Single read-only query tool with basic definition. Tool naming is reasonable (read_only_query is action-oriented), but description and schema have critical gaps. The description is adequate (89 chars, within baseline range), but the schema is severely incomplete: it uses a raw Zod object without proper JSON Schema structure or type declarations visible in the schema definition. Parameter 'sql' lacks detail on format, constraints, or what constitutes a valid query beyond the SELECT prefix check. No output schema is documented, responses are returned as raw JSON strings without type information. Error handling is minimal: only two cases checked (non-SELECT prefix, database error), with generic error messages that don't guide recovery. Resource 'schema' is present but not assessed as a tool per the rubric.

• Convert Zod schema to explicit JSON Schema with 'type': 'object', 'properties' with types, 'required' array, and property descriptions. Example: { type: 'object', properties: { sql: { type: 'string', description: 'SELECT query (max 10000 chars, no UNION/INTERSECT)', minLength: 10, maxLength: 10000 } }, required: ['sql'] }
• Document the output schema explicitly. E.g., 'Returns: { rows: Array<object>, rowCount: number, executionTime: number }' and ensure the actual response matches this contract.
• Add query constraints to the parameter description: 'The SQL SELECT statement (uppercase preferred, max 10000 chars). Must start with SELECT. Allowed: basic WHERE/JOIN/GROUP BY/HAVING/ORDER BY/LIMIT. Disallowed: UNION, INTERSECT, CTE, subqueries deeper than 2 levels, table modification (INSERT/UPDATE/DELETE), system tables.'
• Enhance error responses with actionable guidance. Example: { content: [{ type: 'text', text: 'Invalid query: Only SELECT queries are permitted. If you need to modify data, use a different tool or contact a DBA.' }], isError: true }
• Add result limit enforcement to the tool description and implementation: 'Results are capped at 1000 rows. If you need more rows, use LIMIT and OFFSET to paginate, or export via a dump service.'
• Implement per-item error handling if batch queries are supported later. For a single query tool, ensure the error message identifies the specific issue (syntax error at line X, timeout after 30s, etc.).
• Add tool annotations to the tool registration: annotations: { readOnlyHint: true } to signal to the client that this is a read-only operation.

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

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained