Review candidate changes
Inspect the candidate diff, proof evidence, lineage, and source boundaries.
After workbench skill improve creates a candidate, review the source diff, proof evidence, and version lineage.
See Improve from evidence for the improvement command and Rerun proof evals for another proof attempt.
Inspect history
Show the candidate, its base version, and proof evidence:
workbench skill versions lists recorded package versions. workbench skill diff compares package files. workbench eval show reads run, job, trace, artifact, and file evidence.
Review the diff
Improvement changes package source, not eval source. Files under .workbench/** are the grading standard; changes there are eval changes, not skill improvements.
Look for:
- Changes that address the failing evidence directly.
- New instructions, references, scripts, or assets that remain self-contained in the skill package.
- Removal of brittle or unrelated source changes.
- No accidental edits to
.workbench/**unless you changed the eval on purpose.
Review proof evidence
Proof evidence needs to show the candidate beating the current version on the selected evidence. Inspect the run and job details before publishing:
If proof evidence is too narrow, run a higher-sample eval or expand cases before publishing.