Critical gaps across all three tools. Tool definitions are inferred from test mocks rather than explicit MCP registration. No input schemas visible in source code. Descriptions are present but generic and lack LLM-optimization guidance. The 'initialize' tool violates MCP spec (initialize is a protocol message, not a tool). Tools lack parameter descriptions, error handling guidance, and output schema documentation. No evidence of permission gates, audit logging, or secret injection patterns despite claims of 'access control and audit logging'.

• Define explicit MCP tool schemas in src/core/mcp/router.ts with full JSON Schema: type, properties, required, descriptions for every parameter. Use TypeScript interfaces + @modelcontextprotocol/sdk tool registration.
• Remove 'initialize' tool. MCP protocol handles capabilities negotiation via the handshake; exposing it as a tool is redundant and confusing.
• Add parameter descriptions: 'allowedDrivers' → 'Array of driver names (e.g., ["bigquery", "firestore"]) to filter available tools. Omit to list all.' 'name' → 'Fully qualified tool name (e.g., "bigquery/execute_query"). Use tools/list to discover available names.' 'arguments' → 'Tool-specific arguments as a JSON object. Structure depends on the tool, see tools/list for schema.'
• Document output schemas in tool descriptions. E.g., 'tools/list returns {tools: [{name, description, inputSchema}], total: number}. tools/call returns {content: [{type, text|resource}], isError?: boolean}.'
• Implement error recovery guidance: 'Tool not found. Call tools/list to see available tools.' 'Invalid arguments. Check inputSchema in tools/list result.' 'Execution failed: <reason>. Retryable: <true|false>. Next step: <suggestion>.'

Inferred effective spec: 2026-07-28+.

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