Review PR (final review)

The last gate before a PR is marked ready. Two reviewers read the PR's diff on two separate axes:

• Seams — SEAMS.md: does the change respect the app's documented seams, and do the seam docs still tell the truth?
• Rules — RULES.md: does the code follow .ab-method/rules/code.md, the book's red flags, and the repo's own standards?

Each runs in its own subagent with a fresh context, so neither inherits the context that wrote the code and neither can mask the other. Their reports are shown side by side, never merged or reranked. See CREDITS.md.

Silence is the default output. A PR that follows its seams and rules produces no findings. A finding survives only if it names a concrete cost. Never invent churn to look thorough.

ALWAYS check .ab-method/structure/index.yaml FIRST for paths (task location, docs/seams/, review.md). Look for it under the project root first; only if the project has none (plugin install), read the bundled default at ../../../.ab-method/structure/index.yaml relative to this SKILL.md. Every path it names is relative to the project root. Talk to the user per .ab-method/rules/communication.md.

start-task's inline mode does not change this skill: the reviewers are always subagents.

Process

1. Pin the PR and the diff

• The PR: the number passed in, or the current branch's PR (gh pr view --json number,baseRefName,url).
• No PR (no remote, no gh, or a review before opening one): review the branch instead. The base is the one passed in, else the repo's default branch. Everything below runs the same; what would have been the PR comment goes only into review.md.
• The diff against the merge-base: git diff <base>...HEAD, and git log <base>..HEAD --oneline. Confirm both resolve and the diff is non-empty before spawning anything.

2. Gather context (read only what exists)

• The task's progress-tracker.md (Placement, missions and summaries, ## Protected) and unresolved-questions.md.
• docs/seams/index.md, then only the seam docs the diff touches.
• .ab-method/rules/code.md and the codebase-design skill's RED-FLAGS.md and SEAM-FORMAT.md.
• Repo standards: CLAUDE.md, AGENTS.md, CONTRIBUTING.md, CODING_STANDARDS.md.
• docs/adr/ — nothing an ADR settled is a finding.

3. Spawn both reviewers in parallel

Issue both subagent calls in one message, each read-only, each with a fresh context. On Codex spawn them at the orchestrator's level; on Claude they may nest.

Each prompt carries: the diff command and commit list, its brief (SEAMS.md or RULES.md, pasted in full, or its path with the instruction to read it in full first), the context files from Step 2 that axis needs, and the output rule below. Each returns under 400 words.

Every finding has:

• Where — file and hunk.
• What — the violation, citing its source (seam doc line, rule, red flag, standard).
• Cost — the symptom it causes: change amplification, cognitive load, or unknown unknowns.
• Fix — plain English.
• Class — safe-fix (mechanical, tests stay green, no design decision) or needs-judgment.

Each reviewer also ends with its Checked: evidence line; a clean axis without one didn't run, so re-run it. Near-misses already listed under ## Checked, dropped in review.md are not re-raised.

A recorded black box (a TODO(UQ-n) seam with an entry in unresolved-questions.md) is not a finding.

4. Report side by side

Present the two reports under ## Seams and ## Rules, lightly cleaned. End with one line: findings per axis and the worst issue within each axis. Don't pick a winner across axes.

5. Act — by mode

Autonomous (start-task / start-roadmap):

1. Apply each safe-fix one at a time, re-running the focused tests after each; a fix that goes red is reverted and becomes needs-judgment. Seam-doc edits must pass node .ab-method/scripts/check-docs.js (plugin installs: npx ab-method check-docs).
2. Commit what was kept as refactor(<task>): PR review fixes and push to the PR branch. Nothing kept → no commit.
3. Record every needs-judgment finding in the task's review.md under ## PR review (create the file if absent), and post them as one PR comment (gh pr comment <n> --body-file ...). Clean on both axes → record PR review: clean in review.md, post nothing.
4. Never prompt.

Interactive (create-task / resume-task, or standalone): present the findings, apply what the user approves the same way (tests green, docs gated), commit, push, and record the rest in review.md under ## PR review.

Remember

• Index first; review only this PR's diff, never the whole repo.
• Two fresh-context reviewers, in parallel, reported separately.
• Safe fixes are pushed; judgment calls go to review.md and one PR comment.
• Silent by default.