Workbench Review QA
Use this skill for code review, workflow review, QA verification, release checks, and issue closeout.
Verdicts
PASS: objective is met and evidence is sufficient.
FLAG: useful progress, but a non-blocking issue or missing evidence remains.
BLOCK: objective is not met, unsafe, unverifiable, or materially wrong.
Review Method
- Identify the promised objective.
- Inspect the real output, diff, issue comments, runtime state, or checked-out repo.
- Compare evidence against the objective.
- Verify the smallest real path that proves the claim.
- Report findings first, ordered by severity.
- Check command syntax and live-resource ownership when the work involves Multica CLI mutations.
- If duplicate comments or artifacts exist, identify the primary artifact and explain why.
- For
GOAL_MODE: yes or /goal work, compare the closeout against the locked objective and every required closeout gate.
Findings Format
For problems, return:
- severity,
- location or evidence,
- why it matters,
- required fix or next verification.
For clean reviews, return:
- verdict,
VERDICT_SUMMARY: three lines or fewer,
- evidence checked,
- residual risk or test gap.
Every SDD review should include VERDICT_SUMMARY so the next agent can continue from the review header without re-reading the full review body.
Auto Review Sweeper
For automatic in_review handoffs, use this exact block on each reviewed target:
AUTO_REVIEW
TARGET: <identifier>
VERDICT: PASS | FLAG | BLOCK
VERDICT_SUMMARY: three lines or fewer
EVIDENCE: concrete issue/comment/run IDs, commands, or file paths checked
STATUS_ACTION: done | kept in_review | blocked | no_change
NEXT_ACTION: exact next owner/action, or none
PASS may move the target issue to done only when the original goal is satisfied and no required follow-up remains.
FLAG leaves the issue in in_review with a bounded next action.
BLOCK sets the issue to blocked with blocking evidence.
- If evidence is still arriving, leave the issue unchanged and mark it pending in the sweep summary.
QA Rules
- Do not accept
done based on paraphrase alone.
- Distinguish content failures from workflow/tooling failures.
- If repo access is needed, prefer the issue's project-bound GitHub repo resource and report the commit/branch inspected.
- For Goal Mode tasks,
PASS requires evidence for the locked objective plus all relevant build/test/help-smoke/docs-or-report/git-status gates, or an explicit rationale for each non-applicable gate.
- Treat
file://<LOCAL_WORKBENCH_REPO> as laptop-local fallback only; remote runtimes must flag it as invalid unless explicitly mounted.
- Keep evidence concise and reproducible.
- Verify
Workbench Max remains untouched when a task says it must be preserved.
1---2name: workbench-review-qa3description: PASS, FLAG, and BLOCK review discipline for workflow review, QA verification, release checks, and issue closeout.4---56# Workbench Review QA78Use this skill for code review, workflow review, QA verification, release checks, and issue closeout.910## Verdicts1112- `PASS`: objective is met and evidence is sufficient.13- `FLAG`: useful progress, but a non-blocking issue or missing evidence remains.14- `BLOCK`: objective is not met, unsafe, unverifiable, or materially wrong.1516## Review Method17181. Identify the promised objective.192. Inspect the real output, diff, issue comments, runtime state, or checked-out repo.203. Compare evidence against the objective.214. Verify the smallest real path that proves the claim.225. Report findings first, ordered by severity.236. Check command syntax and live-resource ownership when the work involves Multica CLI mutations.247. If duplicate comments or artifacts exist, identify the primary artifact and explain why.258. For `GOAL_MODE: yes` or `/goal` work, compare the closeout against the locked objective and every required closeout gate.2627## Findings Format2829For problems, return:3031- severity,32- location or evidence,33- why it matters,34- required fix or next verification.3536For clean reviews, return:3738- verdict,39- `VERDICT_SUMMARY`: three lines or fewer,40- evidence checked,41- residual risk or test gap.4243Every SDD review should include `VERDICT_SUMMARY` so the next agent can continue from the review header without re-reading the full review body.4445## Auto Review Sweeper4647For automatic `in_review` handoffs, use this exact block on each reviewed target:4849```text50AUTO_REVIEW51TARGET: <identifier>52VERDICT: PASS | FLAG | BLOCK53VERDICT_SUMMARY: three lines or fewer54EVIDENCE: concrete issue/comment/run IDs, commands, or file paths checked55STATUS_ACTION: done | kept in_review | blocked | no_change56NEXT_ACTION: exact next owner/action, or none57```5859- `PASS` may move the target issue to `done` only when the original goal is satisfied and no required follow-up remains.60- `FLAG` leaves the issue in `in_review` with a bounded next action.61- `BLOCK` sets the issue to `blocked` with blocking evidence.62- If evidence is still arriving, leave the issue unchanged and mark it pending in the sweep summary.6364## QA Rules6566- Do not accept `done` based on paraphrase alone.67- Distinguish content failures from workflow/tooling failures.68- If repo access is needed, prefer the issue's project-bound GitHub repo resource and report the commit/branch inspected.69- For Goal Mode tasks, `PASS` requires evidence for the locked objective plus all relevant build/test/help-smoke/docs-or-report/git-status gates, or an explicit rationale for each non-applicable gate.70- Treat `file://<LOCAL_WORKBENCH_REPO>` as laptop-local fallback only; remote runtimes must flag it as invalid unless explicitly mounted.71- Keep evidence concise and reproducible.72- Verify `Workbench Max` remains untouched when a task says it must be preserved.