ui-test

UI Test — Agentic UI Testing Skill

Test UI changes in a real browser. Your job is to try to break things, not confirm they work.

Three workflows:

• Diff-driven — analyze a git diff, test only what changed
• Exploratory — navigate the app, find bugs the developer didn't think about
• Parallel — fan out independent test groups across multiple Browserbase browsers

How Testing Works

The main agent coordinates — it plans test strategy, delegates to sub-agents, and merges results. Sub-agents do the actual browser testing.

Planning: multiple angles, then execute once

You MUST complete all three planning rounds yourself and output them before launching any sub-agents. Planning happens in your own response — it is NOT delegated to sub-agents. Do not skip ahead to execution.

Round 1 — Functional: What are the core user flows? What should work? Write out each test as: action → expected result.

Round 2 — Adversarial: Re-read Round 1. What did you miss? Think about: different user types/roles, error paths, empty states, race conditions, edge inputs (empty, huge, special chars, rapid clicks).

Round 3 — Coverage gaps: Re-read Rounds 1–2. What about: accessibility (axe-core, keyboard-only), mobile viewports, console errors, visual consistency with the rest of the app?

Deduplicate: Merge all three rounds into one numbered list of tests. Remove overlaps. Assign each test to a group (e.g. Group A, Group B).

Then execute once — launch one sub-agent per group. Each sub-agent receives its specific list of tests to run, nothing more. Sub-agents do not explore or plan — they execute assigned tests and report results.

Output the three rounds, the merged plan, and the group assignments in your response before calling any Agent tool.

Principles for splitting work

• Sub-agents run assigned tests, not open exploration. The main agent hands each sub-agent a specific numbered list of tests. Sub-agents do not plan, explore, or decide what to test — they execute the list and stop.
• The bottleneck is the slowest agent — split work so no single agent has a disproportionate share. Many small agents > few large ones.
• Size the effort to the change — a single component fix doesn't need many agents or many steps. A full-page redesign does. Let the scope of the diff drive the plan.
• No early stopping on failures — find as many bugs as possible within the assigned tests.

Giving sub-agents a step budget

The main agent MUST include an explicit browse step limit in every sub-agent prompt. Sub-agents do not self-limit — they will run until done unless told otherwise.

As a rough heuristic: ~25 steps for a few targeted checks, ~40 for a full page with functional + adversarial + a11y, ~75 for multiple pages or a broad category. Adjust based on what the assigned tests actually require — these are starting points, not rules.

As a rough heuristic: ~25 steps for a few targeted checks, ~40 for a full page with functional + adversarial + a11y, ~75 for multiple pages or a broad category. Adjust based on what the assigned tests actually require — these are starting points, not rules.

Every sub-agent prompt must include:

You have a budget of N browse steps (each `browse` command = 1 step). Count your steps as you go. When you reach N, stop immediately and report:
- STEP_PASS/STEP_FAIL for every test you completed
- STEP_SKIP|<test-id>|budget reached for every test you didn't get to

Do not retry or continue after hitting the budget.
Run only these tests: [numbered list from the merged plan]
Do not explore beyond the assigned tests.
Do NOT generate an HTML report or write any files. Return only step markers and your findings as text.

The main agent should NOT run browse commands itself (except to verify the dev server is up). All testing happens in sub-agents.

When a sub-agent hits its budget, the main agent accepts the partial results as-is. Do not re-run or retry the sub-agent. Include SKIPPED tests in the final report so the developer knows what wasn't covered.

Reporting

Every sub-agent reports back with:

Tests: 8 | Passed: 5 | Failed: 2 | Skipped: 1 | Pages visited: 2

The main agent merges into a final report with:

Tests: 20 | Passed: 14 | Failed: 4 | Skipped: 2 | Agents: 3 | Pass rate: 70%

Do not report "steps used" — browse command counts are implementation plumbing, not a meaningful metric for reviewers.

Testing Philosophy

You are an adversarial tester. Your goal is to find bugs, not prove correctness.

• Try to break every feature you test. Don't just check "does the button exist?" — click it twice rapidly, submit empty forms, paste 500 characters, press Escape mid-flow.
• Test what the developer didn't think about. Empty states, error recovery, keyboard-only navigation, mobile overflow.
• Every assertion must be evidence-based. Compare before/after snapshots. Check specific elements by ref. Never report PASS without concrete evidence from the accessibility tree or a deterministic check.
• Report failures with enough detail to reproduce. Include the exact action, what you expected, what you got, and a suggested fix.

Assertion Protocol

Every test step MUST produce a structured assertion. Do not write freeform "this looks good."

Step markers

For each test step, emit exactly one marker:

STEP_PASS|<step-id>|<evidence>

or

STEP_FAIL|<step-id>|<expected> → <actual>|<screenshot-path>

• step-id: short identifier like homepage-cta, form-validation-error, modal-cancel
• evidence: what you observed that proves the step passed (element ref, text content, URL, eval result)
• expected → actual: what you expected vs what you got
• screenshot-path: path to the saved screenshot (failures only — see Screenshot Capture below)

Screenshot Capture for Failures

Every STEP_FAIL MUST have an accompanying screenshot so the developer can see what went wrong visually.

When a test step fails:

Setup the screenshot directory at the start of any test run:

Rules:

• File name = step-id (e.g., double-submit.png, axe-audit.png, modal-focus-trap.png)
• Store in .context/ui-test-screenshots/ — this directory is gitignored and accessible to the developer and other agents
• For parallel runs, include the session name: <session>-<step-id>.png (e.g., signup-double-submit.png)
• Take the screenshot at the moment of failure — capture the broken state, not after recovery
• For visual/layout bugs, also screenshot the baseline (working state) for comparison: <step-id>-baseline.png

How to verify (in order of rigor)

1. Deterministic check (strongest) — browse eval returns structured data you can inspect. Examples: axe-core violation count, document.title, form field value, console error array, element count.
2. Snapshot element match — a specific element with a specific role and text exists in the accessibility tree. Check by ref: @0-12 button "Save". An element either exists in the tree or it doesn't.
3. Before/after comparison — snapshot before action, act, snapshot after. Verify the tree changed in the expected way (element appeared, disappeared, text changed).
4. Screenshot + visual judgment (weakest) — only for visual-only properties (color, spacing, layout) that the accessibility tree cannot capture. Always accompany with what specifically you're evaluating.

Before/after comparison pattern

This is the core verification loop. Use it for every interaction:

Setup

Avoid permission fatigue

This skill runs many browse commands (snapshots, clicks, evals). To avoid approving each one, add browse to your allowed commands:

Add both patterns to .claude/settings.json (project-level) or ~/.claude/settings.json (user-level):

The first pattern covers plain browse commands. The second covers parallel sessions (BROWSE_SESSION=signup browse open ...). Both are needed to avoid approval prompts.

Mode Selection

Rule: If the target URL contains localhost or 127.0.0.1, pass --local on the first browse open.

Local Mode (default for localhost)

browse open ... --local uses a clean isolated local browser by default, which is best for reproducible localhost QA runs.

Use local-mode variants only when needed:

• browse open <url> --auto-connect — auto-discover an existing debuggable local Chrome. Use this only when the test explicitly needs existing local login/cookies/state.
• browse open <url> --cdp <port|url> — attach to a specific CDP target (explicit local browser attach).

Remote Mode (deployed sites via cookie-sync)

Cookie-sync flags: --domains, --context, --verified, --proxy "City,ST,US"

Workflow A: Diff-Driven Testing

Phase 1: Analyze the diff

Categorize changed files:

Phase 2: Map files to URLs

Detect framework: cat package.json | grep -E '"(next|react|vue|nuxt|svelte|@sveltejs|angular|vite)"'

Phase 3: Ensure the right code is running

Before testing, verify the dev server is serving the code from the diff — not a stale branch.

If testing a PR or specific branch:

If the dev server was already running on a different branch, restart it after checkout.

Find a running dev server:

If nothing found: tell the user to start their dev server.

Verify it actually renders: After browse open + browse snapshot, check that the accessibility tree contains real page content (navigation, headings, interactive elements) — not just an error overlay or empty body. Next.js dev servers can return HTTP 200 while showing a full-screen build error dialog. If the snapshot is empty or dominated by an error dialog, the server is broken — fix the build before testing.

Phase 4: Generate test plan

For each changed area, plan both happy path AND adversarial tests:

Phase 5: Execute tests

For each test, follow the before/after pattern:

Phase 6: Report results

Always browse stop when done.

Phase 7: Generate HTML report

After producing the text report, generate a standalone HTML report that a reviewer can open in a browser. The report embeds screenshots inline (base64) so it works as a single file — no external dependencies.

Why: Text reports are good for the agent conversation, but reviewers (PMs, designers, other engineers) want a visual artifact they can open, scan, and share. Screenshots inline make failures immediately obvious.

How to generate

1. Read the HTML template at references/report-template.html
2. Build the report by replacing the template placeholders with actual test data:

3. For each test result, generate a <details> card. Failed tests should be open by default so reviewers see them immediately:

4. Embed screenshots as base64 so the HTML is fully self-contained:

Read each screenshot file referenced in STEP_FAIL markers, base64-encode it, and embed it as an <img src="data:image/png;base64,..."> in the corresponding test card. For STEP_PASS, only embed a screenshot if one was explicitly taken (e.g., baseline screenshots).

5. Write the final HTML to .context/ui-test-report.html:

6. Tell the user: Report saved to .context/ui-test-report.html and offer to open it.

Rules:

• Failures section comes before passes — reviewers care about what's broken first
• Failed cards are open by default; passed cards are collapsed
• Every STEP_FAIL card MUST have an embedded screenshot — if the screenshot file is missing, note it in the card
• Include the suggestion/fix in each failure card if one was provided
• The report must work offline — no CDN links, no external assets
• Keep the HTML under 5MB — if screenshots push it over, reduce image quality or skip baseline screenshots for passes

Adversarial Test Patterns

Apply these to every interactive element you test. Read references/adversarial-patterns.md for the full pattern library (forms, modals, navigation, error states, keyboard accessibility).

Deterministic Checks

These produce structured data, not judgment calls. Use them as the strongest form of assertion.

For the exact browse eval recipes, read references/browser-recipes.md.

Workflow B: Exploratory Testing

No diff, no plan — just open the app and try to break it. Use this when the user says "test my app", "find bugs", or "QA this site."

Approach

1. Discover the app — read package.json to detect the framework, then open the root URL and snapshot to see what's there
2. Navigate everything — click through nav links, visit every reachable page, note what exists
3. Test what you find — for each page, apply the adversarial patterns below (forms, modals, navigation, keyboard, error states)
4. Run deterministic checks — axe-core, console errors, broken images, form labels on every page
5. Report findings — use STEP_PASS/STEP_FAIL markers, include reproduction steps for failures

Don't try to be systematic about coverage. Just explore like a user would, but with the intent to break things. The agent is good at this — let it roam.

Tips for exploratory runs

• Start with the homepage, then follow the navigation naturally
• Try the 404 page (/does-not-exist) — is it custom or default?
• Look for empty states (pages with no data)
• Test forms with garbage input before valid input
• Check mobile viewport (375px) on every page — does it overflow?
• If the app has auth, use cookie-sync first

Workflow C: Parallel Testing

Run independent test groups concurrently using named browse sessions (BROWSE_SESSION=<name>). Each session gets its own browser. Works with both local and remote mode.

Use when testing multiple pages or categories and you want faster wall clock time.

Read references/parallel-testing.md for the full workflow: session setup, agent fan-out, cookie-sync for auth, and result merging.

Design Consistency

Check whether changed UI matches the rest of the app visually. Read references/design-consistency.md when doing visual or design checks.

Test Categories

Reference guides (load on demand):

• Adversarial patterns — references/adversarial-patterns.md — load when testing forms, modals, navigation, or keyboard a11y
• Browser recipes — references/browser-recipes.md — load when running deterministic checks (axe-core, console, images, form labels)
• Exploratory testing — references/exploratory-testing.md — load for Workflow B (no diff, open exploration)
• UX heuristics — references/ux-heuristics.md — load when evaluating UX quality or citing specific heuristics
• Design system — references/design-system.example.md — template for users to customize
• Design consistency — references/design-consistency.md — load when doing visual consistency checks
• Parallel testing — references/parallel-testing.md — load for Workflow C (concurrent sessions)
• Report template — references/report-template.html — HTML template for Phase 7 report generation

For worked examples with exact commands, read EXAMPLES.md if you need to see the assertion protocol in action.

Best Practices

1. Be adversarial — try to break things, don't just confirm they work
2. Every assertion needs evidence — snapshot ref, eval result, or before/after diff
3. Before/after for every interaction — snapshot, act, snapshot, compare
4. Screenshot every failure — browse screenshot immediately on STEP_FAIL, save to .context/ui-test-screenshots/<step-id>.png
5. Deterministic checks first — axe-core, console errors, form labels before visual judgment
6. For localhost, start with clean local mode — pass --local on the first browse open for reproducible runs; use --auto-connect only when existing local state is required
7. Always browse stop when done — for parallel runs, stop every named session
8. Report failures with reproduction steps — action, expected, actual, screenshot path, suggestion
9. Parallelize independent tests — use Workflow C with named sessions when testing multiple pages or categories on a deployed site

Troubleshooting

• "No active page": browse stop, retry. For zombies: pkill -f "browse.*daemon"
• Dev server not responding: curl http://localhost:<port> — ask user to start it
• browse eval with await fails: Use .then() instead — browse eval doesn't support top-level await
• Element ref not found: browse snapshot again — refs change on page update
• Blank snapshot: browse wait load or browse wait selector ".expected" before snapshotting
• SPA deep links 404: Navigate to / first, then click through
• Remote auth fails: Re-run cookie-sync with --context <id>, try --verified
• Parallel session conflicts: Ensure every browse command uses BROWSE_SESSION=<name> — without it, commands go to the default session
• Session not stopping: BROWSE_SESSION=<name> browse stop. For zombies: pkill -f "browse.*<name>.*daemon"