review

name	review
description	Accept an implementation — gates first, then human judgment — deciding mergeability against the spec's acceptance criteria, and wrapping up with distillation and archiving. Use for review, checking a diff/PR, acceptance, or judging mergeability.

Review

Acceptance, not a line-by-line recheck. What a machine can check goes to the gates; this skill only makes the judgments the gates can't cover.

Prerequisites

• Run the automated gates defined in AGENTS.md first. Gates not green → reject immediately, no human review.
• When the repo has no AGENTS.md or no defined gates, substitute its existing test / lint / build commands; if there's no runnable verification at all, say so honestly and write "verification missing" into the findings itself.
• Find the corresponding spec (default: search docs/specs/). With no requirement source, review by code quality and risk, and state clearly that requirement conformance can't be judged.

Review

1. Against the acceptance criteria: check off each of the spec's Acceptance criteria, with evidence (test, output, code location); "should be fine" is not accepted.
2. Check scope: any creep into Non-goals, any unrelated refactor mixed in.
3. Check risks the gates can't cover: boundary conditions, data migration, compatibility, security, concurrency, error handling.
4. Check tests: whether the critical path and failure path of new behavior are covered.

Findings

Sorted by severity, only actionable issues:

- [P1] Title
  File/location:
  Problem / Impact / Suggestion:

P0 data loss, serious security, core flow unusable; P1 must fix before merge; P2 boundary issue, test gap; P3 minor, non-blocking.

Verdict

One of three: mergeable / revise and re-review / needs re-shape. When there are no issues, say explicitly that no blocking issues were found.

Wrap-up distillation (when the verdict is mergeable)

• Flip the spec's status to done, move it into docs/specs/archive/.
• Stable terms and system boundaries that emerged during implementation go into CONTEXT.md; long-term decisions go into an ADR, with any overturned old ADR marked superseded.
• If this was a large feature, suggest a scoped gc over the changed area.

Completion criteria

• Gate results confirmed; each acceptance criterion has evidence.
• Every finding has a clear impact and suggestion; no style preference disguised as a defect.
• The verdict directly supports the next action.
• When mergeable, distillation and archiving are done or explicitly listed as pending.