Terminal-Bench-RL

Open-source reinforcement learning code for training long-horizon terminal coding agents, scaling from two A100 GPUs to 32 H100s with GRPO and Docker-isolated rollouts.