ANAMNE has 21 well-named tools with consistent verb_noun patterns (ask_why, search_decisions, remember, etc.). All tools have descriptions (avg ~120 chars, within baseline 34-392). Input schemas are present and typed. However, parameter descriptions are sparse or missing for many tools, e.g., 'limit' params lack range constraints; 'tags' arrays lack item descriptions. Output schemas are undocumented (no return type specs visible). Error handling is minimal, ask_why catches exceptions but returns a generic message rather than actionable recovery guidance. No tool declares permissions, audit trails, or idempotency guarantees. Composition is sound (each tool does one thing), but chaining metadata is incomplete.

• Add explicit return type schemas to all tools. Document field names, types, and whether results are paginated. Example: search_facts returns [{id: string, content: string, tags: [string], created_at: ISO8601, activation_score: float}].
• Enhance parameter descriptions with constraints. For 'limit' params, specify range (e.g., '1 - 100, default 10'). For 'threshold', add typical values ('0.6 recommended for grouping similar facts'). For 'tags', clarify item type ('array of lowercase strings, max 20 chars each').
• Implement tool annotations. Mark read-only tools (search_*, list_*, get_*, ask_why) with readOnlyHint=true. Mark destructive tools (forget_fact, consolidate_facts) with destructiveHint=true. Mark idempotent tools (remember, pin_fact) with idempotentHint=true.
• Add dry-run support to consolidate_facts and confirmation step to forget_fact. Return a merge plan before executing; require explicit confirmation for destructive operations.
• Document error responses with recovery guidance. For ask_why, return structured error: {error: 'api_key_missing', message: 'Set ANTHROPIC_API_KEY or GEMINI_API_KEY', recovery: 'Run: anamne doctor'}. Classify errors as retryable, user-fixable, or fatal.