Two tools are explicitly registered via @mcp.tool() decorator in mcp-server/server.py. query_emis has a well-written description (194 chars, within baseline 34-392 range) and clear input schema with description. check_backend_health has minimal description (58 chars) and no input parameters. Both tools lack output schema documentation, the code shows result.get() calls but does not formally document the response structure. Error handling returns structured dicts with 'status' and 'error_message' fields, which is good, but descriptions do not explain error cases or recovery paths. No parameter descriptions beyond the input object itself. LLMs need to know what fields to expect.' Overall definition quality is below average due to incomplete schema coverage and minimal error guidance in tool descriptions.

• Add formal output schema documentation to query_emis and check_backend_health. Use JSON Schema to describe status, timestamp, citation, summary, raw_data, documents fields and their types. Example: {'type': 'object', 'properties': {'status': {'type': 'string', 'enum': ['success', 'error']}, ...}}
• Expand check_backend_health description to 50-100 chars explaining when to call it (e.g., 'before querying EMIS'), what it checks, and what fields the response contains (uptime, version, etc.).
• Classify errors in query_emis responses: mark ConnectionError and Timeout as retryable; HTTPError as user-fixable (may need API key, backend config); other exceptions as fatal. Add error_type field to responses.
• Enhance error messages with recovery guidance. Example: 'Request timed out (retryable). The EMIS portal may be slow. Wait 30s and retry, or try a simpler query.' instead of bare 'Request timed out.'
• Improve query_emis parameter description in the schema. Instead of bare 'Natural language query about EMIS data', add format guidance: 'Natural language question (1-500 chars). Examples: Find recent BBT updates for water treatment, What does EMIS say about waste management legislation? Avoid jargon; phrase as a question or imperative.'
• Consider adding pagination to query_emis if the backend returns many results. Accept optional limit (default 20, max 100) and offset parameters. Document in response whether more results are available.
• Add tool annotations (readOnlyHint: true for both tools). Both are read-only queries, not modifications. This helps agents understand they are safe to retry.

Inferred effective spec: <=2025-11-25.

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