Choose the model for a task
Name a model in plain words and that one task runs on it; pin a seat for every task; or say nothing and let the crew be picked for you.
You should be able to pick a model the way you would pick a colleague: by name, for one job, without touching settings.
- run the refactor on kimi k3
- a quick count on glm 5.3 flash
- every task's worker pinned for a week
Most of the time you should not have to think about models: CodeAF picks a crew for every task. When you do know which model you want for a job, say so in the sentence that asks for the work. There is no flag to remember and nothing to change in settings.
Try it in CodeAF
- Where
In any conversation.
- Type
Run this as a task on kimi k3: count the lines in taskboard/store.py and report the number.Then press Enter.
- You see
The task is proposed and starts. The chat says
Task 1 started on moonshotai/kimi-k3.
What the name changes
The chat turns your words into the catalog's model id, so the short name you would say out loud is enough: kimi k3 became moonshotai/kimi-k3, and glm 5.3 flash became z-ai/glm-5.3-flash. The named model takes the worker seat, the one that does the work, for that task only. The other seats are still picked for you, and the crew line says so, with the pin mark on the seat you chose:
task 1 crew · open-ended · worker ⌖ kimi-k3 (openrouter) · planner glm-5.3 · checker claude-opus-5.5
The task's own page names the model in its head, for example done · 9s · $0.0017 · z-ai/glm-5.3-flash. Your conversation stays on its own model, and the next task is picked the ordinary way again.
For every task: pin a seat
To keep one model in a seat until you change your mind, type /crew pin worker z-ai/glm-5.3-flash. The /crew panel then shows worker ⌖ glm-5.3-flash, and every task uses it. /crew unpin worker hands the seat back to auto. The planner and the checker can be pinned the same way.
When you say nothing
Each seat stays on auto, and /crew shows the model it is likely to get, such as worker auto · likely glm-5.3. The crew is chosen per task from what kind of work it is, weighing quality against price.
Things to build
- A model bake-off: send the same brief twice, "as a task on kimi k3" and "as a task on glm 5.3 flash", and compare the two task pages: result, time and cost.
- Pennies for lookups: "Run this as a task on glm 5.3 flash: count ..." for small questions about the code; our run cost $0.0017.
- A trial week for a new model:
/crew pin worker <model>puts it on every task until/crew unpin worker. - Security work on a chosen model: a rule, "security fixes in this repo always run on openai/gpt-oss-120b", and you never have to say it again. Rules
Go further
Pareto crewing: how the crew is picked when you name nothing.
Coming soon: describe the org you want and CodeAF builds it: teams, managers, budgets, standing orders.