LoRA adapter for a model-routing policy with leases. Given anonymized candidate models, evaluation scores, projected costs, and a coding-agent conversation, the policy picks one candidate and a commitment length: {"action":"B","turns":10}. The caller holds that selection for the committed turns before consulting the policy again.

This is not a general-purpose chat model. It requires the Dari selector prompt, anonymized candidates, and constrained decoding over the exact action × lease-length completion set; unconstrained generate() is not a supported interface.

Breaking change from dari-ai/router-slm: v1 emits bare {"action":"B"} every turn. v2 adds the lease commitment (menu: 5, 10, or 30 turns) and expects cost projections at those horizons in its input. Neither direction is drop-in compatible.

Pin the tag, not main. v40 (the final training checkpoint) is the recommended default.

Terminal-Bench 2.1 (89 tasks, not seen in training), routing a mini-SWE agent among 8 candidates, single attempt per task:

Rank-32 LoRA on Qwen/Qwen3.6-35B-A3B, trained with episodic GRPO on live SWE-bench coding-agent trajectories (reward: task success minus small cost and thinking-length penalties). Inference adapter only — no optimizer state or trajectory data.

Adapter weights are Apache-2.0. Comply with the pinned Qwen base model's terms. The surrounding router remains responsible for candidate eligibility, lease bookkeeping, and rejection of invalid policy output.