Guide / Codex

How to review code written by Codex.

With the skill, or by hand with five commands.

To review a change Codex made, do not rely on the session that wrote it. Run the project's tests, have a different reader look at the diff, and for logic that matters, get tests that execute against the change and against the code before it. arnx does the last part from the command line, so it works with Codex the same way it works with any agent.

Codex can also drive arnx itself through an Agent Skill. arnx skill install puts the skill where Codex's documentation says it looks. We have not yet run the full skill loop inside Codex ourselves, so this guide gives the manual flow first. It depends on nothing but a shell.

The manual flow, for any agent

You need Docker running and a model key, or arnx Cloud.

Once per repository

npm install -g @arnx-dev/arnx
arnx init

Let Codex finish the change. Stage any new files it created with git add. arnx reviews tracked files only, so an unstaged new file is left out.
Run arnx review. On a branch it looks at the commits since the default branch plus uncommitted tracked work. It writes tests, runs each three times in Docker, reruns failures on the code before the change and asks a second model about each.
Run arnx triage, then arnx accept. You mark each issue real, intended, noise or rewrite. Tests of real issues are kept in .arnx/tests/.
Give Codex the fix prompt. arnx prompt prints it: the failing tests, each bug, your answers as intended behavior, and the rule to fix the code and leave the tests alone. Paste it into Codex.
Run arnx check. It reruns the accepted tests with no model calls. Repeat the last two steps until it passes, then run the project's own suite.

The skill in Codex

In the repository

arnx skill install

This writes .agents/skills/arnx/, with SKILL.md and its references. Codex's skills documentation says it scans .agents/skills in every directory from the working directory up to the repository root, and $HOME/.agents/skills for user-wide skills (arnx skill install --user writes there). npx skills add arnx-dev/arnx installs the same folder.

What we have checked: the files land in the directory Codex documents, with the name and description fields it requires. What we have not: a full review, triage, fix and check loop driven by Codex. If you run it, tell us what worked or broke.

The skill gives the agent the same split as the manual flow. The agent runs arnx review once, after saying what it may cost, and later arnx check, arnx prompt and the fix. arnx init, logins, triage, accept and commits stay yours. The details are on the agent page.

Use a different reader

A general rule for agent-written code: the checker should differ from the author. A fresh session helps. A different model helps more. A test that runs helps in another way, because it does not depend on any model's reading being right.

arnx's default harness reviews with Claude and, when you set OPENAI_API_KEY as well, adds a second hunter from OpenAI. You can change every role in .arnx/config.toml, for example to make the reviewer a different vendor from the agent that wrote the code. See models. How much a second model's verdict is worth is measured in can a second LLM judge code review findings?

Limits and trade-offs

The manual flow costs you a few copy and paste steps. In exchange you see every finding before any fix is written.
A Codex sandbox without Docker access cannot run arnx review or arnx check. Run them in your own terminal.
Each review stops writing tests at review_usd, $3.00 by default. Our own reviews ranged from $0.17 to $3.19. arnx check costs no model calls.
Generated tests can be wrong. In one review of SQLFluff the only confirmed finding was a bad test, while a real gap sat under another heading. Read that case.
Python, JavaScript, TypeScript and Rust (early). Changes up to 200 KB.

Try the flow on the checkout demo: clone it, git switch discount-codes, arnx init, arnx review. The quickstart covers each command.

Last reviewed 2026-11-12 against arnx 0.1.11. The release documentation is canonical for your installed version.