cc-pair Has Claude Code and Codex Review Each Other's Code by Trading Numbered Files

Leon Lee's open-source pair command lets one AI coding agent write a change and the other review it, swapping requests, reviews and fixes as Markdown files in a shared folder until the reviewer approves or three rounds pass and both stop to ask you.

By Chris Pirillo · October 31, 2026 · 4 min read

Chris Pirillo at a chessboard reading a postcard with the next move, a tray of outgoing and incoming postcards beside him — Illustration: LockerGnome

The setup works in either direction. Claude Code can write while Codex reviews, or the other way around, and the README calls the roles “symmetric.”

Each review is a thread of files. The author’s request lands as 001-request.md in a lane like claude-to-codex/, the reviewer answers in 002-review.md with a verdict, the author replies with fixes in 003-response.md, and so on. Every git worktree gets its own folder, so two branches don’t mix their reviews.

There are two ways to run it. In auto mode, on macOS or Linux, you type pair start --as claude in your own terminal and get “interactive Claude, automatic Codex reviewer”: each time Claude sends a request, pair starts Codex for one review round and hands the review back inline. Each round has a 900-second limit by default. In manual mode you open both agents in separate terminals and say “be my reviewer” to one of them. Codex in that mode is woken by pair’s MCP server instead of checking for new files on a timer, because “a polling pair wait costs a model call per check.”

The README ends on a real exchange, with a bug planted on purpose. In round 1, Codex flagged that totalPages(5, 2) returns 2 and wrote “Use Math.ceil(count / pageSize).” In round 2 it approved: “I re-read src/paginate.js and ran 41 assertions … All passed.”

The project went up October 7, 2026, is at version 0.7.4 on npm, and had three GitHub stars when we checked. The README doesn’t say how many tokens a typical three-round review costs on each side.

Why it matters

The reviewer reads the code cold. In the README’s example, the reviewer caught that a partial last page went missing from the page count, then re-ran its own checks before approving.

The review trail is plain text. Every request, finding and fix sits as a numbered Markdown file, so you can read the whole argument afterward without either agent’s chat history.

The loop has a stop. After PAIR_MAX_ROUNDS rounds, three by default, the thread escalates to you instead of letting two agents argue forever.

The reviewer’s limits come from instructions. Codex’s source edits as a reviewer “are forbidden by its instructions, not by the OS,” and the README says its guards work “against mistakes and casual prompt injection, not a determined agent.” Auto mode also “passes neither CLI’s permission-bypass flags to reviewers.”

Steal this

Give both sandboxes write access to the channel first. Add ~/.pair to Codex’s writable_roots and to Claude Code’s sandbox allowWrite before the first run, as pair init prints. This is like giving both roommates a key to the mailbox before you start leaving each other notes.
Cap what each review round can spend. Pass --reviewer-args="--max-budget-usd 2" when Claude is the reviewer, as the README’s example does, so a long review can’t run up the bill. Kind of like when you hand a teenager a gas card with a limit on it.
Make the author check every finding. In manual mode, have the author verify each item and either fix it or reject it with a reason, the way the README’s steps describe. This is like a second opinion from a doctor: you still ask why before you schedule the surgery.

Tools like this turn two competing AI coding products into coworkers on the same branch. The design leans on files instead of a chat server, which keeps the record readable and the moving parts few. The open question is cost and time: every round starts or wakes a second agent, and the README has no figures yet for how much that adds to a day of work.

Search

cc-pair Has Claude Code and Codex Review Each Other's Code by Trading Numbered Files

More AI