Single tool 'route_query' has a well-structured input schema with 4 parameters, each with type and description. However, the tool definition in mcp_server.py (tools/list response) is MISSING the input schema entirely, only name and description are returned. Per HARD SCORING RULE, this caps schema at 0. The tool description (67 chars) is adequate but generic. No output schema is documented. The client.py implementation shows sophisticated routing logic, but the MCP interface does not expose it properly. Tool is READ_ONLY (safe), but no error handling guidance is provided for edge cases (e.g., no models match constraints, invalid intent detection).

• Register the input schema in mcp_server.py tools/list response. Include all 4 parameters (prompt, max_latency_ms, cost_weight, allowed_models) with types, descriptions, and constraints (e.g., cost_weight: type=number, minimum=0.0, maximum=1.0).
• Document the output schema in the tools/list response. Declare that route_query returns an object with fields: detected_intent (string), selected_model (string), composite_rank (number), estimated_latency_ms (integer), pareto_candidates (array of {model, composite_rank, estimated_latency_ms, blended_cost_per_m, quality_grade}).
• Fix tools/call handler to extract and pass parameters from the request payload. Change client.route_query() to client.route_query(prompt=args['prompt'], max_latency_ms=args['max_latency_ms'], cost_weight=args['cost_weight'], allowed_models=args['allowed_models']).
• Add input validation and error recovery. If no models pass the latency filter, return an error: 'No models available under max_latency_ms=<value>. Increase max_latency_ms or remove allowed_models filter.' This guides the LLM to adjust constraints.
• Enhance the tool description to explain WHEN to use it. Current: 'Dynamically routes user query to the Pareto-optimal LLM based on intent, latency, and cost.' Better: 'Routes a query to the best LLM based on detected intent (reasoning/realtime/standard), latency budget, and cost preference. Use when you need to optimize for speed, reasoning quality, or cost. Returns the selected model and ranked alternatives.'