The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands via Python's subprocess module to perform hardware discovery (lscpu, nproc), check system resources (/proc/meminfo), and manage container runtimes (docker, podman). These commands are essential for configuring hardware-specific optimizations like NUMA binding and thread affinity.
[EXTERNAL_DOWNLOADS]: The skill retrieves model metadata and serving registry information from well-known sources including HuggingFace (huggingface.co) and GitHub (github.com/vllm-project). It also pulls a validated serving image from a vendor-owned repository (amdih/zendnn_zentorch). These operations are consistent with standard AI development workflows.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied model identifiers and external configuration data from HuggingFace. The risk is mitigated through structured JSON parsing and a mandatory confirmation gate where the agent must present the full deployment plan for human review before execution.
Ingestion points: User-provided model IDs and HuggingFace API metadata.
Boundary markers: The skill explicitly requires the agent to wait for user confirmation (Step 6) after presenting the launch plan.
Capability inventory: Subprocess calls for system tuning, network access for metadata, and container lifecycle management.
Sanitization: The scripts employ shlex.quote for handling file paths and variables in shell contexts.
[SAFE]: No malicious patterns such as credential exfiltration, obfuscation, or persistence mechanisms were found. The skill specifically instructs the agent to report failures and stop instead of attempting to debug or escalate privileges, reducing the risk of unauthorized system changes.

serving-llms-on-epyc