The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches software packages and version metadata from official PyTorch wheel indexes and the standard PyPI registry.
[COMMAND_EXECUTION]: The skill invokes various shell commands using the Python interpreter to inspect the host environment, verify hardware (via nvidia-smi, rocm-smi, and hipcc), and manage packages through pip.
[DYNAMIC_EXECUTION]: The skill performs runtime compilation of CUDA or HIP kernels when the universal wheel distribution is used, which is necessary for the tool's core functionality of model acceleration.
[INDIRECT_PROMPT_INJECTION]: The skill allows users to provide local paths for wheels or source code directories as installation targets, which represents a potential ingestion point for untrusted content if the user provides a malicious directory.
Ingestion points: User-supplied paths provided via <ABSOLUTE_WHEEL_PATH> and <ABSOLUTE_SOURCE_DIR> in SKILL.md.
Boundary markers: None implemented for the content of the provided directories.
Capability inventory: The skill uses subprocess calls to execute pip install and python -c verification scripts.
Sanitization: The skill mitigates this risk by explicitly requiring the user to trust the source and provide approval before execution.

quark-install