12 tools with mixed quality. Naming is verb-forward and clear (health, create_audit, list_runs, get_report). Descriptions exist but are terse (10-50 chars), below the 50-200 char LLM-optimized baseline. Input schemas are present for most tools but lack depth: many parameters have no type info visible (e.g., 'profile' in create_audit is typed 'object' with no nested schema). Output schemas are not documented anywhere in the source. Error handling is minimal, no recovery guidance, no categorization of retryable vs fatal errors. Security concern: create_audit accepts inline 'profile' object without validation hints. Composition is reasonable (single responsibility per tool), but audit_events returns SSE stream without documented structure. Overall: functional but below production baseline.

• Expand all descriptions to 50-200 chars. Add WHEN to call, WHAT it returns, and any prerequisites. E.g., 'List all completed audit runs with summary stats. Call this to review past results or fetch a run_id for get_report(). Returns paginated list with run_id, timestamp, status, and issue count.'
• Document output schemas for all tools. Define fields, types, and ranges. E.g., list_runs returns {runs: [{run_id: string, timestamp: ISO8601, status: 'completed'|'failed', issue_count: int}], total: int, next_cursor?: string}.
• Expand create_audit's 'profile' parameter schema. Define required fields (e.g., audit_type, target_files), constraints (e.g., max file count), and provide a minimal example inline or via a discovery tool.
• Add error handling to create_audit and apply_approvals_once. Document 409 Conflict (audit already running), 400 Bad Request (invalid profile), and 500 Internal Error with recovery guidance for each.
• Document audit_events SSE event structure. Define event types (e.g., 'progress', 'error', 'complete'), fields per type, and example payloads so LLMs can parse streaming responses.