The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/generate_od_defects.py uses subprocess.run() to execute torchrun and a label processing utility. The implementation is secure as it passes arguments as an array and refrains from using shell=True, preventing command injection. It also rigorously validates the existence of all target scripts and input files before execution.\n- [EXTERNAL_DOWNLOADS]: The skill requires specific model checkpoints from Hugging Face (Qwen, NVIDIA Cosmos) and DINOv2. Security is maintained by requiring these assets to be pre-staged in a mounted directory and enforcing an offline runtime environment through environment variables (HF_HUB_OFFLINE=1, TRANSFORMERS_OFFLINE=1) to prevent runtime network access.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes structured JSON generation plans and COCO annotation files. Ingestion points include plan files and tool-generated labels. Sanitization includes strict regex-based ID validation and spatial bounding box verification. The skill has capabilities for file system access and subprocess management for local tasks.\n- [DATA_EXFILTRATION]: Access is restricted to designated input/output directories, and the script actively disables network synchronization for model libraries, ensuring that sensitive model data or generated images are not transmitted to external servers.

tao-generate-od-defects