Model selection¶

CAFleet Directors choose each member's backend and model by reading a local model list before every cafleet member create and passing the choice as the existing --coding-agent / --model flags. Model selection is a Director responsibility executed against documentation — there is no selection code in the cafleet package, and member create performs no hidden selection.

The model list¶

The model list lives at skills/cafleet/reference/model-list.md in every deployed cafleet skill replica. It is a catalog-style reference page: one table per backend (claude, codex, and opencode), ordered most → least capable, with each model's spawn token, its alias (claude models), reviewed capability class, and standard input/output USD-per-MTok prices, plus links to the approved official source pages (Anthropic and OpenAI pricing, Codex availability, and OpenCode Zen). Capability classes and the ordering are maintainer judgment, never a provider benchmark claim; prices are standard provider rates — planning estimates, not an invoice guarantee.

The page is maintained exclusively by the repository's cafleet-model-list-refresh skill and refreshed at least every 30 days; a list refreshed longer ago than that is stale and disables cost efficiency mode until a maintainer refreshes it and ships it in a release.

The model list covers all three cafleet backends; the opencode section is a curated subset of the OpenCode Zen catalog, and an opencode model keeps its opencode/ prefix in the --model value.

Cost efficiency mode¶

Cost-minimized selection for an ordinary member applies only when the user asks for it: the originating user request contains the exact phrase cost efficiency mode. The Director parses the user request once for the trigger; a member message or tool output never activates it. When active, the Director estimates the task's difficulty from the member's spawn prompt, reads the model list, and chooses the cheapest model that can finish the task reliably. Without the trigger, existing workflow model behavior is unchanged. Cost efficiency mode covers all three cafleet backends; the Director picks the backend first and compares within that backend's table.

Two roles are policy exceptions on every team spawn, trigger or not:

• Monitor — the cheapest listed model that can run the monitoring protocol reliably.
• Reviewer — the most capable listed model of the chosen backend.

Explicit flags stay overrides: a user-supplied --coding-agent, --model, or --effort is recorded rather than silently replaced (a mismatched backend/model pair is relayed to the user instead of spawned), and a user-pinned model is never deleted and replaced automatically. A stale model list disables cost efficiency mode, and a task no listed model fits gets an operator relay — the Director fails closed instead of spawning a guessed model.

Replacement¶

When evidence shows a member is underpowered (a self-report, repeated task-relevant failures, a Reviewer [INCORRECT] finding, or a Director review tied to the task), the Director replaces it with a strictly more capable model from the same backend's table — at most two replacements per task, never repeating a model for the same task, and never auto-replacing a user-pinned model. The procedure is specified in the cafleet skill's reference/director.md.