What it does

twoperson sets up a two-person rule for AI coding agents. A builder agent writes a review packet, a JSON record of its goal, head commit, changed files and test results, and publishes it to a shared .twoperson inbox directory. A reviewer agent claims the packet and writes a verdict. There is no server, API key or network involved; the two agents just share a directory. The rule lives in a schema rather than a prompt. A verdict must answer a packet that exists, an approval must name the commit that packet is at, and a report of work as pushed, deployed or restarted is refused without a reference to an approving verdict for that same commit. Test changes must be acknowledged by path, and the changed-files list is recomputed from git. Packets are treated as untrusted input, with a size cap, a strict schema and a scan for credential-like values. It installs with pip. A Stop hook installer adds a signal to Claude Code settings, a macOS launchd watcher can run a reviewer command when a packet arrives, and the inbox is shared across git worktrees. The README is explicit that it does not review code, spawn agents or verify who the reviewer is.

Capabilities

Review packets recording goal, head commit, changed files and test evidence
Approvals bound to one commit, so a rebase or amend makes them stale
Ship reports refused without a matching approving verdict
Strict packet schema with a credential scan
Claude Code Stop hook installer and a worktree-aware shared inbox
Difficulty tiers from low to critical to decide which reviewer sees a packet

When to use it

You run one coding agent to build and another to review the same repository
You want a review recorded against an exact commit instead of relying on a prompt instruction
Your agents work in separate git worktrees and need to share one review inbox