caveman-evidence-review

Load context with caveman_context, pull baseline reports via caveman_report for costs, score, and workflows, then test explanations with caveman_trace_search and inspect representative traces with caveman_trace_get. CLI fallbacks like caveman cloud costs and traces search work when MCP is unavailable.

Measured cost is provider-complete list-price spend, inferred headroom is a per-day estimate of possible savings, and verified savings come from the ledger. The skill keeps these buckets separate and never adds or relabels them.

No. It is strictly read-only and never starts, approves, cancels, or rolls back experiments. Proposed actions are listed as suggestions only and deferred to a separate management skill with a safety gate.

Not by default. It uses metadata, spans, timing, models, token counts, status, and optimizer attribution. Payload retrieval only happens when the user explicitly asks for payload review.

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