The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to install the framework and specialized dependencies from established project repositories and specialized package indices. Examples include the official SGLang GitHub repository and the FlashInfer package index for high-performance CUDA kernels.\n
Evidence: Installation commands like pip install sglang[all] flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/ and git clone https://github.com/sgl-project/sglang.git are found in SKILL.md.\n- [INDIRECT_PROMPT_INJECTION]: As a tool designed to process and serve LLM requests, the workflows naturally involve ingesting external data (e.g., user queries, articles, image paths), which creates an attack surface for indirect prompt injection.\n
Ingestion points: External data enters the context through variables such as user_query, text, article, and new_message as seen in SKILL.md and references/structured-generation.md.\n
Boundary markers: The skill prominently features 'structured generation' as a core defense, using json_schema, regex, and grammar to strictly control and validate the model's output format.\n
Capability inventory: Capabilities include inference serving, multi-GPU processing, and reading multi-modal data (e.g., via sgl.image).\n
Sanitization: The framework focuses on output validation; standard developer practices for input sanitization are assumed for production deployments.

inference-serving-sglang