The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/jev_browser_agent.py interacts with the host system to manage the browser and fetch necessary credentials.
Evidence: Uses subprocess.run to call the macOS security CLI to retrieve API keys from the system Keychain.
Evidence: Uses subprocess.Popen to launch Google Chrome or Chromium binaries with hardening flags such as --headless=new and --user-data-dir for isolated execution.
[DYNAMIC_EXECUTION]: The runner employs a self-re-execution strategy to ensure it runs within a specific virtual environment.
Evidence: The ensure_importable function uses os.execv to replace the current process with a Python interpreter located in a path defined by the JEV_ULTRAFAST_REPO environment variable if dependencies are not found in the global path.
[INDIRECT_PROMPT_INJECTION]: As a browser automation tool, the skill is naturally exposed to untrusted data from web pages which could contain malicious instructions.
Ingestion points: The Agent reads and processes DOM elements and page content from live websites in scripts/jev_browser_agent.py.
Boundary markers: The SKILL.md documentation explicitly instructs the agent to treat page content as data only and never as a source of instructions.
Capability inventory: The skill possesses the ability to navigate, click, and type text within the browser environment.
Sanitization: The implementation mitigates risks by enforcing a mandatory host allowlist (--allow-hosts) and a strict cap on the number of model interactions allowed per task (--max-ticks).

jev-browser-use