Local Review Codex
Testing & Qualitylocal-review-codexLocal Review Codex is an Agent Skill for testing & quality published in the windmill-labs/windmill repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.
Run the CI Codex PR review locally against this branch's unpushed work (committed + uncommitted) before pushing. Same policy, model, and reasoning effort as the codex-pr-review GitHub action.
How to use this skill
Bring this guide into your coding agent with a prompt tailored to the tool you use.
- Open your project in Codex.
- Copy and paste the prompt below.
- Review files and risks before approving any action.
Inspect the Agent Skill at https://github.com/windmill-labs/windmill/blob/HEAD/.agents/skills/local-review-codex/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/local-review-codex, but wait for my approval before writing files or running scripts.
Copying this prompt does not install or run the skill. Review third-party files before use.
Local Codex Review (pre-push)
Runs the exact same review Codex performs in CI (.github/workflows/codex-pr-review.yml),
but locally and scoped to work you have not pushed yet — so you catch what CI would flag
before the PR exists. Use this before git push on a non-trivial change.
Correspondence with CI — identical:
- Policy:
REVIEW.md(severity triage, public-surface checklist, AGENTS.md compliance, test coverage). - Model:
gpt-5.6-sol,model_reasoning_effort="xhigh". - Output: markdown starting with
## Codex Review, findings tagged P0 / P1 / P2 with file:line.
Differences from CI — local-only:
- Scope is the current branch vs
mainat the merge-base, including uncommitted changes (CI reviews a pushed PR diff). - Sandbox is
read-only(CI usesdanger-full-accesson an ephemeral runner). Codex reads the diff and files but cannot modify your working tree. - Fresh context is inherent:
codex execis a separate cold process, so it does not anchor on the current chat session — the same reasonlocal-reviewinsists on a subagent.
Prerequisites
codexCLI >= 0.144.1 installed and authed (codex loginorOPENAI_API_KEY). Older CLIs rejectgpt-5.6-solwith "requires a newer version of Codex". Upgrade withnpm install --global @openai/[email protected](may needsudofor a global install). Keep this in sync with the pin in.github/workflows/codex-pr-review.yml.git fetchthe base ref if it's stale, so the merge-base is accurate.
Run
bash .agents/skills/local-review-codex/run.sh # review vs main (default)
bash .agents/skills/local-review-codex/run.sh <base> # review vs a different base ref
Invoke with bash (or run the executable directly) — the script needs Bash for
set -o pipefail; sh is Dash on Debian/Ubuntu and would fail. If main isn't a
local branch (e.g. a fresh single-branch checkout), the runner falls back to
origin/main automatically.
The script computes BASE_SHA = git merge-base HEAD <base>, feeds Codex REVIEW.md plus a
diff context pointing at git diff <BASE_SHA> (which folds in uncommitted edits), and prints
the review. It writes only temp files — nothing lands in the working tree.
Relaying the result
Print the Codex output verbatim. Do not re-summarize or filter it — the value of a cold Codex pass is surfacing what the current session would rationalize away. Then decide with the user whether to address findings before pushing.
For a Claude-native review instead, use local-review (branch-diff-reviewer subagent). This
skill is the Codex counterpart; run both for independent perspectives.