QA Browser Verify

QA Browser Verify

Use this skill for manual QA passes against a running local application.

Default stance

• Prefer agent-browser over curl for UI verification.
• Verify what the user actually sees: rendered content, overlays, console errors, and basic interactions.
• Stay in QA mode: gather evidence first, describe failures clearly, and avoid editing code unless the user asks for a fix.
• Run agent-browser commands sequentially for a single session. Do not fire multiple commands at the same session in parallel.

Setup

1. Identify the dev server URL from the repo or running processes.
2. If agent-browser is missing, stop and report that blocker.
3. If browser launch fails because Playwright browsers are missing, run:

agent-browser install
npx playwright install chromium chromium-headless-shell

4. Use a dedicated session name, for example:

AGENT_BROWSER_SESSION=qa agent-browser open http://localhost:3000

Smoke-test flow

Run this loop for the home route and any important feature routes:

AGENT_BROWSER_SESSION=qa agent-browser open http://localhost:3000
AGENT_BROWSER_SESSION=qa agent-browser screenshot /tmp/qa-home.png
AGENT_BROWSER_SESSION=qa agent-browser snapshot -i
AGENT_BROWSER_SESSION=qa agent-browser eval 'document.querySelector("[data-nextjs-dialog], .vite-error-overlay, #webpack-dev-server-client-overlay") ? "ERROR_OVERLAY" : "OK"'
AGENT_BROWSER_SESSION=qa agent-browser eval 'document.body.innerText.trim().length > 0 ? "HAS_CONTENT" : "BLANK"'
AGENT_BROWSER_SESSION=qa agent-browser console
AGENT_BROWSER_SESSION=qa agent-browser errors

Interaction checks

• Use fill, type, click, press, and get url to validate a real user path.
• After submitting a form or triggering a flow, wait briefly and capture a second screenshot.
• When useful, confirm the page content with get text, eval, or another snapshot -i.

Example:

AGENT_BROWSER_SESSION=qa agent-browser fill 'input[name="prompt"]' 'Hello'
AGENT_BROWSER_SESSION=qa agent-browser click 'button[type="submit"]'
AGENT_BROWSER_SESSION=qa agent-browser wait 3000
AGENT_BROWSER_SESSION=qa agent-browser screenshot /tmp/qa-after-submit.png

Failure capture

When a check fails:

1. Save a screenshot of the broken state.
2. Capture console and errors.
3. If the request flow matters, inspect network requests.
4. Report the exact route, user action, observed result, and expected result.

Closeout

Finish by closing the browser session:

AGENT_BROWSER_SESSION=qa agent-browser close

Return a concise QA summary:

• what routes were checked
• what interactions were tested
• pass/fail status
• concrete evidence for any failure