Six read-only tools with complete input schemas and reasonable descriptions. All tools follow verb_noun naming (get_*, list_*, query_*, export_*). Descriptions are 50-200 chars, meeting baseline. However, output schemas are undocumented, LLMs cannot plan downstream calls or extract fields reliably. Parameter descriptions are present but minimal (10-30 chars). No error handling guidance, no recovery paths. Tools are well-composed (single responsibility) and idempotent, but lack the depth expected of production tools.

• Document output schemas for all six tools. Example for get_recent_failures: 'Returns list of {task_id: string, error_code: string, error_message: string, ended_at: ISO8601}'. This enables LLM to extract task_id for chaining to get_task_timeline.
• Expand parameter descriptions to 50-200 chars with constraints. E.g., 'hours: Number of hours to look back (1-720, wall-clock time from now). Partial hours are included.'
• Add error recovery guidance. For get_task_timeline: 'Returns null if task_id is unknown. Call list_active_agents() to discover valid task IDs, or query_failure_patterns() to find recent failures.'
• Unify response structure between list_active_agents and get_agent_activity. Both should return agent_id, message_count, task_count, failure_rate so LLM can choose based on time window, not data availability.
• Add pagination to get_recent_failures and query_failure_patterns. Accept offset/limit, return total_count. Cap default limit at 20 and document: 'Results are ordered by recency. Use offset to paginate large result sets.'