The Agent Skills Directory

[SAFE]: The skill utilizes a container image from an official vendor registry (nvcr.io/nvidia/tao/tao-toolkit:7.3.0-data-services) for its core logic, ensuring a controlled execution environment.
[SAFE]: The provided Python script prepare_vlm_bcq_spec.py follows best practices, such as using yaml.safe_dump for writing configuration files and performing standard path validation without dangerous dynamic execution.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface as it processes external predictions_json files. This is a common and necessary pattern for data analysis tools and is considered safe in this context.
Ingestion points: The skill reads model responses and ground truth data from user-provided JSON files via the vlm_bcq action.
Boundary markers: Not explicitly defined in the helper script, though the core logic is isolated in a container.
Capability inventory: The skill uses Read and Bash tools to manage local configuration and initiate containerized analysis.
Sanitization: The helper script correctly uses yaml.safe_dump for spec generation, minimizing the risk of configuration injection.

tao-analyze-gaps-vlm-bcq