| name | sakura-review |
| description | Review final changes against the specification, acceptance contract, and actual verification evidence, then return PASS, PASS_WITH_NOTES, FIX_REQUIRED, or BLOCKED. Use for final completion decisions or independent code review; do not implement a feature from scratch. |
Sakura Review
Role
Inspect correctness, scope, regression risk, security, privacy, compatibility, and completion evidence before delivery, then return one deterministic verdict.
Use When
- The active milestone or complete task has finished verification.
- The user requests review of a working tree, commit, branch, or pull request.
- The task needs a reliable Done decision.
Do Not Use When
- No reviewable diff or implementation exists.
- Critical verification has not run and no valid blocker explains why.
- The user requested implementation or technical research.
Inputs
- The actual diff and relevant code context.
docs/project/PRODUCT_SPEC.md, ACCEPTANCE.md, and EXECUTION_PLAN.md.
- Task state, verification results, error history, and current limitations.
- Replication scope, levels, contracts, matrix, evaluator result, and gaps when an existing system is the target.
Outputs
- Specific findings ordered by severity.
- A location, evidence, impact, and smallest repair direction for every finding.
- One of
PASS, PASS_WITH_NOTES, FIX_REQUIRED, or BLOCKED.
- A handoff to Done,
sakura-fix, or another required stage.
Workflow
- Confirm the review scope and preserve unrelated user changes.
- Inspect the actual diff, surrounding code, and real verification output instead of summaries alone.
- Check every Must acceptance item, failure path, regression risk, security boundary, privacy concern, and compatibility requirement.
- Check for unrelated refactoring, unexpected artifacts, secrets, and evidence gaps.
- For replication work, confirm that target and architecture signals produced the correct L1-L4 requirement, every required level has evidence, and no mandatory gap or approximation is hidden.
- Require every finding to identify a location, explain impact, and provide an actionable repair direction.
- Select one verdict from the severity and completion rules.
- When acting as the parent, write review state and run the completion gate only when the task is eligible to finish. A delegated reviewer returns findings and a report verdict but never writes shared state or Done.
Read review-checklist.md for a formal review. Read completion-evaluation.md before deciding whether the task is complete.
Quality Gate
- The review is based on actual code and verification evidence.
- Findings are concrete and evidence-based, without vague advice or filler praise.
- Blocking and non-blocking issues are clearly separated.
- The verdict matches acceptance, verification, and task state.
- A failed review cannot write Done.
- The parent independently checks delegated findings against the actual diff and remains the only owner of the final review verdict.
- Replication Done is rejected when any capability is under-classified, partial, approximate, lacks required evidence, or has an open mandatory gap.
Failure and Fallback
- Return
BLOCKED when critical code, environment, acceptance, or evidence is unavailable.
- Return
FIX_REQUIRED and hand off to sakura-fix when a blocking implementation problem exists.
- Return to
sakura-spec when expected product behavior is wrong.
- Return
PASS_WITH_NOTES only when remaining items are explicitly non-blocking.
Upstream and Downstream
- Upstream:
sakura-verify, or a user-requested independent review.
- Downstream: Done,
sakura-fix, sakura-spec, or Blocked.
Handoff
Include the verdict, blocking count, key findings, completion evidence, unverified items, remaining risks, and next action. Only PASS or PASS_WITH_NOTES with zero blockers may proceed to Done.