Website Visual Testing

Use this skill to run a website, open it in agent-browser, and verify the UI visually.

Start Here

• Use pty_spawn for any local app, preview server, Storybook, or long-running test target.
• Never start a server with shell backgrounding like &, nohup, or job control.
• Wait for an actual readiness signal with pty_read before opening the site in the browser.
• Use agent-browser snapshot -i before interactions and re-snapshot after navigation or major DOM changes.
• Prefer annotated screenshots when layout, unlabeled controls, or visual state matters.
• Treat visual testing as a UI check, not a replacement for functional or accessibility testing.

Read By Need

Workflow

1. Confirm the target

Determine:

1. whether the user wants a local site, a remote URL, or both
2. which command starts the site if it is local
3. which surfaces matter most: route, component, breakpoint, theme, or bug repro path

If the site is already hosted, skip the PTY startup flow and go straight to browser testing.

2. Start the site with opencode-pty

For local targets:

1. Inspect project scripts or docs to find the narrowest correct start command.
2. Start the server with pty_spawn using:
   • a clear title
   • the correct workdir
   • notifyOnExit: true when useful
3. Use pty_read to watch for:
   • localhost URLs
   • ready messages
   • compile errors
4. If the port is unclear, read the PTY output instead of guessing.

If startup fails, fix the blocker first. Do not continue to browser testing against a broken or half-started app.

3. Open the page in agent-browser

Use stable, explicit navigation:

agent-browser open http://127.0.0.1:3000
agent-browser wait --load networkidle
agent-browser snapshot -i

Prefer 127.0.0.1 or the exact host emitted by the dev server when available.

Before capturing screenshots on asset-heavy pages, ensure the UI is settled:

• wait for app-specific loading indicators to disappear
• wait for text that proves the page rendered
• when needed, use agent-browser eval to wait for document.fonts.ready

4. Inspect visually before interacting

Use both structural and visual views:

1. agent-browser snapshot -i to identify actionable refs
2. agent-browser screenshot --annotate when you need to inspect layout, spacing, icons, overlays, or unlabeled controls
3. agent-browser get title and agent-browser get url to confirm you are on the expected page

Annotated screenshots are especially useful for:

• responsive breakpoints
• modal, dropdown, and tooltip positioning
• icon-only controls
• canvas-like or visually rich areas that the accessibility tree underspecifies

5. Exercise the target states

Move through the smallest path that proves the visual behavior.

Common patterns:

• after every navigation or state-changing click, run agent-browser snapshot -i again
• for forms, fill by ref and capture the empty, focused, invalid, and submitted states when relevant
• for menus or modals, capture both closed and open states
• for bug repros, capture before-action and after-action screenshots

6. Cover key visual dimensions

At minimum, check the views that match the request. Common dimensions:

• desktop viewport
• mobile viewport or named device emulation
• light and dark mode
• happy path plus the visually risky state tied to the bug or feature

Typical commands:

agent-browser set viewport 1440 900
agent-browser screenshot shots/home-desktop.png

agent-browser set device "iPhone 14"
agent-browser screenshot shots/home-mobile.png

agent-browser set media dark
agent-browser screenshot shots/home-dark.png

7. Report findings clearly

When you finish, report:

• the PTY session used and whether the site started cleanly
• the URLs, viewports, and themes you checked
• concrete findings with route and state context
• screenshots captured, if relevant to the task
• anything not verified and why

Rules

• Prefer targeted route or component checks over wandering manual exploration.
• Prefer deterministic states over whatever the live app happens to show.
• Re-snapshot after page changes because @e refs can become stale.
• Do not rely on fixed sleeps when a readiness signal or DOM condition is available.
• If visual noise comes from timestamps, ads, rotating banners, or live counters, call that out explicitly and avoid overclaiming.
• Stop the PTY session with pty_kill when it is no longer needed.

Boundaries

• This skill is for visual browser testing and manual UI verification.
• It does not replace functional tests, accessibility checks, or cross-browser certification unless the user explicitly asks for those.
• If the task requires pixel-diff baselines or CI-grade regression suites, recommend adding dedicated visual regression automation after the manual browser pass.