The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is configured to download pretrained model weights from Hugging Face's public repository as a standard part of the fine-tuning workflow.
Evidence: References to huggingface.co in references/checkpoint-inference-mappings-foundation-stereo.md and references/skill_info.yaml for backbone weight initialization.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied data in the form of annotation files and image datasets, which are used to configure model training and inference actions. This represents a standard ingestion surface for machine learning workflows.
Ingestion points: Path definitions for annotations.txt and dataset folders in SKILL.md and spec templates.
Boundary markers: Absent; the skill relies on the TAO Toolkit's internal structured parsing (YAML/JSON) of these inputs.
Capability inventory: The skill uses the Bash tool to execute docker run commands and the depth_net command-line interface.
Sanitization: Sanitization and validation are managed by the underlying NVIDIA TAO Toolkit binaries and the Docker container environment.

tao-train-foundation-stereo