Four tools with comprehensive schemas and detailed descriptions. All tools have proper input schemas with type definitions and parameter descriptions. Tool names follow verb_noun convention (recommend_model, compare_models, search_models, router_status). Descriptions are substantive (194-250 chars average), exceeding the 10-1024 char baseline. Parameters use Zod validation with enums, min/max constraints, and clear descriptions. Output structure documented via structuredContent pattern. Main gaps: no explicit error recovery guidance in descriptions, no confirmation pattern for potentially expensive operations (recommend_model with ping), and output schemas not formally documented in code comments.

• Add explicit recovery guidance to recommend_model and compare_models descriptions: 'If no models match your constraints, call search_models() to browse available options or relax constraints (e.g., remove provider filters).'
• Document the structuredContent JSON schema for each tool in JSDoc comments. Example: 'Returns {pick: {id, name, cost_usd, time_ms}, runners_up: [...], smartest: {...}, fastest: {...}, cheapest: {...}}'.
• Add toolAnnotations (readOnlyHint, idempotentHint) to tool registration for recommend_model and compare_models to signal they are safe to retry.
• For search_models, add cursor-based pagination support: accept an optional 'cursor' parameter and return 'next_cursor' in structuredContent so agents can fetch all models without hardcoding limit.
• Add a 'dry_run' parameter to recommend_model (default false) to let agents preview recommendations without incurring ping costs, or document the cost upfront in the ping parameter description.