Fleet

Fleet wires an entire per-role routing configuration in one transaction.

python3 scripts/mlx-agent fleet render --path ./router.yaml \
  --assign coding=mlx-community/Qwen3-32B-4bit \
  --assign vision=mlx-community/Qwen3-VL-8B-Instruct-4bit
python3 scripts/mlx-agent fleet apply --path ./router.yaml --from-adoption ./adopt.json
python3 scripts/mlx-agent fleet apply --path ./router.yaml --from-adoption ./adopt.json --confirm --preview-hash <hash>

Inputs

• --assign role=repo (repeatable) for explicit picks, or
• --from-adoption <state.json> to consume a completed adopt handoff's recommendations.

Roles are canonical (general, coding, reasoning, vision, embedding, tool-use); each at most once. Vision defaults to mlx-vlm (:8083), text roles to mlx_lm (:8080); --runtime-map role=runtime overrides.

Guarantees

• Renders one bounded LiteLLM router YAML with a managed header and an exact five-line entry grammar per role.
• Every referenced model is checked against local inventories; --allow-missing records a reviewed warning instead of failing.
• Applies through the identical preview-confirm-receipt-rollback transaction as wire.