CoDRAG plugin defines 5 tools with explicit schemas and descriptions. All tools have input schemas with typed properties and descriptions. However, output schemas are undocumented (tools return generic ToolResult with 'content' or 'error' fields), parameter descriptions lack actionable constraints (e.g., no enum for 'category' in observe tool, no range for 'k' in search), and error handling is minimal (catch-all error returns). Tool names are verb-noun style (context, search, impact, audit, observe) but lack clarity on when to use each vs. alternatives. The 'observe' tool mixes concerns (save note + categorize + link file). Descriptions are adequate (50-150 chars) but lack dependency hints and prerequisites.

• Document output schemas for each tool. E.g., search returns {results: [{file: string, score: number, snippet: string}], total: number, next_cursor?: string}. Include all fields the LLM needs to chain calls.
• Add enum constraints to categorical parameters: audit.categories should list valid options (e.g., ['tech-debt', 'architecture', 'performance', 'security']); observe.category should enforce ['note', 'decision', 'bug', 'pattern', 'assumption'].
• Add bounds to numeric parameters: search.k should be 1 - 50 (default 5); add pagination support (offset/limit) to audit and context if they return large result sets.
• Enhance error messages with recovery guidance. E.g., 'Project not found. Ensure daemon is running at {daemon_url} and project_id is set correctly. Call context() to list available projects.' instead of raw daemon errors.
• Split 'observe' tool into separate tools: save_observation (content, file_path) and tag_observation (observation_id, category) to follow single-responsibility principle and enable flexible composition.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry