Three tools with clear action verbs and reasonable descriptions. All have input schemas with type definitions. However, descriptions lack LLM-optimized guidance on WHEN to use each tool vs alternatives, and parameter descriptions are sparse. The 'delegate' tool's role enum is empty at registration time (populated dynamically), which breaks schema validation. Error handling is present but generic. Output schemas are undocumented.

• Populate role enum statically at server startup or via a discovery tool (e.g., 'list_available_roles'). This enables static schema validation and LLM role discovery without runtime config parsing.
• Document output schema for all three tools. Specify that responses include 'content' array with 'type' and 'text' fields, plus metadata (model name, duration, token counts). Example: {"content": [{"type": "text", "text": "..."}], "isError": boolean}
• Add parameter descriptions to delegate_custom: 'system_prompt' should describe format and constraints (e.g., 'Custom system prompt for the model; appended to any role-specific instructions').
• Enhance tool descriptions with decision guidance: 'Use delegate for preset roles (faster, cheaper). Use delegate_custom only when no preset role fits, requires manual system prompt engineering.'
• Categorize errors in responses: return structured error objects with 'code' (e.g., 'MODEL_NOT_FOUND', 'CONNECTION_ERROR') and 'retryable' boolean, enabling LLM to decide whether to retry, ask user, or escalate.