Four well-named tools with clear, domain-specific descriptions (avg 180 chars). All tools have complete input schemas with type definitions and parameter descriptions. Output structure is documented via code (brief() function). Naming follows verb_noun pattern (search_, get_, list_). However, output schemas are not formally declared in tool registration, only visible in implementation. Error handling returns structured JSON but lacks recovery guidance. No tool annotations (readOnlyHint, idempotentHint). Parameters are well-constrained (enums, min/max bounds). Japanese descriptions are precise and actionable.

• Add outputSchema to each tool registration (e.g., { type: 'object', properties: { subsidies: { type: 'array', items: { type: 'object', properties: { id, title, issuer, ... } } }, total_matched: { type: 'integer' } } }). This enables LLM planning and downstream tool chaining.
• Enhance error messages with recovery hints. E.g., fail(new Error('ID 999 not found')) → fail(new Error('ID 999 not found. Try search_subsidies() to find valid IDs.'))
• Add readOnlyHint: true to all four tools in registerTool() options. This signals to agents that results are safe to cache and calls are idempotent.
• Implement cursor-based pagination for search_subsidies and list_recurring_schedule. Add 'next_cursor' to response and accept 'cursor' parameter. Allows agents to fetch all results without hitting the 100-item cap.
• Move DISCLAIMER to tool descriptions or return it once per session, not per call. Reduces token waste and improves signal-to-noise ratio in LLM context.