Short answer: pipe a git diff into Codex CLI in read-only mode with an explicit review checklist, then run it headless in CI. Reviewing a diff is far cheaper and more reliable than asking it to review a whole repository.

Why review the diff instead of the repo?

A full-repo review pulls in thousands of files and produces comments about code nobody touched. A diff review is scoped, fast, and produces findings the author can act on immediately.

Codex CLI already thinks in diffs because it lives in git. Pointing it at git diff main...HEAD gives it exactly the change set.

How do you run a diff review?

1. Generate the diff: git diff main...HEAD > /tmp/pr.diff.
2. Run Codex in read-only mode so it cannot edit while reviewing.
3. Supply a checklist in the prompt so findings are consistent across runs.
4. Ask for output grouped by severity, not by file.
5. Post the result as a comment on the pull request.

codex --approval-mode suggest \
  "Review this diff. Report only: (1) bugs that would break at runtime, (2) security issues, (3) missing tests. Ignore style. Group by severity." < /tmp/pr.diff

What should the review checklist contain?

Consistency matters more than cleverness. A fixed checklist keeps the agent from writing an essay about naming conventions.

Naming what to ignore is as important as naming what to look for. Without it, agents flood the review with trivia.

How do you run this in CI?

Codex CLI supports non-interactive execution, so it drops into a pipeline step. Set OPENAI_API_KEY as a secret and run it on every pull request after tests pass.

Two practical notes. Cap the diff size — a 5,000-line diff yields shallow comments, so split large PRs. And treat the output as advisory: post it as a comment, never as a required status check, or the agent will block merges with false positives.

Where does AI review fall short?

It is weak on intent. If a change is correct but surprising, the agent may flag it as a bug because it lacks the ticket context. It also cannot judge performance without profiling data — it will guess at complexity rather than measure.

False positives run high on generated code, config files, and migrations. Expect to spend time tuning the checklist in the first two weeks; after that, review noise drops sharply.

This workflow fits teams with real review backlogs and consistent codebases. It is wasted effort on a two-person project where everyone already reads every line.

Quick Summary

• Review the diff, not the repo — smaller input, sharper findings.
• Use read-only mode and an explicit checklist that names what to ignore.
• Group findings by severity so the author can triage quickly.
• Run headless in CI but keep the output advisory, not blocking.
• Expect false positives in the first two weeks while you tune the prompt.

Short answer: install with npm install -g @openai/codex (or Homebrew), run codex in your project, sign in with your ChatGPT account, and keep the default approval mode until you trust it.

What is Codex CLI and how is it different from ChatGPT?

Codex CLI runs locally in your terminal with direct access to the current directory. It can read files, write diffs, and execute shell commands, all with your approval. ChatGPT in a browser can suggest code; Codex CLI can apply it and run the tests.

It is written in Rust, starts in under a second, and supports the GPT-5 family. There is also a VS Code extension and a cloud version, and the CLI shares authentication with them.

How do you install Codex CLI?

1. Requires Node.js 22 or newer, or use the Homebrew formula on macOS.
2. Run npm install -g @openai/codex.
3. Verify with codex --version.
4. Run codex inside a git repository and complete the browser sign-in.
5. Alternatively set export OPENAI_API_KEY=sk-... for headless use.

Codex CLI only works well inside a git repo — it uses git to show diffs and to let you revert. If your project is not under version control, run git init first.

What does a first task look like?

Start with something small and verifiable.

cd my-project
codex
> add input validation to the signup handler and run the tests

Codex will read relevant files, propose a diff, and ask before applying. After it edits, it runs the test command it finds and reports results. The whole loop usually takes one to three minutes for a change of this size.

What are the approval modes and which should you use?

The default asks before file writes and before shell commands. That is the right setting for week one.

Full auto on a repo with production credentials is how people delete the wrong thing. Use a container if you want unattended mode.

What are the real limits?

It is a local agent, so it inherits your machine’s permissions — no sandbox by default. Long tasks can stall on a single failing test it keeps retrying. And it consumes tokens quickly on large repos because it reads many files to build context.

Costs depend on your plan: ChatGPT Plus or Pro includes Codex usage with limits, while API keys bill per token. Heavy use on a big monorepo can run into tens of dollars a day on the API.

Codex CLI suits developers who live in the terminal and want an agent close to their files. It is a poor fit if you dislike reviewing diffs or you want a graphical interface.

Quick Summary

• Install Codex CLI with npm install -g @openai/codex; Node 22+ required.
• It runs best inside a git repository so changes are reviewable and revertible.
• Keep the default approval mode until you trust the agent with your repo.
• Full auto mode has no sandbox — use a container for unattended runs.
• Usage is included with ChatGPT Plus and Pro, or billed per token via API.