The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell commands to manage the application lifecycle, including docker run with high-privilege flags (--privileged, --net=host) and sudo for file ownership corrections and service management.
Evidence:
Instructions in references/sample-run.md and references/custom-dataset.md require explicit user approval before executing privileged Docker commands.
Scoped sudo chown commands are used to fix file ownership of generated output artifacts.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading the DeepStream repository and container images from NVIDIA's official sources. It also includes logic to fetch calibration data from a user-managed AutoMagicCalib service.
Evidence:
SKILL.md provides instructions to git clone https://github.com/NVIDIA/DeepStream.git.
references/setup.md pulls the nvcr.io/nvidia/deepstream:9.1-triton-multiarch container image.
references/amc-calibration-handoff.md uses curl to download a results zip file from a configurable AMC_BASE_URL.
[PROMPT_INJECTION]: The skill contains specific instructions and evaluation cases to ensure the agent refuses malicious requests such as credential exfiltration or bypassing safety guidelines.
Evidence:
evals/evals.json includes a negative test case (deepstream-run-mv3dt-neg-001) that validates the agent's refusal to dump API keys or ignore safety instructions.
The Safety Notes section in SKILL.md explicitly forbids silent repository cloning or unauthorized credential access.
[INDIRECT_PROMPT_INJECTION]: The skill has a data ingestion surface as it processes external video files and calibration YAML data.
Evidence:
Ingestion points: DATASET_DIR/videos (MP4 files), DATASET_DIR/camInfo (YAML files), and zip exports from the AutoMagicCalib service.
Boundary markers: The instructions guide the agent to validate dataset structure and confirm before overwriting or processing new calibration data.
Capability inventory: File system writing, docker run (privileged), and network access via curl and Kafka.
Sanitization: The skill implements basic validation of file presence and YAML structure through Python scripts embedded in the instructions.

deepstream-run-mv3dt