What is pair?
pair is a command-line tool that coordinates code review between Claude Code and Codex, or between agents using the same vendor. One agent makes a change and another reviews it through numbered messages exchanged in a shared directory. Developers can choose an automatic workflow, where pair launches the reviewer for each round, or a manual workflow, where they run both agents themselves. It also supports reviews of plans and design documents, as well as brainstorms before a design exists.
pair Features:
Automatic Reviews: Run an interactive author session while pair launches a reviewer for each round and returns the review inline.
Manual Reviews: Run both agents yourself and exchange requests, reviews, and responses through files in a shared channel.
Review Verdicts: Reviewers provide numbered findings with severity and an approve-or-changes verdict, with unresolved findings escalated after the round limit.
Spec and Brainstorm: Ask agents to review plans and design documents in different formats, or brainstorm ideas with a fixed turn budget and closing summary.
Parallel Worktrees: Run separate review loops for subagents in their own worktrees, with support for multiple reviews at once.
Session Recovery: Resume author conversations, retry failed rounds with requests kept on disk, and inspect history or reset a channel.
pair Benefits:
Independent Feedback: A second agent can examine a change before it is considered finished, helping authors catch issues they may have missed.
Clear Decisions: Numbered findings and explicit verdicts make it easier to see what needs attention and whether a review is complete.
Fewer Manual Handoffs: Automatic mode handles reviewer launches and returns feedback in the author's session, reducing the need to coordinate separate review rounds by hand.
Parallel Progress: Separate worktree channels let teams keep reviews for different branches apart while several subagent tasks proceed at once.
Recoverable Work: Saved requests and resumable conversations make it possible to continue after a failed review round or interrupted session.
Use Cases:
Code Change Review: Have Codex inspect a change authored with Claude Code, or use a same-vendor pairing to review an implementation before merging.
Plan Validation: Ask a second agent to examine a Markdown plan or an OpenSpec change before beginning implementation.
Early Brainstorming: Explore possible approaches with an agent pair when a project has ideas but no settled design yet.
Feature Branches: Start a review session in a dedicated worktree so feedback on a feature branch stays separate from other work.
Subagent Reviews: Serve subagents' worktrees and review several parallel tasks while keeping each task's review loop in its own worktree.

