Six tools with complete input schemas and descriptions. Naming follows verb_noun pattern (list_, search_, get_, execute_). Descriptions are clear and action-oriented (avg ~80 chars), explaining WHAT and WHEN. All parameters have types and descriptions. Output schemas are documented via Pydantic contracts (DatasetInfo, QueryResult, SchemaResult, RetrievedContext). No critical security issues, all tools are read-only with hard limits (500-row cap on execute_sql, 50-value limit on get_distinct_values). Gaps: descriptions could be more LLM-optimized (10-200 char sweet spot); no error recovery guidance in descriptions; no tool annotations (readOnlyHint, idempotentHint); parameter constraints (e.g., top_k range, limit bounds) not formally declared in schema.

• Add tool annotations to FastMCP server: @server.tool(readOnlyHint=True, idempotentHint=True) for all six tools. Clients will infer safety properties without extra documentation.
• Declare parameter bounds in schema: add minValue/maxValue to top_k (1 - 20), limit (1 - 50), max_rows (1 - 500). Update descriptions to reference these bounds: 'Maximum number of distinct values to return (1 - 50, default 5)'.
• Enhance error descriptions with recovery guidance. E.g., execute_sql: 'Execute one validated read-only SQLite query with a hard 500-row cap. If dataset_id is invalid, call list_datasets() first. If SQL is malformed, sqlglot will reject it, check syntax and retry.'
• Document output schemas in tool descriptions. E.g., search_context: 'Returns RetrievedContext with fields: ddl (string), business_evidence (list), question_sql_examples (list).' Or link to a schema reference.
• Clarify search parameter semantics in get_distinct_values: 'Optional substring filter for column values (case-insensitive). E.g., search="john" returns all values containing "john".' Specify exact matching behavior.

Inferred effective spec: 2026-07-28+.

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