The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides commands to download model weights and configuration files from Hugging Face using the huggingface_hub CLI. These downloads target the official nvidia/ organization on Hugging Face (e.g., nvidia/NV-CodonFM-Encodon-1B-v1). As Hugging Face is a well-known service and NVIDIA is the verified vendor for this skill, these operations are considered safe and represent the primary function of the skill.
[COMMAND_EXECUTION]: The instructions include standard development commands for environment setup, including Docker container management (docker run, docker exec, docker ps), Python virtual environment creation (python3.11 -m venv), and package installation via pip. These commands are localized to the user-provided repository directory and are standard practices for ML model deployment.
[INDIRECT_PROMPT_INJECTION]: The skill contains instructions for processing external data (supplied source archives and checkpoint metadata). While this represents a surface for indirect prompt injection, the skill explicitly warns the agent to use standard library tools (like zipfile) without full extraction for inspection and provides clear boundaries for supported model architectures, reducing the risk of accidental obedience to embedded instructions.
[DYNAMIC_EXECUTION]: The skill uses python -c to verify CUDA availability and GPU access. This is a common and safe pattern for verifying hardware prerequisites in machine learning environments and does not involve executing untrusted strings or external code dynamically.

codonfm-setup