The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for executing various shell commands to set up the environment, install dependencies, and manage services. Examples include pip install vllm, docker run, and vllm serve.
[PRIVILEGE_ESCALATION]: Troubleshooting documentation suggests using sudo ufw allow 8000 to modify system firewall rules to allow network traffic on specific ports.
[REMOTE_CODE_EXECUTION]: The documentation recommends the use of the --trust-remote-code flag when serving models. This configuration allows the library to execute arbitrary Python code contained within model configuration files downloaded from remote repositories like Hugging Face.
[INDIRECT_PROMPT_INJECTION]: The skill describes a batch processing workflow in SKILL.md that ingests input from an external file (prompts.txt) and processes it using the llm.generate method. The instructions do not implement input sanitization or boundary markers, which presents a surface for indirect prompt injection if the input file contains malicious instructions.

serving-llms-vllm