Crew Workflow
Frontier-level coding results from cheap or local models. /crew-supervised and /crew-self-managed hand one task to them in a private copy of the project. Koda runs the checks itself and applies work that passes to the working tree, uncommitted.
Supervised
Used by /crew-supervised. Without a leader and a worker model, it opens this page and runs nothing.
Leader model
Plans each task, reads the verdicts and decides what to do after a failure. Every leader turn runs on it, whatever model the session uses.
Worker model
The cheap model that reads the project and writes the code.
Reasoning
How hard the worker model thinks. Empty leaves its default.
Stronger model
One attempt on this model when the worker's fixes fail.
Review model
Optional. Reviews the task, the test and the diff once the check passes, and can only turn an OK into not verified.
Budget per task
What one task may spend. A model with no price offers tokens or none.
Retries after a failure
How many narrower retries the leader may send for a task that failed. 0 reports the failure at once.
Skills
Lets the worker and research read your skills. The leader never does.
Memory
Lets the worker and research use the project's memory: recall past fixes and store new ones. The leader never does.
Self-managed
Used by /crew-self-managed. Without a worker model, it opens this page and runs nothing.
Worker model
The cheap model that plans, manages and writes the code.
Reasoning
How hard the worker model thinks. Empty leaves its default.
Worker rounds
The most rounds that run before a last worker finalizes the work.
Open tasks
The longest task list kept.
Skills
Lets the worker and ideation read your skills. The manager never does.
Memory
Lets the worker and ideation use the project's memory: recall past fixes and store new ones. The manager never does.
Project
How a private copy is made.
Dependency folders
One per line. Linked into each copy instead of installed. Build output is never linked.
While it runs
Parallel jobs
The most crew jobs running at the same time, across all sessions. More wait in a queue and start as others finish.
Recent crews and spend
Crew jobs in this project that wait for your decision, and what each model has spent.
Defaults and config keys
Supervised settings are in [crew.supervised] in config.toml:
| Setting | Default | Key |
|---|---|---|
| Leader model | Not chosen. Required. | leader |
| Worker model | Not chosen. Required. | worker |
| Reasoning | Model default | reasoning |
| Stronger model | None | stronger |
| Review model | None | review |
| Budget per task | $0.50. Tokens default to 1,000,000. | budget = dollars, tokens or none. budget_dollars, budget_tokens |
| Retries after a failure | 1 | retries |
| Skills | On | skills |
| Memory | Off | memory |
Self-managed settings are in [crew.self_managed]:
| Setting | Default | Key |
|---|---|---|
| Worker model | Not chosen. Required. | worker |
| Reasoning | Model default | reasoning |
| Worker rounds | 10 | rounds |
| Open tasks | 12 | open_tasks |
| Skills | On | skills |
| Memory | Off | memory |
Settings for both workflows are in [crew]:
| Setting | Default | Key |
|---|---|---|
| Dependency folders | node_modules, .venv, venv, vendor | dependency_folders |
| Parallel jobs | 3 | parallel_jobs |
A dependency folder cannot be build output (dist, build, out, .next, .nuxt, target, coverage), an absolute path or a path with ... See Crew.