The Agent Skills Directory

[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it ingests untrusted data in the form of YAML configuration files and JSONL benchmark results which directly influence the shell commands the agent is instructed to execute.
Ingestion points: scripts/compare_benchmark_results.py (reads JSONL results) and scripts/validate_cookbook_configs.py (reads YAML configurations).
Boundary markers: Absent. There are no specific delimiters or instructions to ignore embedded commands within the processed data fields.
Capability inventory: While the included Python scripts do not invoke dangerous primitives themselves, the overall skill workflow requires the agent to execute complex shell commands (Docker, Python modules) constructed from the data.
Sanitization: scripts/compare_benchmark_results.py performs basic escaping of pipes and newlines for Markdown table integrity. scripts/validate_cookbook_configs.py provides a validation layer that checks rendered command flags against known framework surfaces.
[SAFE]: The skill demonstrates safe handling of sensitive information, providing explicit instructions to avoid printing secrets like HF_TOKEN in logs or artifacts and recommending the use of environment variable inheritance for containers.
[SAFE]: External resources such as Docker images and documentation links target well-known and reputable organizations in the machine learning ecosystem (e.g., NVIDIA, Cloudflare, vLLM, SGLang projects).

llm-serving-auto-benchmark