The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download model weights from Hugging Face using the huggingface_hub library. The weights (nvidia/C-RADIOv2-B) are hosted by the official vendor organization.
[COMMAND_EXECUTION]: Provides templates and instructions for executing the visual_changenet command-line interface within NVIDIA TAO Docker containers. These commands involve standard volume mounts and GPU acceleration parameters.
[DYNAMIC_EXECUTION]: Documentation notes that the segmentation task requires compiling custom CUDA kernels (MultiScaleDeformableAttention) upon first execution. This is a standard procedure for this specific model architecture and is handled within the isolated container environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an ingestion surface via CSV dataset files (dataset.classify.train_dataset.csv_path).
Ingestion points: CSV files and image directories are processed during training and evaluation actions (referenced in SKILL.md and references/skill_info.yaml).
Boundary markers: None explicitly defined in the provided instruction templates.
Capability inventory: The skill has Read Bash tool access and can execute arbitrary Docker commands (defined in SKILL.md and references/local-docker.md).
Sanitization: Data parsing is performed by the visual_changenet binary inside the NVIDIA TAO container. While this represents a theoretical surface, it is inherent to the skill's primary function of model training.

tao-train-visual-changenet