The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the 'okou computer-use' command-line interface to interact with the host operating system. This includes listing applications, opening bundles, retrieving application state, and simulating user inputs like clicks, typing, and key presses.
[DYNAMIC_EXECUTION]: Node.js is used to execute inline script snippets via the '-e' flag. These scripts are used to process and filter JSON data generated by the computer-use tools, such as searching for specific applications by name or extracting accessibility tree paths from temporary files.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection because it ingests and acts upon data from external applications that may contain attacker-controlled content.
Ingestion points: The agent reads 'appState' accessibility trees and screenshots from the local filesystem (e.g., /tmp/vm0/computer-use/...) which represent the current state of third-party desktop applications.
Boundary markers: There are no explicit delimiters or warnings instructed to the agent to treat content found within the accessibility tree or screenshots as untrusted data.
Capability inventory: The skill provides a high degree of control over the desktop environment, including the ability to type text, press keys, and navigate browsers.
Sanitization: No evidence of sanitization or filtering of external application content (like window titles, web page text, or element labels) is present before the agent processes the data to decide its next action.

computer-use