38 tools with inconsistent quality. Most have basic descriptions (50-150 chars) and input schemas, but lack depth, error guidance, and output documentation. Naming is verb-forward but often generic (e.g., 'ask_local_tool', 'ask_remote_tool'). No tool descriptions explain WHEN to use them vs. alternatives, what they return, or how to handle failures. Parameters lack validation constraints (enums, ranges, formats). No evidence of output schemas or pagination support. Security-sensitive tools (ab_start, cap_ab_start, event_publish) lack permission gates or audit trail documentation. Composition is poor: multiple tools do similar things (consult vs. cross_examine, ask_local_tool vs. ask_remote_tool, ab_* vs. cap_ab_*) without clear differentiation.

• Consolidate overlapping tools: merge ask_local_tool + ask_remote_tool into a single 'query_surgeon' tool with a 'surgeon' enum parameter (neurologist|cardiologist). Reduce cognitive load and eliminate LLM disambiguation.
• Add output schemas to all tools. Document return types in tool descriptions: 'Returns: {stream_id: string, patterns: string[]}' for event_subscribe. Enable LLMs to chain calls and extract required fields.
• Expand descriptions to 100-200 chars with WHEN/WHY context. Example: 'Scan code for complexity risk vectors (cyclomatic complexity, nesting depth, function size). Use before refactoring to identify high-risk areas. Returns: {risks: [{type, location, severity}]}'.
• Add enum constraints and validation rules to parameters. 'depth' should be enum ['full', 'quick']; 'mode' should be enum ['single', 'iterative', 'continuous']; 'verdict' should be enum ['control_wins', 'treatment_wins', 'inconclusive']. Document in descriptions.
• Document permission requirements and audit trails for WRITE tools. Add to descriptions: 'Requires: write:experiments scope. Logged to audit trail with user_id, timestamp, test_id, verdict.' Implement server-side permission checks before execution.