The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_anomalygennext_amp.py uses subprocess.run to launch the anomalygen.scripts.auto_mask_placement.roi_place module. The execution uses a list format (preventing shell injection) and takes parameters like model_id from the подготовке configuration file. This is a standard pattern for multi-step AI processing pipelines.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from Parquet files and YAML configurations, which define file paths, anomaly types, and model parameters. While this introduces an indirect injection surface, the scripts include extensive validation logic for mask contents, image dimensions, and bounding box integrity to mitigate malformed data risks.
[DYNAMIC_EXECUTION]: The test suite (scripts/tests/) uses importlib.util to dynamically load and execute the skill's scripts for verification purposes. Additionally, the run_amp action dynamically constructs command-line arguments to invoke internal Python modules.
[EXTERNAL_DOWNLOADS]: The skill references models from Hugging Face (nvidia/Cosmos3-Nano, Qwen/Qwen3-VL-8B-Instruct). However, the run_amp script explicitly sets HF_HUB_OFFLINE="1" and TRANSFORMERS_OFFLINE="1", ensuring the skill operates on local assets rather than performing runtime downloads from the public internet.

tao-prepare-anomalygennext-inputs