Crew splits the work between models, and Koda's runtime checks and validates every result.
/crew-supervisedA senior lead with a junior team. A frontier model defines the work and the acceptance test, a lower-cost model builds it, and the lead steps in only when results fall short.
/crew-self-managedA team that runs itself. One lower-cost model plans, assigns, builds and signs off. No premium model is involved, so spend stays predictable.
Checked and validated by Koda's runtime, never on a model's word.
Everything a crew uses is chosen once, on one page: Preferences → Crew Workflow. Each workflow has its own section. Until its models are chosen, the command opens that page and runs nothing.
| 1. Open the page | Press ⌘, (Ctrl+, on Windows and Linux) and pick Crew Workflow, or type a crew command before its models are set and the page opens for you. |
| 2. Choose a worker | Pick a provider, then the model that will do the work. Any low-cost hosted model or a local one works. Each workflow has its own worker model. |
| 3. Choose a leader | Supervised only: the Leader model that plans each task and reads the verdicts. Self-managed needs no leader. See the two workflows above. |
| 4. Add a backstop | Optional, Supervised only. A stronger model gets one attempt when the worker's fixes fail, and a review model checks the test and the diff before anything is applied. |
| 5. Set a budget | Supervised: a cap per task in dollars, or in tokens for a local model with no price. $0.50 by default. Self-managed is bounded by its worker rounds. |
| 6. Start a crew | Type /crew-supervised or /crew-self-managed and the task in any session. Settings save as you change them. |
A worker reads, searches, edits and runs commands inside its copy, and asks you when it is stuck. Other tools are refused, and so are edits to .git, .koda, lockfiles and linked folders. The leader can't touch code at all.
The copy is part of the project, so with Changes in project set to Ask before each, every worker edit waits for you, with the file shown in the copy, and the crew's clock pauses while it waits.
For a bug fix, the reproduction test must fail before the fix and pass after it. A test that lost lines after it failed is flagged in the verdict.
Set a Review model and a fresh copy reads the task, test and diff after the check passes. Only a first line of PASS lets the work be applied.
A budget per task in dollars, or tokens for a local model with no price, and a limit on retries. What each model spent is on the settings page.
Up to three crew jobs run together by default; more wait in a queue and start as others finish.