CAG's tool definitions exhibit significant quality gaps across naming, descriptions, and schemas. While 8 tools are explicitly registered with basic descriptions and parameter schemas, the implementation has critical issues: (1) Tool naming lacks clear verb-noun patterns, 'cag_agent_task_tool' is a multi-action dispatcher that violates single-responsibility principle; (2) Descriptions are generic and lack actionable context for LLM selection, e.g., 'Create a background or foreground task' does not explain WHEN to use foreground vs background or what the implications are; (3) Parameter schemas are present but incomplete, missing enums for 'mode' (foreground/background) and 'agent_name', no range/format constraints on strings; (4) Output schemas are not documented anywhere in the provided source, LLMs cannot infer what fields task creation returns; (5) Error handling is not evident in the source, no recovery guidance, categorization, or actionable error messages visible. The tool definitions are skeletal and would force LLMs to guess at behavior.

• DELETE cag_agent_task_tool (tool 8), it is redundant with tools 1 - 7 and violates single-responsibility. Tools 1 - 7 already provide dedicated list, get, result, wait, wait_any, and cancel operations.
• For cag_agent_create_task: Declare agent_name and mode as formal enums in the input schema, not just description. Add enum: ['claude', 'gemini', 'codex', 'cursor', 'antigravity', 'acp', 'pi'] for agent_name; enum: ['foreground', 'background'] for mode.
• Document output schema for cag_agent_create_task. Specify: returns { task_id: string, status: string, agent_name: string, mode: string, created_at: ISO8601_string }. This enables downstream tool chaining (e.g., agent knows to pass task_id to cag_agent_task_wait).
• Expand cag_agent_create_task description from current ~50 chars to 150 - 200 chars. Include: 'Creates an async task to run a specified AI agent with a prompt. Use foreground for interactive responses (blocks until completion); use background for long-running tasks (returns task_id immediately). Returns task_id for polling or waiting.'
• Add pagination to cag_agent_task_list: introduce limit (default 20, max 100) and offset parameters. Return { tasks: [...], total_count: int, has_more: boolean }. Update description: 'Lists up to <limit> non-expired tasks (paginated by creation time). Use limit and offset for large result sets.'
• For cag_agent_task_result: clarify in description that task must be in terminal state (completed or failed). Add error guidance: 'If task is still running, use cag_agent_task_wait first.'