Microduck RL

Close the loop: spec → sim → package → Hub → robot. You are training a 61→14 PPO policy, not an LLM. Hub publish is not the end — deploy is in scope. A published repo is a deliverable for whoever owns a duck, not a baked-in official weight.

Canonical playbook: AGENTS.md in pollen-robotics/microduck_rl. This skill is the job ticket; that file wins if they disagree.

$SKILL = directory that contains this SKILL.md. Always invoke scripts with that absolute path. The skill id is microduck-skill.

Preconditions

1. Run $SKILL/scripts/doctor.sh (add --clone --sync if there is no checkout). It must print doctor ok.
2. Default compute is Hugging Face Jobs. Require HF_TOKEN (or hf auth login) and MICRODUCK_HF_NAMESPACE (user or org). No namespace → Jobs prompts and the agent hangs.
3. Local train only if MICRODUCK_TRAIN_LOCAL=1 and CUDA is present. Smoke uses the same switch.
4. Stop if doctor fails. Do not invent another trainer (Isaac Sim, AutoTrain, Colab-only hacks).

Hard rules