Two tools with basic schemas and minimal descriptions. record_latency and get_percentiles have input schemas with typed properties, but descriptions are under 50 chars and lack context on WHEN to use them, WHAT they return, or dependencies. No parameter descriptions beyond property names. No output schema documentation. No error handling guidance. Naming is verb-noun compliant but generic. Missing idempotency hints, retry guidance, and structured output documentation.

• Expand tool descriptions to 100-200 chars. Example: 'record_latency: Records the execution time of an agent step in milliseconds. Call this after each step completes to build a latency profile. Returns OK on success. Retries will append duplicate samples, use idempotent_record_latency if deduplication is needed.'
• Add parameter descriptions in the schema. For step_name: 'Unique identifier for the agent execution step (e.g., prompt_formatting, llm_generation). String, 1-100 chars.' For latency_ms: 'Execution duration in milliseconds. Must be >= 0. Decimal values allowed.'
• Document the output schema for get_percentiles. Include field types and ranges: 'Returns {step_name: string, sample_count: int, min_ms: float, avg_ms: float, p50_ms: float, p90_ms: float, p95_ms: float, p99_ms: float, max_ms: float}. Returns null if step_name has no samples.'
• Add input validation and error messages. Reject empty step_name with 'step_name cannot be empty'. Reject negative latency_ms with 'latency_ms must be >= 0, got {value}'. Return structured error objects with actionable guidance.
• Add an idempotent variant or document non-idempotency. Either implement deduplication (e.g., record_latency_idempotent with a timestamp key) or add to description: 'WARNING: This tool is non-idempotent. Retries will record duplicate samples. Use with caution in retry loops.'