The Fabric SQL Assistant server registers 6 tools with explicit schemas and descriptions visible in mcp_server.py. All tools have basic name-verb patterns (configure_, discover_, ask_, get_, execute_, get_) and input schemas with JSON types. However, several definition quality gaps prevent a higher score: (1) Output schemas are completely undocumented, tools return TextContent but the actual data structure (field names, types, pagination) is unknown; (2) Error handling descriptions are absent, tools catch exceptions but don't guide LLM recovery; (3) Some parameter descriptions lack constraint information (e.g., 'sql' parameter has no length limits, SQL injection warnings, or expected result bounds); (4) Tool descriptions lack dependency hints and composition guidance; (5) Parameter relationships are undocumented (e.g., ask_database with use_auto_schema=true depends on discover_schema being called first, but this is not stated). Average tool description length is ~100 chars, within the acceptable range (p10=34, p90=392), but descriptions omit WHEN and WHY to use tools. All 6 tools have input schemas with types present, but 2 tools (get_current_config, discover_schema) have zero or minimal required parameters, which is appropriate but underscore simple behavior. Per-tool scores below reflect individual gaps.

• Document output schemas for all 6 tools. Define the TextContent response structure: list fields that will appear in the 'text' field (e.g., for discover_schema: 'Returns JSON with keys: tables (array of {name, columns (array of {name, type})}), last_updated (ISO8601).').
• Add SQL injection warnings and constraints to execute_sql_query description: 'WARNING: SQL queries are executed as-is. Validate inputs to prevent injection. Max query length: 10000 characters. Only SELECT queries recommended; DDL/DML restricted by default.'
• Enhance ask_database description with dependency: 'Recommended: call discover_schema first to cache table metadata. If use_auto_schema=true (default), schema is auto-discovered on first call (slower). Returns natural-language summary of results; use include_raw_data=true for structured output.'
• Add per-parameter validation rules to descriptions: table_name should state 'Must be a valid table name in the connected database (alphanumeric + underscore, case-sensitive)'; limit_rows should state 'Range: 1-10000; default 100; values >10000 may timeout.'
• Implement structured error responses with recovery hints. Instead of 'Error in execute_sql_query: ...', return JSON with {error_type, message, recovery_action, retryable}. E.g.: {error_type: 'syntax_error', message: 'Invalid SQL: ...', recovery_action: 'Check SQL syntax; call ask_database for natural-language help', retryable: false}.
• Add permission checks. Gate configure_database and execute_sql_query behind authorization. Return permission_denied error if caller lacks 'write:database' or 'execute:sql' scope.