The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation provides commands to download and execute installation scripts for the qwen-cua-driver tool from the author's official GitHub repository (QwenLM/qwen-code). These patterns include /bin/bash -c "$(curl -fsSL ...)" and irm ... | iex for Windows.
[PERSISTENCE]: The skill documents management commands for the qwen-cua-driver CLI (qwen-cua-driver autostart enable) that allow the user to register a Windows Scheduled Task to automatically start the driver daemon upon interactive logon. Similar capabilities for systemd units are mentioned for Linux.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it ingests untrusted data from GUI accessibility trees, screen contents, and web pages.
Ingestion points: Untrusted data enters the agent context through GUI metadata (accessibility trees) and screenshots processed in SKILL.md and BROWSER.md.
Boundary markers: The instructions do not define strict delimiters for this data, though SKILL.md explicitly warns the agent to treat all application-provided text and labels as untrusted.
Capability inventory: The skill has high-privilege capabilities including simulated user input (keyboard/mouse) and filesystem mutation, implemented via subprocess calls to the qwen-cua-driver binary.
Sanitization: There is no automated sanitization of external content, placing the responsibility for safe interpretation on the agent's internal logic.
[COMMAND_EXECUTION]: The skill operates by executing the qwen-cua-driver CLI tool through shell commands to perform GUI automation tasks, including application launching, window state inspection, and input delivery.

cua-driver