---

name: atari-breakout-pong description: "Guides standard Atari Breakout and Pong DQN/PPO workflows with preprocessing, devices, checkpoints, W&B logging, and safe synthetic smoke checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

Atari Breakout/Pong DQN and PPO

Use this sub-skill when a task is about the repository's standard Atari Breakout/Pong workflows: DQN, PPO, ALE/Gymnasium preprocessing, Nature CNN models, DQN replay buffers, PPO vector rollouts, device choice, W&B logging, checkpoints, --test, or benchmark interpretation.

Do not use this sub-skill for Montezuma's Revenge, Pitfall, PrivateEye, RND, Go-Explore, restore-state search, or robustification curricula; route those to the hard-Atari exploration owner instead. Do not use it for CartPole PPO/DQN or GridWorld formulas.

Start here

1. Read preprocessing-and-devices when the user asks about --env breakout|pong, --device auto|cpu|cuda|mps, wrappers, frame stacking, ROM/display requirements, or per-life versus per-game returns.
2. Read algorithm-and-run-guide for the DQN and PPO workflow contracts, checkpoint names, command-shape examples, training constants, W&B behavior, and benchmark caveats.
3. Read troubleshooting before diagnosing ROM install failures, W&B login/network issues, missing checkpoints, device errors, replay-buffer sampling errors, headless rendering, or unexpectedly different scores.
4. Run or inspect scripts/atari_basic_smoke.py for a safe synthetic check of the Nature CNNs, DQN replay stacking/masking, PPO GAE, and tiny gradient updates. The helper intentionally avoids ALE env construction, ROM downloads, W&B credentials, and long training.

Runtime boundaries

• Source workflow labels: 3-atari/1-dqn.py is the Atari DQN workflow label; 3-atari/2-ppo.py is the Atari PPO workflow label; 3-atari/env.py is the preprocessing/device helper label. Treat these labels as provenance only, not as files this sub-skill needs future agents to open.
• This sub-skill is self-contained for guidance and smoke verification. If a user explicitly asks to run an actual repository checkout, apply the CLI contract documented in the references to that checkout's Atari entrypoints.
• Full Breakout/Pong training is expensive: the documented training budget is 10M agent steps, and real env reset requires Atari ROM availability. Prefer the bundled smoke helper when the task is only to validate model/replay/GAE logic or environment-independent usability.

Quick synthetic check

From this sub-skill directory:

python scripts/atari_basic_smoke.py --help
python scripts/atari_basic_smoke.py --device cpu

Use --device auto only when you want the helper to choose CUDA, then MPS, then CPU using the same priority as the Atari workflows.