The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands via scripts like scripts/check_artifacts.py (using bash -n for syntax validation) and scripts/pull_push_image.sh (using docker for image management). These are well-contained and necessary for the deployment workflow.
[EXTERNAL_DOWNLOADS]: The skill fetches model metadata and documentation from docs.vllm.ai, container images from quay.io, and model weights from huggingface.co or modelscope.cn. These are well-known, trusted services for AI development, and the skill provides explicit instructions to verify these sources using -k only after hostname validation.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external model documentation and user-provided configuration paths, creating a potential attack surface for indirect instructions.
Ingestion points: Documentation pages from docs.vllm.ai, model repository IDs, and user-supplied OBS/SWR paths.
Boundary markers: The skill includes mandatory security guidelines in SKILL.md requiring URL whitelisting and input validation before any data is passed to tools.
Capability inventory: API interactions in create_infer_service.py, file operations in upload_code_to_obs.py, and container operations in pull_push_image.sh.
Sanitization: The skill instructions mandate the use of regex and whitelist validation for all user-supplied and externally-fetched inputs before command interpolation.
[SAFE]: The skill follows security best practices by recommending environment variables or protected files for IAM token handling, explicitly checking for 600 file permissions, and providing clear troubleshooting for resource cleanup.

huawei-cloud-modelarts-vllm-ascend-model-deploy