The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and analyze untrusted data from external sources, including live web pages, screenshots, and terminal output. This is a known attack surface where malicious content on a target website could attempt to influence the agent's logic.
Ingestion points: Live web content (via browser_vision), browser console logs, and user-provided images/files.
Boundary markers: The skill enforces the use of structured JSON schemas (e.g., visual_qa_verdict/v1, render_capture_manifest/v1) for all outputs, which helps maintain a boundary between untrusted data and agent instructions.
Capability inventory: The skill has the ability to interact with browsers via browser_cdp and browser_console and can perform file-based operations for logging evidence.
Sanitization: The instructions explicitly warn against performing destructive actions (like checkout or deletion) and mandate the redaction of credentials and PII from all captured evidence.
[COMMAND_EXECUTION]: The skill makes use of platform-provided browser automation tools, specifically browser_cdp (Chrome DevTools Protocol) and browser_console. These are used appropriately within the scope of the skill's primary purpose (web auditing and viewport emulation) and are gated by platform-level tool permissions.

omh-visual-qa