visual-web-qa

Verify rendered behavior in a browser instead of inferring visual correctness from code. Choose the smallest set of evidence that directly answers the request.

1. Classify the requested behavior as stable, temporal, or exact-state verification.
2. Resolve a valid local, preview, or explicitly requested target URL.
3. Choose representative pages, states, viewports, and themes.
4. Capture browser evidence with meaningful waits and fresh element refs.
5. Share requested artifacts with sendFiles and verify delivery success.
6. Report the exact target, evidence, result, findings, and limitations.

Use screenshots for stable rendered states. Use short videos for timing, motion, loading, or sequence. A purely temporal request may use video without a redundant screenshot. DOM checks support visual evidence; they do not replace it when the user asks how something looks.

agent-browser --session visual-qa screenshot /tmp/visual-qa.png
agent-browser --session visual-qa screenshot --full /tmp/visual-qa-full.png
agent-browser --session visual-qa screenshot --annotate /tmp/visual-qa-issue.png

• Check one to four representative pages or states unless the user requests broader coverage or the change spans more templates.
• Choose the viewport-theme combinations most likely to expose the issue instead of building an exhaustive matrix by default.
• State clearly when only part of the requested surface was verified.

Use this order:

1. User-provided URL
2. Running local development server
3. Repo-native server such as pnpm dev
4. Preview deployment
5. Production as an explicitly requested or read-only baseline
6. file:// output only when the site renders correctly without a server

Do not check production and claim an unmerged change is present. If no valid target is reachable, report blocked and name the missing server, preview, build, or authentication requirement.

Wait for the state that proves progress instead of adding arbitrary delays:

agent-browser --session visual-qa wait --url "**/expected-path"
agent-browser --session visual-qa wait --text "Expected text"
agent-browser --session visual-qa wait --load networkidle
agent-browser --session visual-qa wait 100

• Prefer URL or text waits for route and content changes.
• Use networkidle only when the action triggers real network activity.
• Use a short fixed wait only when an animation has no semantic completion signal.
• Run snapshot -i after navigation or significant DOM changes before using element refs.

For initial loading behavior, start recording before the first navigation:

agent-browser --session visual-qa set viewport 1440 900
agent-browser --session visual-qa record start /tmp/visual-qa-load.webm "$URL"
agent-browser --session visual-qa wait --load networkidle
agent-browser --session visual-qa record stop

For post-load interactions, explore first, then start recording. record start creates a fresh browser context and reloads the page, so discard earlier refs and run snapshot -i again before interacting. Stop recording as soon as the target behavior is captured. Never end a QA run with an active recording.

Use sendFiles for every artifact the user should receive:

{
  "files": [
    { "path": "/tmp/visual-qa.png" },
    { "path": "/tmp/visual-qa-load.webm" }
  ]
}

• Claim an artifact was shared only when sendFiles succeeds in this turn.
• If delivery fails or sendFiles is unavailable, report the error and saved paths without claiming attachment success.

Never capture or share credential entry, session tokens, customer data, or unrelated sensitive UI state. If reaching the target requires exposing that data, use an existing safe authenticated session or report the authentication requirement as a limitation or blocker.

Report:

• Target: exact URL verified
• Evidence: screenshots, videos, and state checks gathered, with why each was chosen
• Result: pass, issues found, or blocked
• Findings: specific rendered behavior observed
• Limitations: requested pages, states, viewports, or themes not verified

Use pass only when the captured evidence matches the requested behavior without an obvious scoped regression. Use issues found for broken layout, incorrect motion, flicker, missing assets, or invalid states. Use blocked when no safe reachable target exists.

Never generalize beyond the evidence collected, and never claim a rendered change looks correct without opening a browser and gathering supporting evidence.