AI-powered adversarial UI testing via the browse CLI. Analyzes git diffs to test only what changed, or explores the full app to find bugs. Tests functional correctness, accessibility, responsive layout, and UX heuristics. Use when the user asks to test UI changes, QA a pull request, audit accessibil

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: 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: The main agent merges into a final report with: 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: or 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 Target Mode Command Auth localhost / 127.0.0.1 Local browse open <url local None needed (clean isolated local browser by default) Deployed/staging site Remote browse open <url remote Browserbase credentials; use contexts where supported 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: File pattern UI impact What to test .tsx , .jsx , .vue , .svelte Component Render, interaction, state, edge cases pages/ , app/ , src/routes/ Route/page Navigation, page load, content, 404 handling .css , .scss , .module.css Style Visual appearance (screenshot), responsive form , input , field Form Validation, submission, empty input, long input, special chars modal , dialog , dropdown Interactive Open/close, escape, focus trap, cancel vs confirm nav , menu , header Navigation Links, active states, routing, keyboard nav Non UI files only None Skip — report "no UI tests needed" Phase 2: Map files to URLs Detect framework: cat package.json grep E '"(next react vue nuxt svelte @sveltejs angular vite)"' Framework Default port File → URL pattern Next.js App Router 3000 app/dashboard/page.tsx → /dashboard Next.js Pages Router 3000 pages/about.tsx → /about Vite 5173 Check router config Nuxt 3000 pages/index.vue → / SvelteKit 5173 src/routes/+page.svelte → / Angular 4200 Check routing module 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](references/report template.html) 2. Build the report by replacing the template placeholders with actual test data: Placeholder Value {{TITLE}} Report title for <title tag (e.g., "UI Test: PR 1234 — OAuth Settings") {{TITLE HTML}} Report title for the visible <h1 . If a PR URL is available, wrap the PR reference in an <a tag so it's clickable (e.g., UI Test: <a href="https://github.com/org/repo/pull/1234" PR 1234</a — OAuth Settings ). If no URL, use plain text same as {{TITLE}} . {{META}} One line context: date, app URL, user, branch {{TOTAL TESTS}} Total STEP PASS + STEP FAIL count {{AGENT COUNT}} Number of sub agents that ran {{PASS COUNT}} Number of STEP PASS {{FAIL COUNT}} Number of STEP FAIL {{PASS RATE}} Integer percentage (e.g., "92") {{RATE CLASS}} good (≥90%), warn (70–89%), bad (<70%) {{FAILURES SECTION}} HTML for failed test card

ui-test