The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the yq binary from the official GitHub repository (github.com/mikefarah/yq) during workflow execution to facilitate YAML configuration patching. GitHub is a well-known and trusted service.
[COMMAND_EXECUTION]: Shell scripts are used to orchestrate complex tasks, including IsaacSim rendering, Cosmos3 training, and inference. These operations include standard environment manipulations such as updating the PATH variable to include locally downloaded tools.
[DATA_EXFILTRATION]: Credential preflight scripts verify that the Hugging Face token (HF_TOKEN) provided in the environment has the necessary permissions to access gated model repositories (nvidia/Cosmos3-Nano, nvidia/Cosmos3-Edge) by performing authenticated API probes to Hugging Face, a well-known service.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided configuration files (defect_spec.jsonl) that may contain natural language prompts for region-of-interest (ROI) detection.
Ingestion points: The defect_spec.jsonl file is read in scripts/anomalygen_train.sh and scripts/anomalygen_generate.sh.
Boundary markers: No explicit boundary markers are present in the prompt interpolation.
Capability inventory: The workflow environment allows shell script execution and file writing to S3 buckets.
Sanitization: No sanitization is performed on the input prompt strings, as they are intended for model guidance.
[DYNAMIC_EXECUTION]: Several orchestration scripts use embedded Python heredocs to perform runtime patching of configuration files based on environment variables and user-supplied parameters.

physical-ai-defect-image-generation-v1-1