The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses standard command-line tools for benchmarking and process management.
Executes vllm serve to start the inference engine.
Executes vllm bench serve to initiate the performance measurement.
Uses curl for health checks and kill for process cleanup, which are standard administrative actions for this use case.
[EXTERNAL_DOWNLOADS]: The skill references the installation of the vllm package via pip.
vLLM is a widely recognized, open-source library for high-throughput LLM serving.
The models referenced (e.g., Qwen, Llama, OPT) are standard models hosted on HuggingFace, typical for benchmarking tasks.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with LLM endpoints by sending synthetic prompts.
Ingestion points: The --dataset-name random parameter in vllm bench serve generates synthetic data internally (SKILL.md).
Boundary markers: Standard CLI argument structure.
Capability inventory: Includes starting/stopping server processes and making local network requests (curl).
Sanitization: The skill relies on the vllm tool's internal handling of synthetic data generation.

vllm-bench-random-synthetic