The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to execute docker run commands. These commands invoke the NVIDIA TAO data-services container to perform signed distance calculations and threshold sweeps. The execution includes necessary parameters like --gpus all and host volume mounting (-v $WORKSPACE:$WORKSPACE) which are required for the skill's intended computational purpose.
[EXTERNAL_DOWNLOADS]: The skill setup logic resolves a specific Docker image tag using a local versioning script and performs a docker pull. This targets official NVIDIA TAO Toolkit images, which are trusted vendor resources aligned with the skill's functionality.
[DATA_EXPOSURE]: The skill includes a packaging hook (hooks/rca-package.sh) that automatically copies session transcript logs and project settings into the output directory when an analysis report is generated. While this involves accessing the local .claude directory, it is a functional requirement for Root Cause Analysis (RCA) workflows to preserve the context of the model evaluation.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from inference results (inference.csv) and training configurations (train.yaml).
Ingestion points: Data is read from user-provided experiment directories into the data-services container and local Python analysis blocks.
Boundary markers: The skill does not use explicit boundary markers or "ignore instructions" warnings when interpolating data into the final RCA_Report.md.
Capability inventory: The skill has capabilities for Docker execution, file system writes (via cp and Python), and network access (via docker pull).
Sanitization: External content is parsed using structured data libraries like pandas and pyarrow, which mitigates common injection risks during processing.

tao-analyze-gaps-visual-changenet