Critical gaps across all dimensions. Tool descriptions are verbose (200-400+ chars) but lack actionable structure for LLM selection. Input schemas are present but lack type definitions for most parameters, all four tools define parameters as strings without validation constraints, enums, or ranges. No output schemas documented. Error handling is absent. Tool naming is descriptive but some combine multiple concerns (q_cockpit bundles elicitation + observability + HTTP server). No evidence of idempotency, permission gates, or secret injection patterns. This server is not production-ready for agent integration.

• Add enum constraints to 'mode' parameter in validate_audit: enum=['legacy', 'manifest']. Add enum to 'level' in all tools: enum=[0, 1, 2]. Document valid ranges for port in q_cockpit: minimum=1024, maximum=65535.
• Rewrite descriptions as: '[WHAT] Generates audit compliance report. [WHEN] Call after running audit scans; use validate_audit first to check for gaps. [RETURNS] Structured report object with fields: audit_report_path, plan_path, issues_path, evolution_path.' Apply to all four tools.
• Document output schemas for each tool. Example for generate_report: {type: 'object', properties: {audit_report_path: {type: 'string'}, plan_de_mejora_path: {type: 'string'}, remediation_issues_path: {type: 'string'}, propuesta_evolutiva_path: {type: 'string'}}}.
• Split q_cockpit into three tools: (1) start_elicitation_session(project_root, topic, doc_path) → session_id; (2) query_mission_cockpit(session_id, query_type) → results; (3) inspect_artifacts(session_id, artifact_type) → artifact_data. Remove HTTP server management from MCP tool layer.
• Add error handling: validate level ∈ [0,1,2], check manifest_path exists before reading, verify output_path is writable, return {error: 'Invalid level: got 5, must be 0|1|2', retryable: false} for invalid inputs. Implement pattern:recovery-guide.