The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple Python scripts (detect.py, cpu_tune.py, validate.py) utilize subprocess.run(shell=True) to execute system discovery commands such as lscpu, nproc, and grep. Additionally, the skill instructions in SKILL.md use eval to ingest environment variables produced by the cpu_tune.py script.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch metadata and configuration from established external sources. It downloads model registry information from the vLLM project's official GitHub repository to verify architecture support and fetches model metadata from Hugging Face APIs using urllib.request in scripts/check_model.py and scripts/estimate_memory.py.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the processing of untrusted model identifiers.
Ingestion points: User-provided <model> and <port> parameters passed to the serve scripts.
Boundary markers: Absent; inputs are interpolated directly into shell command strings without delimiters.
Capability inventory: The skill possesses capabilities for shell execution, container management (docker/podman), and hardware configuration.
Sanitization: No validation or escaping of user-supplied strings was observed before interpolation into shell commands.
[REMOTE_CODE_EXECUTION]: scripts/detect.py contains logic to execute discovery commands on remote systems via ssh if a host is specified. While intended for remote hardware detection, this capability involves executing shell commands on external targets.

serving-llms-on-epyc