CascadeGateway has three tools with basic schemas and descriptions, but significant quality gaps. Tool names lack clear action verbs (query_local_5090, cascade_llm are vague about what they return). Descriptions are present but generic (194-char baseline; these average ~120 chars, borderline). Parameters lack type specificity, system_prompt has no type annotation visible in schema. No output schema documented. Error handling returns raw exception strings ('Error communicating with local 5090: {e}') rather than actionable guidance. No parameter validation, no enum constraints, no pagination support. The get_cascade_metrics tool has empty inputSchema properties but no description of what 'metrics' means. Overall, definitions are skeletal, functional but not LLM-optimized.

• Rename tools to verb_noun pattern: 'run_local_inference', 'invoke_cascade_gateway', 'fetch_cascade_metrics'. This clarifies intent and helps LLMs disambiguate.
• Add explicit output schemas to all tools. E.g., query_local_5090 returns {type: 'object', properties: {result: {type: 'string', description: 'LLM response text'}, model_used: {type: 'string', enum: ['local-5090', 'gemini']}, tokens_used: {type: 'integer'}}}.
• Replace raw exception strings with structured error objects: {error: 'service_unavailable', message: 'Local RTX 5090 not responding on port 8000', recovery: 'Retry in 30s or check Ollama status', retryable: true}.
• Add parameter descriptions and constraints: system_prompt should have maxLength: 2000, description: 'System context for the LLM (max 2000 chars). Defaults to coding assistant prompt.'
• Document get_cascade_metrics return structure in description: 'Returns {tokens_saved_local: number, offload_percentage: number (0-100), biasing_state: string (active|idle), gateway_status: string (ready|unavailable)}. Call to monitor cascade efficiency and decide when to offload to cloud.'