20 tools with mixed quality. Most have basic descriptions (50-150 chars) and partial schemas. Critical gaps: 8 tools lack visible input schemas (learn, export, save-evals, history, neighbourhood, consolidate, contradictions inferred from test files, not src/mcp/tools.ts). Parameter descriptions present but generic. No output schemas documented. Error handling absent. Tool names are verb-based but some are ambiguous (e.g., 'recall' vs 'retrieve', 'check' vs 'status'). Composition issues: overlapping tools (search/recall/neighbourhood all find related content differently). No pagination guidance despite search accepting limit/offset. Destructive tools (unlearn, gc) lack confirmation patterns.

• Add explicit input schemas to learn, export, save-evals, gaps, history, neighbourhood, consolidate, contradictions in src/mcp/tools.ts. Currently inferred from test files only.
• Document output schemas for all tools. Specify returned fields, types, and pagination info (e.g., search returns {results: [{kbid, title, score}], total, hasMore}).
• Rename ambiguous tools: 'check' → 'check_integrity', 'gc' → 'run_garbage_collection', 'status' → 'get_status'. Add 10-20 char descriptions clarifying when to use each.
• Add confirmation pattern to unlearn and gc: require a confirm_id parameter or return a dry-run result first. Prevent accidental data loss.
• Consolidate overlapping tools: merge recall, neighbourhood, contradictions into a single 'find_related' tool with a 'relation_type' enum parameter (semantic, structural, contradictory).