| name | autoreview |
| description | Run an available second-model review helper for closeout. Use when the user requests autoreview, Codex/Claude review, or pre-ship external review; not ordinary self-review. |
Auto Review
Run a structured closeout review helper against the current change. This is advisory code review, not approval routing.
This skill is the reviewer role from clean-context delegation: a second opinion on changes. When the host exposes a fork/subagent tool, a clean-context delegate is the preferred dispatch; the helper script below is the concrete implementation when it is available.
Based on the OpenClaw autoreview skill; this package bundles the workflow, not the helper script.
Entry protocol
- Find an available helper, in order:
.agents/skills/delivery/autoreview/scripts/autoreview
skills/delivery/autoreview/scripts/autoreview
~/.codex/skills/agent-scripts/autoreview/scripts/autoreview
- any project-documented
autoreview helper path
- If no helper exists, report blocked with the missing helper. Do not invent a reviewer command.
- Use Codex review by default unless the user explicitly requested another engine.
- Do not push just to review. Push only when the user requested push, ship, or PR update.
Skill composition
- Run after
tdd, diagnose, or pi-extensions-helper has produced local validation evidence; review is not a substitute for tests.
- Feed accepted findings back to the relevant specialist with trigger, finding artifact, next skill, and expected fix validation signal: bugs to
diagnose, missing behavior coverage to tdd, API/package concerns to pi-extensions-helper.
- If the user requested ship after review, hand the clean review result and test receipts to
git-commit-push.
- If Goal mode is active, record accepted/rejected findings and final clean review as Goal evidence.
Contract
- Treat findings as advisory. Never blindly apply them.
- Verify every accepted finding by reading the real code path and adjacent files; when
codebase-map-understand.md exists and a finding depends on cross-module impact, consult the codebase map for caller/path leads and verify them in source.
- Read dependency docs/source/types when a finding depends on external behavior.
- Reject speculative edge cases, broad rewrites, and fixes that over-complicate the codebase.
- Prefer small fixes at the right ownership boundary.
- If a review-triggered fix changes code, rerun focused tests and rerun autoreview.
- Stop when the helper exits 0 with no accepted/actionable findings.
- Do not run nested reviewers, built-in
codex review, or reviewer panels from inside the review.
- Multi-reviewer panels are opt-in only: use them when explicitly requested or risk justifies the spend.
Pick target
Dirty local work:
<autoreview-helper> --mode local
Use local mode only when the patch is actually staged, unstaged, or untracked in this checkout.
Branch or PR work:
<autoreview-helper> --mode branch --base origin/main
If an open PR exists, prefer its actual base:
base=$(gh pr view --json baseRefName --jq .baseRefName)
<autoreview-helper> --mode branch --base "origin/$base"
Committed single change:
<autoreview-helper> --mode commit --commit HEAD
Use commit mode for already-landed or already-pushed work on main; reviewing clean main against origin/main is usually empty after push.
Options
--parallel-tests "<focused test command>" can run tests and review together after formatting.
--prompt-file and --dataset add review context.
--panel or --reviewers codex,claude run opt-in reviewer panels.
--model and --thinking set per-engine model/effort when explicitly needed.
- Keep helper mode as
auto unless target selection needs a precise mode.
Verification gate
Before final response:
- accepted findings were verified against source;
- rejected findings have a concise reason;
- focused tests/proof ran after any fix;
- final autoreview rerun is clean, or remaining findings are consciously rejected with reasons.
Final report
Include:
- review command used;
- tests/proof run;
- findings accepted/rejected and why;
- final clean review result, or why remaining findings were intentionally rejected.
Do not run an extra review solely to improve final wording.
Shared contract
Follow the shared skill contract for repo study, dirty-worktree hygiene, verification evidence, safe handoffs, and safety defaults.