← All projects

Project

gitaccord

Git hooks that help parallel AI coding agents coordinate before they land changes.

Overview

gitaccord is a Git-native coordination protocol for teams of AI coding agents working in parallel. It helps agents share their plans and reasoning, recognize overlapping changes, and avoid silently overriding another agent's work.

Rather than introducing a new command-line workflow, gitaccord uses ordinary Git branches, lightweight coordination files, and Git hooks to make collaboration visible and enforce agreements before changes are pushed.

How it works

  • Each agent works on its own branch and records a plan.
  • Agents exchange notes about overlapping work and maintain a shared project direction.
  • Commit messages include the reasoning behind each change.
  • Git hooks surface overlaps and enforce an agreement before a change is pushed.
  • Overrides are explicitly declared for human review.
  • Each agent's acceptance tests help protect the behavior it owns.

Experimental results

In approximately 300 test runs with Claude Code (Sonnet) and pi (GLM 5.3 Flash), the project reported that silent feature overrides fell from 24–28% with plain Git to 0% with gitaccord. The spike also detected all designed clashes in those tests.

These are results from the project's current experimental test setup, not a claim of production-wide guarantees. Token overhead and hook latency under heavier multi-agent load remain areas for improvement.

Status

The design has been approved and the initial spike is complete; implementation is starting. The repository includes the design specification, prototype, and test harness.