Guide / Claude Code

How to review code written by Claude Code.

Three layers, from a second read to tests that run.

Review Claude Code's work in layers. First let it run the project's own tests and type checks. Then get a second read of the diff, which Claude Code can do itself with /code-review. Then, for changes where behavior matters, get tests that execute against the change. The last layer is what arnx adds, as a plugin or a skill.

The reason for layers: the session that wrote the code tends to share its own blind spots. A fresh read catches some of them. A test that runs catches a different set, because it does not depend on anyone's reading being right.

What Claude Code reviews on its own

These exist today, according to Anthropic's documentation:

/code-review in the terminal, with /review as an alias. It reports findings as text. --fix applies them to the working tree and --comment posts them on a pull request. Local review docs.
/code-review ultra, a research preview that runs reviewer agents in a cloud sandbox. Ultrareview docs.
/security-review for security issues in a change.
Code Review, a managed service that reviews pull requests on GitHub for Team and Enterprise plans. Its docs say a review works from reading the diff and the surrounding code, without building or running your code. Code Review docs.

These are a good second read and cost no setup. A side-by-side with sources is in arnx and Claude Code's native review.

Adding tests that run

arnx writes tests for the change, runs them in Docker, reruns failures on the code before the change and has a second model challenge each one. Install the CLI, then give Claude Code the skill in one of two ways.

In your shell

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

arnx skill install writes .claude/skills/arnx/ (and .agents/skills/arnx/ for other agents). Commit them so teammates get the same workflow. Or install it as a plugin inside Claude Code:

In Claude Code

/plugin marketplace add arnx-dev/arnx
/plugin install arnx@arnx

arnx init is yours to run: it writes the config that names your models. With ANTHROPIC_API_KEY set, every job runs on Claude and you pay Anthropic directly. Your Claude subscription does not cover these API calls. Without a key, use arnx Cloud.

Who does what

The skill tells Claude Code which steps are its own and which are yours. Claude runs the review once, then stops. You judge the findings. Claude fixes only what you accepted.

Claude Code	You
Says what the review may cost, then runs it once	`arnx init`, logins, consent
Reports findings and questions, changes no code	`arnx triage`: mark findings, answer questions
`arnx check`, `arnx prompt`, the fix, `arnx check` again	`arnx accept`
Runs each touched project's full suite and reports	Commit and push

Claude never edits accepted tests. If it did, arnx check would refuse to run with exit 4. The full rules are on the agent page.

What a session looks like

You ask Claude Code for a change. When it is done, it tells you the review budget it read from .arnx/config.toml and runs arnx review. The first review also sets up the project, in at most 10 minutes.
It lists each finding with its input, actual and expected behavior, each question from arnx word for word, and the cost.
You run arnx triage and arnx accept in your own terminal, then tell Claude.
Claude runs arnx check, reads arnx prompt, fixes the code and checks again until the accepted tests pass. After three failed rounds it stops and reports.
It runs the project's own suite and gives you a summary. You commit the fix with the accepted tests.

A coding agent fixed arnx's own code from an arnx fix prompt. One result is in the path escape arnx found in its own remove command.

Limits and trade-offs

The review stops for you. That is deliberate: the tests are model-written, and in our grading many confirmed findings tested intended behavior. See the counts.
If Claude Code runs in a sandbox that cannot reach Docker, it cannot run the review or the check. It hands you the command.
The skill tells agents typical reviews cost $0.02 to $0.50. Each stops writing tests at review_usd, $3.00 by default.
A reviewer and the author sharing a model family may share blind spots. arnx's default second hunter and fallback use another vendor when you set that key. See models.
Python, JavaScript, TypeScript and Rust (early) only. Code in a project with no test suite is reported as not reviewed.

To try it without touching your own code: clone the checkout demo, switch to discount-codes, and ask Claude Code to review the branch with arnx. Setup details are in the quickstart.

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