---

name: core-workflows description: "Build and debug Acme core dm_env loops, specs, wrappers, logging, counting, observers, and simple custom Actor/Learner components." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Acme Core Workflows

Use this sub-skill when an agent needs to assemble or debug the framework-level Acme loop before choosing a JAX, TensorFlow/Sonnet, or distributed backend.

Read First

• Start with Core API contracts for exact signatures, method ordering, and import surfaces.
• Use Core workflows for adapting Gym-like environments, building a minimal Actor, wiring EnvironmentLoop, wrappers, loggers, counters, and observers.
• Use Core troubleshooting when loop bounds, specs, dtypes, wrapper behavior, logger timing, or optional backend imports fail.
• Run check_core_imports.py to inspect the local Acme installation without requiring JAX/TF/Reverb/Launchpad execution.

Scope

This sub-skill covers:

• acme.core.Actor, Learner, VariableSource, Saveable, and Worker component contracts.
• acme.specs aliases and make_environment_spec(environment) for dm_env environments.
• acme.EnvironmentLoop and acme.environment_loop.EnvironmentLoop single-environment interaction.
• Core wrappers from acme.wrappers, especially GymWrapper, SinglePrecisionWrapper, CanonicalSpecWrapper, EnvironmentWrapper, and wrap_all.
• acme.utils.counting.Counter, acme.utils.loggers, acme.utils.observers, and signal-safe loop termination basics.

Route elsewhere:

• Replay adders, Reverb tables, datasets, and iterator pipelines belong in replay-and-data.
• JAX networks, builders, experiments, and agent-specific JAX behavior belong in jax-agents.
• TensorFlow/Sonnet agents and Launchpad-backed TF examples belong in tf-agents.

Quick Routing

• If the user has a custom environment, first make it implement dm_env.Environment or wrap it with acme.wrappers.GymWrapper, then verify observation_spec(), action_spec(), reward_spec(), and discount_spec().
• If the user has a loop failure, check EnvironmentLoop.run() bounds, action/spec dtype compatibility, and whether optional imports are failing before the loop starts.
• If the user needs metrics, prefer Counter, a Logger implementation, and EnvLoopObserver instances instead of modifying EnvironmentLoop internals.
• If the user asks for a complete learning agent, keep this sub-skill focused on the Actor/Learner interface and route backend-specific network or replay decisions to the backend/data sub-skills.