How it works
How it works
The Planner writes the PRD. You accept it. The Conductor runs each level of the DAG in parallel while observers watch the bus. A successful run ends in a pull request.
- Step 1
Plan
The Planner explores the target directory and turns your goal into a plan: a PRD of stories with acceptance criteria, test commands, retries and dependencies. Cycles are rejected.
- Step 2
Review
The plan opens on the Review screen. Refine it with feedback, accept it, or quit.
--dry-runsaves the PRD toprd.jsonwithout running it. - Step 3
Run
The Conductor runs the DAG one level at a time. Every story in a level gets its own story agent on the run branch; each agent is told to build the project before it commits.
- Step 4
Finish
At the end the TUI shows time saved by running in parallel. The Finalizer opens one pull request per successful run unless you pass
--no-pror--no-push.
When a story fails
A story gets its retries first. When it fails for good, the Surgeon asks a model (opus by default) for a replan: split the story, insert a prerequisite, rewire dependencies, or skip it. The Conductor applies it at the next level boundary.
A run counts as successful only when every original story passes. If a story is dropped without a replacement, the run ends with a "did not complete the goal" verdict. --no-surgeon-llm keeps a deterministic, skip-only Surgeon; --no-surgeon turns replans off.
Run a whole goal as parallel agents
cars runs on your machine, in your repository, with your Claude CLI login. There is no hosted service and no account.