Chaos Mesh server has 13 tools with complete input schemas and descriptions, but quality is inconsistent. All tools have verb-noun naming (pod_kill, container_kill, etc.) and parameter descriptions. However, descriptions are generic and lack actionable context for LLM selection. No output schemas documented. Critical gap: no error handling guidance, no recovery paths, and no confirmation/dry-run patterns for destructive operations (pod_kill, network_partition, delete_experiment). Parameters like 'mode' and 'value' are poorly explained, LLMs cannot infer valid mode values from descriptions alone. No enum constraints despite clear enumerated options (mode: one|all|fixed|fixed-percent|random-max-percent). Risk annotations present but not integrated into tool definitions.

• Add enum constraints to 'mode' parameter: mode: enum[one, all, fixed, fixed-percent, random-max-percent]. Update description to explain each option and when to use it.
• Document output schemas for all tools. Example: 'Returns: {experiment_name: string, namespace: string, status: string, created_at: ISO8601}'. Include fields needed for chaining (e.g., experiment_name for delete_experiment).
• Add confirmation/dry-run pattern for destructive tools. Introduce a 'dry_run: bool' parameter (default false) or a separate 'preview_experiment' tool that shows what would happen without executing.
• Document parameter dependencies explicitly. For 'value', add: 'Meaning depends on mode: if mode=fixed, value is pod count (int); if mode=fixed-percent, value is percentage (0-100); if mode=one or all, value is ignored.'
• Add error handling guidance to each tool description. Example: 'Errors: service not found (try list_services), invalid mode (must be one of: one, all, fixed, fixed-percent, random-max-percent), invalid duration format (use 5m, 10s, etc.)'

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry