xreview

Cross-agent code review for Claude Code and Codex: when one agent finishes substantial work, the other one reviews it, read-only, in a fresh session, from the real git diff. It is one plugin that installs in both tools from this repository.

• The agent decides when. The xreview skill tells the implementing agent to request a review before it reports substantial or risky work as done: multi-file logic, refactors, non-obvious bug fixes, and anything touching data splits, preprocessing, training, evaluation, metrics or job scripts. Questions, docs and small edits get no review. You can always ask for one, or say no.
  • Claude Code picks skills by the topic of the request, so the plugin also adds a one-line reminder of this rule when a Claude Code session starts. In the trigger evals, Claude reviewed its own finished work in 0 of 3 runs without the reminder and in 3 of 3 with it, and still did not review questions or typo fixes (3 of 3 each).
  • Codex triggered the skill from its description alone.
• The reviewer sees what actually changed. xreview collects staged, unstaged and untracked changes from git itself, code first. The author's notes are labelled as unverified claims.
• The reviewer sees only what git would commit. It works in a private copy of the repository: git-ignored data and outputs, files that look like credentials, and paths you mark private in .xreview.json never reach it, because whatever it reads goes to its provider. You can allow specific ignored outputs, such as a metrics file, under readable.
• The reviewer cannot change anything. Codex's shell runs under a permission profile that can read only the review copy (not your home folder, other repositories or /tmp), with no writes or network, and without your Codex config, so without your MCP servers, plugins, hooks and connected ChatGPT apps (GitHub, Gmail, Drive, Slack...). Claude Code runs in the copy with only Read, Grep and Glob, and without settings files, hooks, plugins or MCP servers.
• Reviewer models: Codex reviews with GPT-6 Astra at medium effort, and Claude Code with Opus 5.5 at high effort. Either can be changed per machine (configuration).
• Reviews run in the background, so no agent's command timeout kills them. Answers must match a JSON schema. A crash, timeout, malformed answer or cannot_review is never treated as an approval.
• Built for research code too. Besides correctness, the reviewer checks for data leakage, train/validation/test separation, checkpoint and hyperparameter selection, metric correctness and baseline fairness.

Install

You need both CLIs logged in on the machine where the agents run: codex login and claude auth login. The desktop apps' sign-ins do not cover the command-line tools.

Claude Code, so that Codex reviews Claude's work:

/plugin marketplace add Armin-Saadat/xreview
/plugin install xreview@xreview --scope user

On Claude Code 2.1.275 or later, /plugin install xreview --marketplace Armin-Saadat/xreview does both in one step.

For project, declare the marketplace in the project as well, so collaborators who open the repository are offered both:

/plugin marketplace add Armin-Saadat/xreview --scope project
/plugin install xreview@xreview --scope project

Codex, so that Claude Code reviews Codex's work:

codex plugin marketplace add Armin-Saadat/xreview
codex plugin add xreview@xreview

That enables it for every project. To use it in some projects only, turn it off in ~/.codex/config.toml and on in each project's .codex/config.toml, which you can commit. Codex applies project settings only in projects you have trusted.

# ~/.codex/config.toml
[plugins."xreview@xreview"]
enabled = false

# <project>/.codex/config.toml
[plugins."xreview@xreview"]
enabled = true

Collaborators still install the plugin once themselves.

Check each machine once. Ask your agent to "run the xreview doctor with --live", or run it from a clone of this repository:

python3 skills/xreview/scripts/xreview.py doctor --live

It checks both CLIs and their logins, that Codex's sandboxed shell can read the review copy and nothing else, and runs one real review per reviewer. That review uses a little of each account's quota.

Without a plugin system, copy skills/xreview/ into ~/.claude/skills/ (Claude Code) or ~/.agents/skills/ (Codex) for every project, or into a project's .claude/skills/ or .agents/skills/.

Use

Work as usual. When the agent finishes something substantial, it starts a review, waits, checks every finding against the code, fixes what holds up and reports the verdict, the fixes, any disputed findings and the path of review.md. To ask for a review yourself, use /xreview:xreview in Claude Code, $xreview in Codex, or just say "get a second opinion on this".

After fixes, the agent reviews again with --followup, so the reviewer first checks its previous findings; findings that predate the change are reported separately and never block. --paths reviews one piece of a larger change.

Put a .xreview.md at a repository root to give every review there rules a reviewer cannot infer, such as "the test split never feeds model selection". Put a .xreview.json there to keep data away from reviewers, or to let them read selected outputs:

{"private": ["data/", "splits/*_ids.csv"], "readable": ["outputs/*/metrics.json"]}

readable entries take effect once committed. Models, timeouts, permissions, what is sent, and troubleshooting are in skills/xreview/references/setup.md.

HPC and container hosts. Codex's sandbox cannot start on some of them (bwrap: ... Operation not permitted, or unsupported host mount). doctor reports this, and xreview never runs Codex without its sandbox. Instead, set "codex": {"tools": "repo"} in that machine's ~/.config/xreview/config.json: Codex then has no shell and reads the review copy through xreview's own read-only tools (list, read and search files, plus git history). Claude Code reviews need no sandbox and work there unchanged.

Get updates

Claude Code tracks every commit on main. Codex updates when version in plugin.json changes.

Repository layout

Development

python3 tools/validate.py
python3 -m unittest discover -s tests -v
claude plugin validate .

Read AGENTS.md before changing the skill. CI runs the first two on every push.

License

MIT