The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches container images from NVIDIA's NGC registry (nvcr.io), clones source code from official NVIDIA GitHub repositories, and downloads models and configurations from HuggingFace's public hub. These operations target trusted organizations and well-known services.
[COMMAND_EXECUTION]: The skill executes various shell commands to manage the integration lifecycle, including docker run, docker exec, and git clone. It also runs Python-based tools like pytest, pylint, and flake8 for codebase validation within isolated containers.
[CREDENTIALS_UNSAFE]: The workflow requires authentication for private models and registries. It correctly instructs the agent to prompt the user for an HF_TOKEN and an NGC API Key rather than storing them in the skill itself. It uses standard authentication mechanisms such as huggingface_hub.login and docker login.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) as it processes untrusted data in the form of user-provided HuggingFace model IDs or URLs. It includes a manual validation step where the agent must verify the model's pipeline_tag and architecture before proceeding with the implementation.
Ingestion points: $ARGUMENTS, HuggingFace model ID, and model URL.
Boundary markers: None explicitly defined for user-supplied model identifiers.
Capability inventory: The skill has extensive capabilities including file system writes, network access (cloning/pulling), and command execution (docker, pip).
Sanitization: Includes checks to ensure the model is a Computer Vision model and provides an ONNX export sanity check gate.
[COMMAND_EXECUTION]: Provides instructions to use sudo rm -rf to clean up root-owned build artifacts (e.g., *.egg-info, build/) generated during containerized development. This is a standard procedure for managing host-mounted volumes in Docker-based development workflows.

tao-port-huggingface-model