The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates shell commands for launching containers and managing workloads using docker run, kubectl, and ssh. It employs shlex.quote() and json.dumps() in its Python templates to ensure that user-provided parameters and structured JSON payloads are correctly escaped before being executed in a shell environment.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of container images from the NVIDIA Container Registry (nvcr.io) and model weights from the HuggingFace Hub. These operations are essential to the skill's primary function and target well-known, trusted service providers.
[CREDENTIALS_SAFE]: The skill implements a robust 'Secrets rule' that prohibits prompting the user for sensitive credentials. Instead, it instructs the agent to read secrets such as HF_TOKEN, WANDB_API_KEY, and TAO_API_KEY from environment variables, which is a standard security practice for managing secrets in automated workflows.
[DATA_EXPOSURE]: The skill maintains a local service registry at /tmp/tao-inf-ms-state.json to track running services and their endpoints. Access is limited to the local system, and the skill includes instructions to route requests based on specific Job IDs, preventing accidental data exposure to the wrong service instance.

tao-run-inference-service