---

name: core-extension-and-logging description: "Use keras-rl core fit/test lifecycle, processors, callbacks, logging, utilities, and environment checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

keras-rl Core Extension and Logging

Use this sub-skill when the task is about the shared keras-rl agent lifecycle or cross-cutting extension surfaces rather than a specific algorithm family.

Read when

• You need to call or debug Agent.fit(...), Agent.test(...), compile(...), callback wiring, verbose, visualize, nb_max_episode_steps, action_repetition, or start_step_policy.
• You are writing a custom Processor, custom Env, old-Gym API adapter, multi-input observation processor, reward/action transformer, or whitening normalizer.
• You need training logs, JSON log visualization, model checkpoints, WandbLogger, FileLogger, TrainIntervalLogger, TestLogger, or Visualizer.
• You need clone_model, clone_optimizer, huber_loss, WhiteningNormalizer, or a safe installed-environment compatibility check.

Route elsewhere

• For DQN, Double DQN, Dueling DQN, SARSA, CEM, replay memory, and discrete policy construction details, use the sibling discrete-control sub-skill.
• For DDPG, NAF, actor/critic/value/mu/L model construction, continuous-action random processes, and MuJoCo/Pendulum patterns, use the sibling continuous-control sub-skill.

Use this sub-skill

1. Check the exact shared APIs and callback/log schema in references/api-reference.md.
2. Follow the applicable lifecycle, processor, logging, visualization, and environment-check workflows in references/workflows.md.
3. If imports, plotting, callbacks, Gym wrappers, or processor shapes fail, use references/troubleshooting.md.
4. To plot a FileLogger JSON without relying on the original repository, use scripts/visualize_keras_rl_log.py.
5. To inspect installed keras-rl/Keras/backend compatibility before agent work, use scripts/check_keras_rl_env.py.

Compatibility note

keras-rl is legacy Keras 2.x code. Prefer a legacy Keras backend and check compatibility before investing in training. A Theano CPU backend is often the safest compile-only route for this package family; TensorFlow-era imports may work while agent construction can still hit legacy symbolic-tensor behavior. Modern tf.keras/Keras 3 stacks should be treated as incompatible until proven by a local smoke check.