audit-implement

Run clerk audit run with --base set to the ref the work started from, a one-sentence --brief, and the original --request verbatim. The command scopes the diff, spawns one agent per applicable lens, refutes each finding, and returns a ranked JSON report.

The audit runs semantic and guidelines lenses per language, plus test-integrity when tests changed, and concurrency or performance lenses only when the scoping pass finds real signals. Skipped lenses are reported in lenses_not_run rather than passing silently.

Yes, a coding harness must be on PATH because the round spawns headless agents through whichever it finds. Without one, clerk audit run refuses to start rather than reporting a clean audit; you can also pass --harness-cmd.

Every runtime claim goes to an independent refuter that must reproduce it by execution, such as a failing test or race run; unreproduced claims are dropped. Quality claims must cite a guideline or consistent practice at a specific line to survive.

The round runs in its own process, so a timeout stops only the wait, not the round. Use clerk audit wait to block until it lands, clerk audit status to check liveness, or re-run clerk audit run to resume from the recorded phase.

Skip it for tiny diffs like a two-line change, since a round spawns 10-20 agents over 15-25 minutes. Also fix a red test suite first, because lenses assume the suite passes and auditing broken work wastes the pass.