
2. Updated MCP instructions with correct function names
- Replaced non-existent manage_task with actual functions
- Added complete function signatures for all tools
- Improved workflow documentation with concrete examples
This fixes the issue where AI agents were confused by:
- Wrong function names in instructions (manage_task vs update_task)
- Inconsistent parameter patterns across update functions
- TypedDict magic that wasn't clearly documented
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Critical improvements to MCP server reliability and client experience:
Error Handling:
- Created MCPErrorFormatter for consistent error responses across all tools
- Provides structured errors with type, message, details, and actionable suggestions
- Helps clients (like Claude Code) understand and handle failures gracefully
- Categorizes errors (connection_timeout, validation_error, etc.) for better debugging
Timeout Configuration:
- Centralized timeout config with environment variable support
- Different timeouts for regular operations vs polling operations
- Configurable via MCP_REQUEST_TIMEOUT, MCP_CONNECT_TIMEOUT, etc.
- Prevents indefinite hangs when services are unavailable
Module Registration:
- Distinguishes between ImportError (acceptable) and code errors (must fix)
- SyntaxError/NameError/AttributeError now halt execution immediately
- Prevents broken code from silently failing in production
Polling Safety:
- Fixed project creation polling with exponential backoff
- Handles API unavailability with proper error messages
- Maximum attempts configurable via MCP_MAX_POLLING_ATTEMPTS
Response Normalization:
- Fixed inconsistent response handling in list_tasks
- Validates and normalizes different API response formats
- Clear error messages when response format is unexpected
These changes address critical issues from PR review while maintaining
backward compatibility. All 20 existing tests pass.
Update MCP server to use the new modular tool structure:
- Projects and tasks from existing modules
- Documents and versions from new modules
- Feature management from standalone module
Remove all feature flag logic as separated tools are now default.
- Rename src/mcp to src/mcp_server for clarity
- Update all internal imports to use new path
- Create features/projects directory for modular tool organization
- Add separate, simple project tools (create, list, get, delete, update)
- Keep consolidated tools for backward compatibility (via env var)
- Add USE_SEPARATE_PROJECT_TOOLS env var to toggle between approaches
The new separate tools:
- Solve the async project creation context loss issue
- Provide clearer, single-purpose interfaces
- Remove complex PRP examples for simplicity
- Handle project creation polling automatically
