The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/jev_gui_agent.py utilizes subprocess.Popen to launch the cua-driver binary and subprocess.run to call /usr/bin/open on macOS. These operations are required for the skill to observe the desktop and interact with applications as intended.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it ingests untrusted data from the GUI environment, such as window titles and element labels. This data is used as context for model-driven decisions.
Ingestion points: The element_rows function in scripts/jev_gui_agent.py and observe in scripts/coagent_cu.py collect labels and roles from active desktop windows.
Boundary markers: The skill relies on external model processing (jevkit), and explicit prompt boundary markers are not present in the local script logic.
Capability inventory: The agent has the authority to click elements, type text, scroll, and open specific apps or URLs.
Sanitization: The skill implements an is_sensitive filter for UI labels and uses validation functions like safe_url and safe_app_name to vet plan steps before execution.
[DATA_EXFILTRATION]: To facilitate GUI navigation, the skill transmits UI metadata, including element labels and current goals, to external services such as the Jev model and OpenRouter. While the skill attempts to filter sensitive labels locally, this represents a flow of desktop information to remote endpoints.

jev-computer-use