About this project
Commitarium provides a local-first environment for orchestrating coding agents that plan, implement, and review software together. The system is built around a coordinator that manages projects, features, workflow transitions, and agent sessions, storing operational state in SQLite. It supports both a deterministic simulation mode for testing and an opt-in mode using real AI agents (Codex or Claude).
In the real-agent mode, the coordinator connects to persistent conversations with selected lead and reviewer agents. The workflow involves goal clarification, planning, implementation, and review cycles. The system integrates with Forgejo, a self-hosted Git forge, to serve as the source of truth for plans, reviews, and pull requests. Each agent operates with its own scoped Forgejo identity, ensuring an audit trail of commits and structured PR work. The coordinator verifies agent outputs, such as commit hashes and PR heads, before advancing the workflow.
The architecture typically runs via Docker Compose with services for the coordinator, Forgejo, and worker agents. It supports isolated provider profiles for Codex and Claude, allowing independent authentication and configuration. The system includes features for restart recovery, where active sessions can be resumed after a coordinator restart. It also offers project import capabilities, allowing users to associate existing Git repositories with the workspace. The design emphasizes user control, with options for automatic or manual approval of merges and workflow steps.
Comments
0 Rating appears after 10 ratings
Sign in to join the discussion.