rlab

Tune one checked-in rlab SB3 PPO or A2C recipe from durable training completion signals without launching checkpoint evaluations. Use when the user points to a recipe and asks to tune, optimize, autoresearch, improve sample efficiency, maximize training return, find the best hyperparameters, or make training behavior stable across seeds. Runs a bounded fixed-rung beast-3 search, confirms the winner on five untouched training seeds, and patches only the pointed leaf recipe.

Launch, continuously monitor, and when explicitly authorized repair queue-backed rlab training from a checked-in goal and recipe, defaulting to beast-3. Use when the user asks to launch, run, start, execute, follow, watch, monitor, diagnose, fix, or harden an rlab training recipe, run, or research goal. Stay attached through terminal events, report W&B URLs immediately, preserve failed-run evidence, and return final statistics, the best artifact, and the exact playback command.

Start and execute a persistent repair-mode goal that stress-tests the complete rlab queue, fleet, training, checkpoint-evaluation, W&B publication, artifact, playback, and cleanup lifecycle. Use when the user explicitly asks to harden, stress-test, repeatedly exercise, or fix the fleet experiment lifecycle until it is reliable, especially with bounded SuperMarioBros-NES PPO runs across Level1-1 through Level1-4.

Composite rlab release workflow for trained checkpoints. Use when the user asks to upload, publish, release, or promote a checkpoint/model to Hugging Face, especially RL checkpoints with gameplay preview videos. Coordinates Hugging Face model-card publishing with video preview requirements and YouTube upload of the same preview video.

Upload or update YouTube model preview videos from rlab, especially RL gameplay previews associated with Hugging Face model cards. Use when the user asks to upload a local video to YouTube, update an existing video description/title/tags/privacy, add a video to a playlist, produce OAuth authorization flow instructions, or ensure model-card preview videos have clean clickable model links.