The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references container images hosted on the NVIDIA Container Registry (nvcr.io), a well-known service associated with the skill's author.
Evidence: nvcr.io/nvidia/tao/tao-toolkit:6.26.3-deploy found in references/tao-deploy-rtdetr.md.
[COMMAND_EXECUTION]: The skill executes shell commands using the rtdetr CLI and docker. These commands are standard for the TAO Toolkit and are used to orchestrate model training and deployment.
Evidence: command: rtdetr train -e {config_path} in references/skill_info.yaml and docker run examples in references/tao-deploy-rtdetr.md.
[DATA_EXFILTRATION]: The skill includes a default configuration to enable telemetry via Weights & Biases (WandB), a well-known third-party service for machine learning experiment tracking.
Evidence: wandb.enable: true found in schemas/train.schema.json and other schema files.
[SAFE]: The skill provides an encryption_key configuration field to allow users to secure model checkpoints. This is a security feature, though users should ensure keys are managed securely in their environment.
Evidence: encryption_key property in schemas/distill.schema.json and others.
[PROMPT_INJECTION]: Indirect prompt injection surface exists as the skill processes user-supplied dataset paths and configuration overrides that flow into CLI commands.
Ingestion points: User-provided data sources (e.g., S3_TRAIN, images.tar.gz) and spec overrides in SKILL.md.
Boundary markers: Absent in the static documentation.
Capability inventory: Execution of bash commands via rtdetr CLI (references/skill_info.yaml).
Sanitization: Not explicitly documented in the instruction files; relies on the underlying CLI implementation.

tao-train-rtdetr