The Agent Skills Directory

[EXTERNAL_DOWNLOADS]:\n
Source: https://github.com/SwinTransformer/storage/releases/download/v1.0.8/swin_tiny_patch4_window7_224_22k.pth\n
Context: Referenced in SKILL.md as the pretrained_backbone.\n
Behavior: Fetches model weights from a well-known public repository on GitHub.\n- [COMMAND_EXECUTION]:\n
Source: references/tao-deploy-oneformer.md\n
Behavior: Provides documentation for docker run commands to execute the TAO Toolkit.\n
Context: Utilizes official container images from the vendor's registry (nvcr.io) to perform machine learning tasks.\n- [SAFE]:\n
Category 8 (Indirect Prompt Injection) Surface Analysis:\n
Ingestion points: Dataset paths and S3 bucket identifiers in SKILL.md.\n
Boundary markers: Present; the skill employs strict JSON schemas (schemas/) and spec templates to structure all agent-provided data.\n
Capability inventory: Bash tool usage is documented for running official TAO subcommands within isolated containers in references/skill_info.yaml.\n
Sanitization: External inputs are validated against the schema properties before being passed to the training engine.\n
Secret Management: The skill correctly uses empty string placeholders for sensitive parameters like encryption_key, ensuring no credentials are hardcoded.

tao-train-oneformer