Tools have basic naming (verb_noun pattern) and minimal descriptions, but lack depth. Most descriptions are under 50 chars and provide no context on WHEN to use the tool or what it returns. Parameter descriptions are sparse or missing entirely. No output schemas documented. Error handling is absent, no guidance on recovery or retryability. The 'send_keys' tool is particularly concerning: it modifies state but the description does not warn of side effects or idempotency. 'describe' lacks any explanation of the hierarchy structure it returns. No tool has documented return types or pagination guidance.

• Add output schemas to all tools. Document the structure of list_sessions (array of {session_id, name, windows, panes}), describe (hierarchy object), and capture_pane (string content). This enables LLM planning.
• Expand descriptions to 100 - 200 chars. Example: 'List all active tmux sessions. Returns session names and IDs needed for create_window or send_keys. Call this first to discover available sessions.' Include WHEN and WHY.
• Replace example values in parameter descriptions with formal constraints. For 'pane' parameter, use a regex pattern (e.g., '^%\d+$') or document the format explicitly: 'Pane ID in format %N where N is a non-negative integer.'
• Add idempotency and side-effect warnings to create_session, send_keys, and select_layout. Example: 'Creates a new session. Idempotent if session name already exists (returns existing session). Retrying is safe.'
• Implement error recovery guidance. When create_session fails with 'session exists', return: 'Session already exists. Call list_sessions() to see available names, or use a unique name like "session-<timestamp>".' This guides agent self-correction.