Core concepts
The pipeline
The default agent flow, its branches, and its revision loop.
Every project ships with a default pipeline — a graph of agent steps with conditions on the edges. The default flow is:
Planner → Developer → Reviewer → QA → Requirements → Merger (end)With branches back to the Reviser:
- Reviewer approves → continue to QA; requests changes → Reviser.
- QA passes → continue; fails → Reviser.
- Requirements met → merge; not met → Reviser.
- Reviser applies fixes and loops back to the Reviewer (up to a max, then escalates).
The steps
| Step | Agent | What it does |
|---|---|---|
| Planner | plan | Breaks the ticket into per-repo tasks. |
| Developer | execute | Writes code and opens a PR (Claude Code). |
| Reviewer | review | Code review — approves or requests changes. |
| QA | qa | Runs your configured tests/build; pass or fail. |
| Requirements | requirements | Checks the PR actually solves the ticket. |
| Reviser | revise | Applies review/QA/requirements feedback, loops back. |
| Merger | merge | Merges the PR (the end state; auto-merge is off by default). |
| Rebaser | rebase | Resolves conflicts when the branch gets new commits. |
Triggers
GitHub events can drive steps directly. By default:
- branch pushed → Rebaser
- PR comment → Reviser (address the comment)
Customizing it
The pipeline is fully editable in the visual pipeline editor — add steps, rewire branches, change the loop, and wire triggers. You can also manage it as code with the CLI.