The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages a browser lifecycle and interacts with the operating system through multiple shell scripts.
Executes scripts/check-deps.sh, scripts/ensure-browser.sh, and scripts/close-browser.sh to manage environment and process states.
Launches a Google Chrome or Chromium process with the --remote-debugging-port=9222 flag enabled.
Relies on the agent-browser CLI tool to send commands to the browser instance.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic script execution for specific automation tasks.
scripts/close-browser.sh executes a Python script using a heredoc (PYEOF) to interact with the browser's WebSocket debugging interface without requiring external libraries.
The instructions direct the agent to use agent-browser --cdp 9222 eval "<js>" to execute custom JavaScript within the page context for extracting specific DOM elements or controlling media playback.
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to ingest and process data from external websites, it possesses an inherent attack surface for indirect prompt injection.
Ingestion points: Content enters the agent's context through snapshot (accessibility tree), get text, screenshot images, and the results of eval scripts in SKILL.md.
Boundary markers: The instructions do not define specific delimiters or warning headers to isolate untrusted web content from the agent's core instructions.
Capability inventory: The skill has the capability to write files (saving screenshots to /tmp/), navigate to arbitrary URLs, and execute JavaScript in a browser.
Sanitization: There is no explicit sanitization or filtering of the fetched web data within the provided scripts before it is processed by the AI.

web-access