incident-lens has well-structured tool definitions with clear naming (verb_noun pattern), comprehensive parameter descriptions, and proper schema validation via Zod. All 5 tools follow consistent patterns and include time-range parameters with sensible defaults. However, output schemas are not explicitly documented in the visible code, tool handlers return JSON-stringified results without formal schema declarations. Error handling is minimal; tools return raw results without recovery guidance. Parameter constraints are good (enums, min/max) but some descriptions could be more prescriptive about format expectations.

• Document output schemas for each tool. Define the structure of metrics, logs, errors, deployments, and timeline objects. Include field types, required fields, and examples.
• Add error recovery guidance. When a backend is unavailable, return: 'Prometheus unreachable. Check config.yaml and verify URL/auth. Retry with a different time range or check get_deployments() for recent changes.'
• Enforce and document result limits. Cap query_logs at 1000 entries and get_errors at 100 groups. Document pagination strategy (cursor or offset) if results exceed limits.
• Specify time format constraints in parameter descriptions. E.g., 'ISO 8601 (2024-01-15T10:30:00Z) or relative (1h ago, now-2h, 30m). Relative times are parsed relative to the current server time.'
• Add tool annotations to the schema. Mark all tools with readOnlyHint=true and idempotentHint=true so LLMs know they are safe to retry without side effects.