The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install standard machine learning libraries including mamba-ssm, causal-conv1d, torch, and transformers. It also references downloading pre-trained models from official HuggingFace repositories (e.g., state-spaces/mamba-2.8b). All referenced sources are well-known organizations in the AI research community.
[COMMAND_EXECUTION]: The documentation includes shell commands for package installation using pip and for running benchmark scripts. These commands are standard for setting up a development environment for the Mamba architecture and do not contain malicious payloads or unnecessary privilege escalations.
[INDIRECT_PROMPT_INJECTION]: The skill outlines workflows for loading pre-trained models and processing external input (prompts) for inference. While this creates a surface for indirect prompt injection, it follows standard large language model usage patterns and does not include unsafe interpolation or missing boundary markers in its implementation examples. As a documentation-focused skill, the risk is negligible.
Ingestion points: Pre-trained model weights from HuggingFace and user-provided text prompts in generation examples.
Boundary markers: Standard tokenizer-based processing is used.
Capability inventory: No file-write or network-send capabilities are present in the provided scripts; capabilities are limited to local model inference and training calculations.
Sanitization: Standard tokenization and generation constraints (max_length, temperature) are applied.

mamba-architecture