Agent Pair — Development skill for Claude Code

agent-pair

A skill that lets Claude Code and OpenAI Codex CLI review each other's code. It works from either side.

The agent you are using is the **host**; the other one is the **peer**, called in the background through its CLI. There are two ways to split the work:

Mode	Writes code	Reviews
A: I write, peer reviews	host	peer (read-only)
B: peer writes, I review	peer	host (does not change code)

A reviewer who is not the author is far more likely to spot what the author cannot see. That is why the roles are kept strictly apart and the reviewer never fixes code on the side.

Features

• One session across rounds: continued via codex exec resume / claude -p --resume, so the peer remembers earlier feedback and each round only needs the delta.
• Resumable and clearable sessions: every run reports the peer session id, which you can continue later with --resume; --clear deletes every peer session the current session created in one go.
• Tight permissions: the reviewer is read-only; the writer can only edit workspace files and cannot commit.
• Recursion guard: the peer runs with AGENT_PAIR_PEER set, so it never calls back into the host.
• Replies in your language: the skill is written in English, but the host always talks to you in the language you use.

Requirements

• git, bash
• The peer's CLI, logged in:
  • from Claude Code: codex with codex exec resume support; --clear also needs codex delete
  • from Codex: claude
• jq is recommended; python3 is used to parse JSON when it is missing

Installation

Claude Code:

git clone https://github.com/natsu1211/agent-pair.git ~/.claude/skills/agent-pair

Codex: put it in Codex's user-level skill directory. If you use both, clone once and symlink:

mkdir -p ~/.agents/skills
ln -s ~/.claude/skills/agent-pair ~/.agents/skills/agent-pair

Update: `git -C ~/.claude/skills/agent-pair pull`

Usage

Just ask in