What problem does it solve? Caveman's report-only repository observations describe aggregate usage shapes but do not tell you what to change or whether a change is safe. This Skill turns those observations into an operator-chosen, evidence-first optimization workflow with a paired baseline/candidate evaluation, preventing premature edits, invented savings claims, and unsafe actuation. ## Core Features & Use Cases - Observation Reading: Runs caveman opportunities list and reads only the report_only_observations array, preserving server-provided titles and observations verbatim while rejecting retired profile ids. - Operator-Gated Candidate Design: Requires an explicit operator choice before inspecting callsites, then proposes one minimal candidate change with a paired eval using identical fixed inputs for both arms. - Guarded Application and Reporting: Applies only the approved candidate, runs the paired evaluation plus repository checks, reverts on quality regression or inconclusive results, and reports outcomes without converting token reductions into dollar savings. - Use Case: A Caveman report shows a tool-output-size-profile observation. The operator selects it, the Skill identifies the responsible callsite, proposes a minimal change, runs baseline and candidate on the same fixture, and reports whether the candidate earned adoption. ## Quick Start Ask the assistant to list the current Caveman report-only observations and help evaluate an optimization candidate for one of them.
