The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run and subprocess.Popen within runtime/cli/executor.py and runtime/cli/async_jobs.py to execute the YOLO command-line interface and manage background training tasks. It also uses subprocesses to launch Gradio or Streamlit web interfaces in runtime/cli/launcher_handlers.py. These operations are fundamental to the skill's purpose of orchestrating machine learning workflows.
[EXTERNAL_DOWNLOADS]: Communicates with OpenAI-compatible APIs (including OpenAI and Alibaba DashScope) for multimodal visual reasoning as described in runtime/multimodal/runtime.py. It also references datasets and taxonomies from well-known sources like HuggingFace and OpenDataLab for open-world object detection.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted image data and user prompts.
Ingestion points: Untrusted data enters the agent context via inputs.source (image paths/URLs) and inputs.prompt which are processed in the multimodal reasoning pipeline.
Boundary markers: Prompt templates in the assets/prompts/ directory use clear delimiters (e.g., User task: {{USER_PROMPT}}) to separate user input from system instructions.
Capability inventory: The skill can execute shell commands via subprocess and write artifacts/logs to the local filesystem.
Sanitization: The skill employs json_safe serialization and strict path resolution logic in runtime/cli/normalize.py to mitigate common injection risks.
[DYNAMIC_EXECUTION]: Uses importlib.import_module in several modules (e.g., runtime/cli/device.py, runtime/cli/moe_tools.py) to load the ultralytics framework and repository-internal helpers. This is used for legitimate code organization and feature detection.

yolo-master-agent