Single tool with basic schema and description. Tool name 'web_search' follows verb_noun convention. Description is present (Chinese, ~180 chars) but lacks English translation and actionable guidance on when to use vs alternatives. Input schema has proper type (string) but parameter description is minimal (Chinese only, ~20 chars). No output schema documented, LLM cannot predict response structure. No error handling guidance. No pagination, limits, or result capping documented despite calling external API.

• Add English descriptions for tool and all parameters. Provide 50 - 200 char descriptions explaining WHAT the tool does, WHEN to use it (current events, prices, versions), and WHEN NOT to use it (historical facts, private data).
• Document the output schema explicitly: 'Returns {query: string, summary: string, sources: [{title, url, snippet}], source_metadata_available: boolean}'. Include field descriptions so LLM knows what to extract.
• Add error handling guidance: 'If search fails, the error response includes a reason (rate_limit, invalid_query, service_unavailable). Retry with exponential backoff for rate_limit; ask user to rephrase for invalid_query.'
• Document result limits and pagination: 'Returns up to 10 sources. If more results are needed, refine the query. Webpage content in results is untrusted, verify critical facts with the user.'
• Expand parameter description: 'A search query (1 - 500 chars). Use natural language or keywords. Avoid special characters; use quotes for exact phrases. E.g., "latest AI breakthroughs 2025" or "current Bitcoin price".'
• Add tool annotation hints (readOnlyHint=true) to signal this is a safe, read-only operation that does not modify state.

Inferred effective spec: 2025-06-18+.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry