The Agent Skills Directory

[SAFE]: The skill documents standard practices for multi-GPU training using torchrun and uv. All instructions focus on legitimate training workflows and troubleshooting of NCCL (NVIDIA Collective Communications Library) issues.
[COMMAND_EXECUTION]: The skill suggests running training commands like uv run isaaclab train_multigpu. These are standard entry points for the Isaac Lab framework and are consistent with the skill's stated purpose of facilitating reinforcement learning training.
[REMOTE_CODE_EXECUTION]: While the skill provides a Python script (nccl_probe.py) for debugging, this is a local diagnostic tool used to test NCCL synchronization. The code is transparently provided in the documentation and requires the user to manually create the file, posing no hidden execution risk.
[PROMPT_INJECTION]: No evidence of instructions designed to bypass agent safety filters or override system constraints was found. The instructions are technical and educational in nature.
[DATA_EXFILTRATION]: No network operations to unknown or suspicious domains were detected. The mentioned environment variables (e.g., NCCL_P2P_DISABLE) are standard configuration flags for the NVIDIA NCCL library to handle hardware-specific transport issues.

isaaclab-training-multi-gpu