PufferLib RL Guide

What it does

PufferLib speeds up reinforcement-learning simulation and training. This skill keeps you on one explicit version profile at a time — the published PyPI release 3.0.0 versus the redesigned upstream 4.0 source — so you never mix incompatible imports, configs, and CLI flags.

• Correct wrapping of Gymnasium / PettingZoo environments via GymnasiumPufferEnv and PettingZooPufferEnv
• Vectorization workflow: start on Serial, then move to Multiprocessing, with a checklist for shapes, masks, and agent slots
• Policy contracts for Torch modules, recurrent composition, and nativize_dtype usage
• A dry-run workflow that emits a strict-JSON training plan and validates it before anything executes
• Checkpoint inspection by hash and metadata only, never calling torch.load
• Hard rules for W&B / Neptune credentials: named env vars only, never in args, configs, or logs

Every bundled CLI is standard-library only, CPU-only, synthetic, and network-free.

Who it's for

• RL researchers and engineers running large-scale environment simulation with PufferLib
• Teams porting existing Gymnasium or PettingZoo environments into PufferLib
• Security-conscious setups that must vet third-party checkpoints, ROMs, or environment packages
• MLOps folks who care about reproducibility: separate seeds, pinned commits, committed lockfiles

Examples

1. Validate an environment contract: run python3 scripts/env_contract_validator.py --steps 64 --episodes 8 --seed 42 to check spaces, dtypes, finite rewards, seeding, and the terminated vs truncated distinction.
2. Plan before you train: scripts/train_template.py --profile pypi-3.0.0 --device cpu --total-timesteps 10000 emits a config plan without launching, and validate_plan.py rejects mixed-version options, unbounded resources, dotted env paths, and coupled train/eval seeds.
3. Vet a downloaded checkpoint: scripts/inspect_checkpoint.py checkpoint.pt --expected-sha256 ... hashes and classifies the file without unpickling, so you only load it in a sandbox after the digest matches.

Install the skill I found on Claude Skill Mart.
Copy the skills/pufferlib folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/pufferlib/.
When it's done, tell me in one line what this skill can do.