Single tool 'execute_skill_action' has a minimal description (89 chars, below 194-char baseline) that lacks WHEN to use it, prerequisites, or state-modification warnings. Input schema present but parameter 'query_payload' has only a generic description ('Input execution payload'). No output schema documented, LLMs cannot predict response structure. No error handling guidance. Tool name is vague (execute_skill_action) and does not follow verb_noun pattern clearly. No parameter constraints, enums, or validation rules visible. The implementation returns hardcoded success responses, masking real behavior.

• Rename tool to 'reap_cloud_resources' or 'cleanup_ephemeral_instances' to clearly signal the action and match the domain (cloud resource cleanup).
• Expand tool description to 150 - 250 chars: state WHAT it does (identifies and terminates idle cloud resources), WHEN to use it (cost optimization, cleanup workflows), and WHAT it returns (count of resources reaped, cost savings, audit URL). Include: 'This operation is destructive and irreversible, resources are terminated immediately.'
• Document output schema explicitly in the tool definition: { 'reaper_execution_id': string, 'idle_resources_discovered_count': integer, 'orphaned_volumes_reaped_count': integer, 'idle_gpu_instances_terminated_count': integer, 'projected_monthly_cost_saved_usd': number, 'reaper_audit_manifest_url': string }.
• Replace generic 'query_payload' parameter with explicit, typed parameters: 'max_idle_age_hours' (integer, 1 - 720, default 24), 'target_resource_tags' (object with string key-value pairs, e.g. {'env': 'ephemeral-ci'}). Add descriptions for each.
• Add error handling: return structured errors with actionable guidance. E.g., 'No resources matched filters. Available tags: [list]. Adjust max_idle_age_hours or target_resource_tags and retry.' Include retry classification (retryable vs. user-fixable).