What problem does it solve? Running reinforcement learning experiments with an AI agent often lacks structure: environments are unverified, training launches without approval, results are not reproducible, and state is lost on interruption. This Skill provides a deterministic, file-backed state machine that coordinates environment onboarding, evidence retrieval, training implementation, supervised execution, and result packaging with explicit approval gates. ## Core Features & Use Cases - Dual Workflow Profiles: Supports application_delivery (turn custom environment code into a verified, runnable training package) and research_iteration (reproduce and improve a baseline under a frozen comparison protocol). - Guarded Experiment Execution: Enforces an authority model separating build, experiment, control, and dependency permissions, so full training only launches with an exact run-plan approval digest. - Recoverable State Machine: Uses append-only hash-chained JSON events, compare-and-swap state transitions, and validation scripts so runs can resume safely after interruption. - Bounded Subagent Coordination: Spawns scoped child workers for environment onboarding, verification, evidence retrieval, training builds, and telemetry supervision, with the main orchestrator validating every artifact. - Use Case: A researcher provides a custom Gym environment and asks for a PPO training pipeline. The Skill verifies the environment contract, gathers evidence, builds checkpointable training code, runs a smoke test, and packages the verified deliverable without launching unauthorized full training. ## Quick Start Ask the agent to initialize an AutoRL run with the application_delivery profile for your custom environment and produce a verified training package with dry_run build authority.
