The server defines 3 tools with explicit schemas and descriptions. All tools have Zod-based input schemas with type definitions and descriptions. However, there are significant gaps in error handling guidance, output schema documentation, and parameter validation. Naming follows verb_noun convention (agent-query, direct-query, clear-memory), which is correct. Descriptions exist for all tools but are relatively brief (106-149 chars) and lack recovery guidance for failures. Parameters are typed (string, number, uuid) with descriptions, but lack validation rules, constraints, and enum declarations where applicable. Output schemas are not formally documented, responses are JSON-stringified text blocks without structured type declarations. The server lacks actionable error messages and does not follow pattern:recovery-guide or pattern:confirmation-request for the destructive clear-memory tool.

• Declare explicit output schemas for all tools. Instead of returning JSON-stringified text, return structured objects with typed fields. E.g., agent-query should return {answer: string, toolInteractions: array, sessionId: string}.
• Add a confirmation step to clear-memory. Implement a pattern like 'Do you want to clear session {sessionId}? Reply with confirm_clear_<sessionId>' before executing the destructive operation.
• Enhance error messages with recovery guidance. E.g., if agent-query fails with 'Agent service unavailable', return: 'Agent query failed: service unavailable. Try direct-query instead, or check server logs.'
• Document parameter constraints in descriptions. For direct-query limit, state: 'Maximum number of results to return (1-100, default 5)'. For query, state: 'Search query (1-500 characters)'.
• Add pagination support to direct-query. Return {results: array, total: number, next_cursor?: string} so LLMs can handle large result sets without context explosion.
• Clarify tool selection in descriptions. agent-query description: 'Use this to synthesize information and follow multi-turn reasoning. Slower but more accurate for complex questions.' direct-query description: 'Use this for fast, direct similarity search without multi-turn reasoning.'
• Add tool annotations via MCP extension. Annotate clear-memory with destructiveHint=true and idempotentHint=false to help MCP clients warn before invocation.