browse

Invoke the browse skill with a target URL and it opens the page in the Aside browser or a headless Chromium fallback, then saves a JPEG screenshot to a report directory. The agent reads the screenshot back so you see it inline.

The browse skill installs a console hook before navigation that captures console.error calls, uncaught exceptions, and unhandled promise rejections. After loading the page it prints a CONSOLE_ERRORS line with the collected array, verbatim.

Yes, when driving the Aside browser on macOS the skill works inside your real tabs with your existing cookies and sessions. It never reads cookies or types passwords; if a sign-in wall appears, you sign in yourself and the agent re-runs the step.

The skill falls back to gstack's own headless Chromium daemon, exposed through the $B command. The same cookbook steps translate one-to-one: goto, snapshot, click, fill, screenshot, and console error collection all work without Aside.

On local targets (localhost, .test) mutating actions may proceed directly. On any non-local origin the skill stops and asks for explicit consent once per run before any submit, create, delete, or settings change, and it never follows logout or delete links.

Dropdown items rendered in React portals or popovers without ARIA roles are absent from Playwright's accessibility tree. The skill's cursor-interactive scan detects clickable non-ARIA elements and floating containers to surface those items as @c refs.