Critical gaps across all three tools. Tool definitions lack proper input schemas (no JSON Schema visible), descriptions are minimal (10-50 chars, well below 194-char baseline), and parameter documentation is absent. The 'ai_prompt' tool has two parameters but neither has descriptions. 'respond_permission' parameters lack type info and constraints. No output schemas documented. Error handling is absent, no recovery guidance or categorization. Tools appear to be inferred from handler code rather than explicitly registered with full MCP schemas.

• Add explicit JSON Schema for each tool input. Example for 'ai_prompt': {"type": "object", "properties": {"prompt": {"type": "string", "description": "The prompt to send to Claude (1-4000 chars)"}, "wait": {"type": "boolean", "description": "If true, block until response received; if false, return immediately with message ID"}}, "required": ["prompt"]}
• Expand tool descriptions to 100-200 chars. Example: 'ai_prompt' → 'Send a prompt to Claude via Orchestra and optionally wait for the response. Use this to ask questions, generate content, or analyze data. Returns the assistant's response or a message ID if async.'
• Add parameter descriptions and constraints. For 'respond_permission', add enum constraint: decision must be one of 'approve' or 'deny'. Document what happens on each decision.
• Document output schemas. Specify what 'list_features' returns (e.g., array of feature objects with name, description, enabled fields). Specify what 'ai_prompt' returns (text response, token count, model used).
• Add error handling guidance. Example: 'If Claude is unavailable, retry after 30 seconds. If the prompt is rejected for policy reasons, ask the user to rephrase. If the session is invalid, call list_features to refresh state.'