The Agent Skills Directory

EXTERNAL_DOWNLOADS (LOW): The skill instructs the user to install mamba-ssm and causal-conv1d via pip. While these are widely recognized research packages, their source (GitHub state-spaces) is not within the pre-defined 'Trusted Organizations' list. Standard supply chain risks apply.
REMOTE_CODE_EXECUTION (LOW): The skill utilizes MambaLMHeadModel.from_pretrained and AutoTokenizer.from_pretrained to download model weights and configurations from HuggingFace. This pattern involves executing code to load remote assets, which is standard for the transformers ecosystem but technically involves remote data processing.
DYNAMIC_EXECUTION (MEDIUM): In training-guide.md, the skill demonstrates the use of torch.load() for model checkpoints. This function uses Python's pickle module, which is susceptible to arbitrary code execution if a user is tricked into loading a malicious checkpoint file from an untrusted source.

mamba-architecture