← ClaudeAtlas

3-reviewlisted

Run an independent code review of an implemented work unit via a fresh code-reviewer subagent, then drive the fix loop to a merge-ready state. Use after /2-implement passes the gate.
ChristopherHoward1/agentic-coding · ★ 0 · Code & Development · score 70
Install: claude install-skill ChristopherHoward1/agentic-coding
# /3-review — independent review Input: a work unit whose worktree branch `wt/<slug>` passed the gate. ## Steps 1. **Assemble artifacts, not transcripts:** the diff (`git diff main...wt/<slug>`), the plan path, and nothing else. Reviewers must never see the implementation conversation or your own commentary on the diff. 2. **Spawn both reviewers.** Launch the `code-reviewer` agent (fresh thread) with the diff, the plan path, and the worktree path, instructed to review per its own definition. From the primary checkout, run `scripts/codex-review.sh <slug>` so Codex reviews the branch plan and diff from cold, read-only context and writes `work/<slug>/codex-review.md`. Its exit-code contract is: 0 = APPROVE, so record the sentinel; 1 = REQUEST CHANGES, so route the artifact's findings to the followup flow; >=2 = tooling error, not a verdict. Never record a sentinel from prose in the artifact after a non-zero exit; fix the tooling and re-run. 3. **Route findings:** - Substantive findings from either reviewer → back to the implementer via `/2-implement`'s followup flow (same worktree). - Trivial mechanical fixes (typo-grade) → you may fix them directly, but then the re-review in step 4 is mandatory, because you just became a writer. 4. **Re-review after any change:** spawn a *new* fresh code-reviewer thread and re-run `scripts/codex-review.sh <slug>` on the updated diff. Repeat until both approve or 3 rounds total — after 3, escalate the open findings to the Owner rather t