The Agent Skills Directory

[SAFE]: The skill enforces robust credential management by explicitly instructing the agent not to prompt the user for secret values like HF_TOKEN, WANDB_API_KEY, or TAO_API_KEY. It directs the agent to generate code that reads these secrets from environment variables, preventing sensitive data exposure in chat logs or shell history.
[COMMAND_EXECUTION]: All shell commands constructed for container management incorporate shlex.quote() for user-provided inputs such as model_path and job_id. This correctly mitigates the risk of command injection vulnerabilities.
[SAFE]: External resources, including container images, are sourced from the official and trusted NVIDIA registry (nvcr.io).
[SAFE]: Service state is managed locally in /tmp/tao-inf-ms-state.json, providing a structured discovery mechanism for microservice endpoints without external data exfiltration.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it processes untrusted output from the inference containers. 1. Ingestion points: references/request.yaml (processing choices[0].message.content). 2. Boundary markers: Absent. 3. Capability inventory: SKILL.md (Bash, Write, Read tools). 4. Sanitization: Absent. This is an inherent property of multimodal inference tasks.
[SAFE]: The workflow includes mandatory user checkpoints for sampling parameters, ensuring human-in-the-loop validation of the microservice's configuration.

tao-run-inference-service