What problem does it solve? Turning a clarified AutoRL task into runnable code is risky: choosing the wrong RL framework, installing dependencies without approval, or launching full training prematurely can waste compute and break reproducibility. This Skill converts evidence-backed task artifacts into a concrete, approval-gated implementation path. ## Core Features & Use Cases - Framework Selection: Chooses between reusing an existing local framework, adding a maintained RL library (Stable-Baselines3, CleanRL, Tianshou, RLlib), cloning a reference repo into third_party/, or writing a minimal native implementation, with rationale recorded in implementation_plan.md. - Dependency & Approval Gating: Writes dependency_plan.md with exact commands, version/commit pins, risks, and rollback notes before any pip install, git clone, or large download. - Bounded Verification: Runs only the smoke tests allowed by the execution boundary (generate_only, dry_run, or runtime_allowed) and reports results in smoke_test_report.md. - Use Case: Given a task card, evidence report, and decision packet for a Gymnasium-based PPO task, the Skill inspects the workspace, selects Stable-Baselines3, writes env adapter, training and evaluation entrypoints, and runs a one-step dry-run smoke test without starting full training. ## Quick Start Use the rl-framework-implementer skill to set up the RL framework and implement the training path for the task defined in runs/<task-id>/task_card.md.
