Two well-named search tools with excellent descriptions and comprehensive parameter schemas. Tool names (search_project_history, search_global_history) follow verb_noun convention and clearly distinguish scope. Descriptions are 200+ characters, context-rich, and explain WHAT, WHEN, and WHY. All 8 input parameters are typed and documented. However, output schemas are not explicitly documented in the visible code, only mentioned in docstrings as 'returns dict'. The schema completeness is inferred from the Pydantic model_dump() call but the actual structure is not visible. Error handling is minimal, no recovery guidance, no invalid input examples, and no edge case documentation (e.g., what happens if query is empty, or threshold is out of bounds). Both tools are READ_ONLY with appropriate risk marking, but lack the defensive input validation patterns common in production tools.

• Document the complete output schema with field types and structure. Example: {matched_messages: [{text: string, score: float, timestamp: ISO8601, session_id: string, context: {before: Message[], after: Message[]}}], total_matches: int, offset: int, has_more: bool, project: string|null}
• Add error recovery guidance: 'If threshold is too high (no results), try lowering it to 0.1. If a date is invalid, use ISO 8601 format: 2025-01-15 or 2025-01-15T10:30:00Z.'
• Add parameter constraints in descriptions: threshold: 'Float between 0 and 1, default 0.2. Higher values (0.8+) return only high-confidence matches; lower values (0.1-0.3) are permissive.'; max_results: 'Max 100, default 10. Large result sets (50+) slow semantic search, prefer narrow queries with low threshold.'; offset: 'Zero-indexed, must be >= 0. Use with max_results for pagination.'
• Document why defaults exist: 'Default context_before_after=3 provides conversational context without overwhelming results. Adjust for longer multi-turn discussions (context_before_after=5-10).'
• Clarify pagination strategy: 'Use offset + max_results to paginate. Call again with offset=prev_offset+max_results until has_more=false. Results are deterministic if query and threshold are unchanged.'
• Add idempotent/readonly hints (tool annotations) if the MCP framework supports them to signal LLMs these are safe to retry.
• Include an example call in the description for discoverability: 'search_project_history(query="authentication bug", threshold=0.5) searches project history for security discussions.'