Pact — AI skill for Claude Code

pact

A language-agnostic orchestrator for running multiple AI coding agent CLIs (Claude Code, GitHub Copilot CLI, Codex) in parallel on the same repository, without them fighting each other.

**What "without fighting" means here:** each agent gets its own git worktree (real filesystem isolation, not just a hope that they touch different files), conflicts get detected and surfaced instead of discovered the hard way, and `merge-all` gets N agents' work back onto one branch automatically wherever it safely can. The file-lease/messaging layer that lets agents coordinate is advisory, not enforced -- it makes overlapping work visible and communicable between agents, it doesn't stop an agent that never checks it from editing a file another agent already claimed. Worktree isolation and merge-all's conflict handling are the parts that are real guarantees; coordination is a convention agents opt into.

**The full loop, end to end:** isolate each agent in its own git worktree → prepare dependencies before the agent's first command → launch the agent with coordination tools wired in automatically → track every claim, message, and merge in a queryable operation log (`pact history`) → merge completed work back onto one branch, sequenced by risk (small, low-risk changes first) → verify an AI-proposed conflict resolution against a real test command before ever accepting it (Arbiter). Six real subsystems working together, not just a lock server, not just a worktree wrapper, and not just an MCP integration -- coordination is the one most visible from the wire protocol, but it's one piece of the loop, not the whole tool.


*(That's `pact demo`'s actual, real output -- zero-cost, zero-agent-CLI-call, run it yourself in 5 seconds. See Getting started below for the real thing, with real agents.)*

**[Getting started guide](GETTING_STARTED.md)*