caveman--caveman-evidence-review

Load context with caveman_context, then pull overview, costs, score, workflows, and verified_savings reports via caveman_report. Test the leading explanation with bounded trace searches and inspect representative traces before reporting findings.

Use caveman_trace_search with closed filters like workflow, model, or cost bounds over a bounded window, then compare a suspect cohort against a control cohort. Inspect a few high-signal traces with caveman_trace_get to confirm the pattern.

No, it is strictly read-only and never starts, approves, cancels, or rolls back experiments. Proposed actions are reported as proposals only and deferred to the caveman-manage skill for lifecycle review.

No, payload retrieval stays off by default. Metadata, spans, timing, models, token counts, status, and optimizer attribution are sufficient unless the user explicitly requests payload review.

The review stops and asks the user to run caveman login or select a project. It never guesses an organization id or reads outside the project selected by Caveman context.