You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
feat(code-review): add a ReviewBench harness for the review pipeline
The harness runs the Standard, Quick and Deep review presets as a
ReviewBench reviewer image. It reuses the production prompt builders,
parser and dedupe key. Pure helpers move from executor.ts to unitPure.ts,
so the harness bundle has no database code. executor.ts re-exports them.
On the 25-PR test set, Standard scores 66.7% high-severity grounded F1
at 100% precision with a local judge. FINDINGS.md has the results and
the prompt experiments that did not generalize.