browser-qa

name	browser-qa
description	QA web applications using Chrome DevTools MCP. Use when testing running apps, verifying acceptance criteria against a live UI, checking for console errors, evaluating UI behavior, or running regression checks. Requires the chrome-devtools MCP server to be connected. Complements dm-work:review (code-level) with runtime verification.

Browser QA

Test running web applications by actually navigating, clicking, and asserting — not just reading code.

Scope

This skill covers standard web applications testable via Chrome DevTools. It is NOT suitable for:

• WebGL/Canvas-heavy apps (real-time games, 3D renderers) — CDT snapshots can't inspect canvas internals. Use manual verification with screenshots, or project-specific tools.
• Native mobile apps — use platform-specific QA tools (e.g., Maestro for iOS/Android).
• CLI tools — use bash-based testing, not browser QA.

When the evaluator encounters projects outside browser-qa's scope, it should mark runtime criteria as UNTESTABLE and note the recommended verification method.

When to Use

• After implementation work, before merge — verify acceptance criteria against the live app
• Post-merge regression checks — exercise key user flows after code lands
• Investigating visual or interaction bugs — reproduce in browser, inspect state
• Evaluator protocol — the evaluator uses this skill to grade against sprint contracts
• Design quality checks — screenshot and visually evaluate for AI slop patterns

Prerequisites

• Chrome DevTools MCP connected (chrome-devtools-mcp plugin)
• App running locally — dev server started (the QA agent does NOT start the server)
• Acceptance criteria available — from bead, spec, or derived from the change

QA Workflow

Step 1: Resolve Acceptance Criteria

Find testable criteria from (in priority order):

1. Bead — bd show <id> → check the design field for acceptance criteria
2. Spec file — check docs/ or history/ for the spec that spawned this work
3. Explicit — user or orchestrator provides criteria in the prompt
4. Derived — if nothing else, derive basic criteria from the change:
   • Page loads without console errors
   • Changed elements are visible and interactive
   • No network request failures
   • No regressions in surrounding UI

Step 2: Navigate and Snapshot

navigate_page(url: "http://localhost:<port>/<path>")
take_snapshot()

The snapshot returns a text representation of the page's accessibility tree. Each interactive element has a UID you can target with click, fill, etc.

Always snapshot before interacting — UIDs are ephemeral and change on navigation.

Step 3: Exercise User Flows

For each acceptance criterion, drive the corresponding user flow.

Navigate and verify content:

navigate_page(url: "<target URL>")
wait_for(text: ["<expected content>"])
take_snapshot()    # confirm element is present in a11y tree

Fill and submit a form:

take_snapshot()                                    # get current UIDs
fill(uid: "<input-uid>", value: "<test value>")    # type into field
click(uid: "<submit-uid>", includeSnapshot: true)  # submit + get new state
wait_for(text: ["<success indicator>"])             # wait for result

Multi-field forms (more efficient):

fill_form(elements: [
  { uid: "<email-uid>", value: "[e-mail]" },
  { uid: "<name-uid>", value: "Test User" },
  { uid: "<role-uid>", value: "admin" }
])
click(uid: "<save-uid>", includeSnapshot: true)

Keyboard interaction:

press_key(key: "Enter")           # submit
press_key(key: "Escape")          # dismiss modal
press_key(key: "Tab")             # focus next
press_key(key: "Control+A")       # select all

Drag and drop:

drag(from_uid: "<source>", to_uid: "<target>", includeSnapshot: true)

Step 4: Assert Expected Behavior

After each flow step, verify using the cheapest sufficient method:

Text/element presence (cheapest — prefer this):

take_snapshot()
# Scan the a11y tree for expected text, roles, states

Visual verification (use when layout/styling matters):

take_screenshot()
# Review for correct layout, colors, typography, spacing

Console errors (always check):

list_console_messages(types: ["error"])
# Should be empty. Warnings are informational, errors are failures.

Network request success:

list_network_requests(resourceTypes: ["fetch", "xhr"])
# Verify status codes (200, 201, 204). Flag 4xx/5xx.

Custom JS assertions (for state that isn't visible in the a11y tree):

evaluate_script(expression: "document.querySelector('.toast.success') !== null")
evaluate_script(expression: "window.__APP_STATE__.user.role === 'admin'")

Step 5: Regression Checks

Beyond the specific acceptance criteria, check for collateral damage:

1. Console errors — list_console_messages(types: ["error"]) on every page visited
2. Network failures — any 4xx/5xx in list_network_requests()
3. Responsive — emulate(viewport: "375x667,mobile,touch") then re-check key elements
4. Dark mode (if supported) — emulate(colorScheme: "dark") then screenshot
5. Accessibility — lighthouse_audit(mode: "snapshot") for quick a11y score

Step 6: Report Results

Structure the report as pass/fail per acceptance criterion:

## QA Report: <bead-id or description>

### Acceptance Criteria Results

| # | Criterion | Result | Notes |
|---|-----------|--------|-------|
| 1 | User can navigate to /settings | PASS | |
| 2 | Form validates email client-side | PASS | Shows red border + message |
| 3 | Save button disabled during submit | FAIL | Button stays enabled |
| 4 | Success toast after save | PASS | |
| 5 | No console errors | PASS | 0 errors |

### Regression Checks

- Console errors: 0
- Network failures: 0
- Mobile responsive: PASS
- Lighthouse accessibility: 92

### Verdict: FAIL (1/5 criteria failed)

Failing criteria:
- #3: Save button not disabled during form submission. The button remains
  clickable, allowing duplicate submissions.

Integration with beads: If a criterion fails, create a linked bead:

bd create --title="Save button not disabled during submission" --type=bug --priority=2 --deps discovered-from:<parent-bead>

---

CDT Tool Reference

Tool	Purpose	When to use
navigate_page	Go to URL, back/forward/reload	Start of each flow
take_snapshot	A11y tree with element UIDs	Before any interaction; for text/structure assertions
take_screenshot	Visual capture (PNG)	Layout/styling verification; design quality checks
click	Click element by UID	Buttons, links, toggles
fill	Input/textarea/select by UID	Single form fields
fill_form	Multiple fields at once	Multi-field forms (more efficient)
type_text	Raw keystrokes (no targeting)	When you need to type into the focused element
press_key	Key/combo (Enter, Escape, Ctrl+A)	Submit, dismiss, shortcuts
wait_for	Wait for text to appear	After navigation or async operations
drag	Drag element to element	Drag-and-drop UIs
list_console_messages	Console output (filter by type)	Error checking — always run
list_network_requests	Network activity (filter by type)	API call verification
get_network_request	Single request details + body	Deep API response inspection
get_console_message	Single console message details	Investigating specific errors
lighthouse_audit	A11y, SEO, best practices scores	Quick accessibility check
evaluate_script	Run JS in page context	Custom assertions on app state
emulate	Viewport, dark mode, network, geo	Responsive/a11y/offline testing

Context Efficiency Tips

• Prefer take_snapshot over take_screenshot — text is far cheaper than images in context
• Use take_screenshot only when you need to verify visual appearance (layout, colors, styling)
• wait_for before asserting — async UI needs time to settle after interactions
• list_console_messages(types: ["error"]) is the cheapest smoke test — run it on every page
• For SPAs, always wait_for expected content after navigation — don't assume instant render
• Use includeSnapshot: true on click/fill to get the post-action state in one call instead of two
• Save screenshots to files (filePath) when running multiple checks to avoid bloating context
• lighthouse_audit(mode: "snapshot") is faster than mode: "navigation" (no reload)

Patterns

Quick Smoke Test

Cheapest possible verification — use when you just need to know the page works:

navigate_page(url: "<url>")
wait_for(text: ["<any expected text>"])
list_console_messages(types: ["error"])    # should be empty

Form Submission Flow

navigate_page(url: "<form page>")
take_snapshot()
fill_form(elements: [<fields>])
click(uid: "<submit>", includeSnapshot: true)
wait_for(text: ["<success message>"])
list_console_messages(types: ["error"])
list_network_requests(resourceTypes: ["fetch"])   # verify API call

Visual Regression Check

navigate_page(url: "<page>")
wait_for(text: ["<loaded indicator>"])
take_screenshot(filePath: "qa-screenshots/desktop.png")
emulate(viewport: "375x667,mobile,touch")
take_screenshot(filePath: "qa-screenshots/mobile.png")
emulate(colorScheme: "dark")
take_screenshot(filePath: "qa-screenshots/dark-mobile.png")

Design Quality Spot Check

navigate_page(url: "<page>")
take_screenshot(fullPage: true, filePath: "qa-screenshots/full-page.png")
# Review screenshot for AI slop patterns:
# - Generic purple/blue gradients over white cards
# - Template-default typography and spacing
# - Identical component styling across unrelated sections
# - Over-reliance on rounded corners + drop shadows + blur
# - No distinct identity or mood — looks like every other AI-generated site