---

name: config-hyperparams description: "Read, write, and statically validate RL Zoo hyperparameter configs and CLI override grammar." disable-model-invocation: true metadata: disco-role: operating license: MIT

config-hyperparams

Use this sub-skill when the task is about RL Zoo hyperparameter configuration, not about running a training job. It is the routing point for reading, editing, linting, or explaining YAML/Python config files and StoreDict-style CLI override strings.

Use this for

• Inspecting or editing algorithm hyperparameter config files such as a2c.yml, ppo.yml, sac.yml, td3.yml, tqc.yml, trpo.yml, or a custom YAML file.
• Checking default, atari, and environment-specific entries before a training command uses --conf-file.
• Translating between YAML config files and Python config files/modules that expose a top-level hyperparams dictionary.
• Validating normalize, schedules, policy_kwargs, env_kwargs, monitor_kwargs, env_wrapper, vec_env_wrapper, callback, frame_stack, and n_timesteps shapes.
• Explaining --hyperparams, --env-kwargs, and --eval-env-kwargs key:value override grammar and its eval trust boundary.

Do not use this for

• Launching, resuming, or benchmarking actual training/evaluation runs; hand off to the training-cli or evaluation-and-artifacts sub-skill.
• Designing Optuna search spaces or interpreting study artifacts; hand off to the tuning-optimization sub-skill.
• Implementing or debugging wrapper/callback class internals; hand off to the custom-components sub-skill.

Runtime references

• Configuration semantics covers loader precedence, special keys, schedules, normalization, wrapper/callback shapes, and StoreDict override parsing.
• Hyperparameter files covers algorithm file families, reusable YAML/Python patterns, and concrete examples.
• Troubleshooting covers missing fallback entries, indentation and quoting issues, eval trust boundaries, env-id/package mismatches, and normalization/vector-wrapper interactions.
• Static validator checks config shape without training, network access, or mutation. Python configs are statically inspected by default; execute/import them only with --import-python when the caller explicitly accepts that trust boundary.

Suggested workflow

1. Identify whether the input is YAML, a Python file, or an importable Python module containing hyperparams.
2. Resolve the target --env and whether the runtime should choose an exact entry, atari, or default fallback.
3. Use the configuration and hyperparameter-file references to edit or explain values.
4. Run the bundled validator before composing a training command, especially after changing wrapper/callback lists or policy_kwargs strings.
5. If validation requires actual environment registration, class importability, or a real run, route to the appropriate sibling sub-skill instead of training from here.