The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches model weights and configuration from Hugging Face and installs the vLLM serving engine. These downloads target well-known, official repositories (huggingface.co and the vLLM project GitHub) and are required for the skill's primary function.
[COMMAND_EXECUTION]: Instructions involve executing shell commands on Runpod Pods to create virtual environments, install dependencies, and launch model servers. These commands are localized to the user's provisioned infrastructure and include safety checks for CUDA compatibility and system resources.
[CREDENTIALS_UNSAFE]: The skill handles sensitive tokens including HF_TOKEN and RUNPOD_API_KEY. It incorporates robust protective instructions: tokens must be read from environment variables, never passed via command-line arguments (to prevent exposure in process lists), and are explicitly forbidden from being printed in the agent's output.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of Hugging Face model cards and model responses during the verification phase. This risk is mitigated by the requirement to verify file integrity using SHA-256 hashes and the use of dedicated, non-evaluative testing scripts to validate model readiness.

runpod-vllm