Critical gaps across all dimensions. Tool definitions exist but lack essential structure. No visible input schemas in source code (recall_mcp/tool_surface.py not provided). Descriptions are present but generic (10-50 chars, well below 194-char baseline). No parameter type definitions visible. No output schema documentation. 22 tools with minimal metadata make this unsuitable for production LLM agents. STDIO transport compounds the issue.

• Add explicit JSON Schema input definitions for all 22 tools. Include type, description, enum constraints, and min/max for every parameter. Example: { 'query': { 'type': 'string', 'description': 'Search query (1-500 chars)', 'minLength': 1, 'maxLength': 500 } }
• Expand tool descriptions to 100-250 characters. Answer: What does it do? When should the LLM call it instead of similar tools? What does it return? Example: 'Search the memory corpus for facts matching a query. Use this for general discovery; use recall_reasoning_query for multi-step reasoning. Returns up to 20 results with confidence scores and provenance.'
• Document output schemas for all tools. Specify returned fields, types, and whether results are paginated. Include IDs and references needed by downstream tools (e.g., if recall_search returns result_id, recall_evidence must accept result_id).
• Rename ambiguous tools: recall_reasoning_query → recall_query_with_reasoning, recall_reasoning_projection → recall_project_reasoning_outcome, recall_reasoning_proposals → recall_generate_reasoning_proposals. Make distinctions obvious from names alone.
• Add permission gates and audit logging to destructive tools (recall_forget, recall_apply_fact, recall_calibration_publish). Implement dry-run or confirmation patterns. Document required scopes (e.g., 'write:memory', 'delete:memory').