Browser Test — Code-as-Action Browser Verification

Verify web features by writing Playwright scripts, not by imagining what a browser would show. The script is the durable artifact: it lands in the project's test library and re-runs on every future change to that flow. (Concept adapted from Microsoft's Webwright: treat the browser as a disposable environment your code spawns; the workspace — scripts, logs, screenshots — is the state.)

State layout

.claude/idev/browser-tests/
├── scripts/        # one script per flow — the accumulated E2E library
├── artifacts/      # screenshots, console/network logs per run (gitignorable)
└── reports/        # structured test reports

Setup detection (in order)

1. Project already has @playwright/test in package.json → use it (npx playwright test). Respect the project's playwright.config.*.
2. Python playwright importable → use the sync API (python3 script.py).
3. Neither → tell the user the one-time setup (npm i -D @playwright/test && npx playwright install chromium) and ask before installing anything.

Script conventions

• Reuse first: grep scripts/ for the flow before writing a new script. Adapt and parameterize an existing one rather than duplicating it.
• One flow per script, named for the flow: login-flow.spec.ts, checkout-smoke.spec.ts.
• Base URL from env, never hardcoded: const BASE = process.env.BASE_URL ?? 'http://localhost:3000' — scripts must run against dev, staging, or CI unchanged.
• Assert real outcomes, not just "page loaded": visible text, URL after navigation, row counts, toast messages. Prefer role/label selectors (getByRole, getByLabel) over CSS chains — they survive refactors.
• Capture evidence on every run: screenshot at each key step and on failure; collect console errors and failed network requests:

import { test, expect } from '@playwright/test';
const BASE = process.env.BASE_URL ?? 'http://localhost:3000';

test('user can log in', async ({ page }) => {
  const consoleErrors: string[] = [];
  page.on('console', m => m.type() === 'error' && consoleErrors.push(m.text()));
  page.on('requestfailed', r => consoleErrors.push(`NET ${r.url()}`));

  await page.goto(`${BASE}/login`);
  await page.getByLabel('Email').fill(process.env.TEST_USER ?? 'test@example.com');
  await page.getByLabel('Password').fill(process.env.TEST_PASS ?? 'test');
  await page.getByRole('button', { name: 'Sign in' }).click();

  await expect(page).toHaveURL(/dashboard/);
  await page.screenshot({ path: '.claude/idev/browser-tests/artifacts/login-ok.png' });
  expect(consoleErrors, `console/network errors: ${consoleErrors}`).toHaveLength(0);
});

• Credentials via env vars only (TEST_USER, TEST_PASS) — never commit secrets into scripts; note required vars at the top of the script.
• Headless by default; headed only while debugging a script.

Failure discipline (the important part)

When a run fails, classify before acting:

• Script bug (wrong selector, missing wait, bad assumption) → fix the script and re-run. Max 3 repair iterations; then report what's blocking.
• App bug (console error, failed request, wrong behavior) → do NOT silently change the app to make the test pass. Capture the evidence (screenshot, error text, request) and report it as a finding.
• Environment issue (server not running, missing test data) → report what's needed; see the project's run conventions.

Report format

Write to reports/<YYYY-MM-DD>-<flow>.md:

# Browser Test Report: <flow> — <date>
Target: <BASE_URL>   Script: scripts/<name>   Result: PASS | FAIL

| # | Check | Result | Evidence |
|---|-------|--------|----------|
| 1 | login redirects to /dashboard | PASS | artifacts/login-ok.png |
| 2 | zero console errors | FAIL | "TypeError: x is undefined" |

## App bugs found
- <error + repro + evidence path>  (also log per the lessons-learned skill if recurring)

## Not covered
- <flows/states this run did not exercise>

Report only what actually ran — no inferred results. If a check didn't execute, it's "not covered", not "pass".

Anti-patterns

• Verifying UI changes by reading the JSX/HTML and declaring it works — that's what this skill exists to replace.
• Writing a throwaway script and deleting it — save it to scripts/; the library is the point.
• waitForTimeout(3000) sprinkled to "fix" flakiness — wait for conditions (expect(...).toBeVisible(), waitForURL) instead.
• Testing through 10 UI steps what one API call could set up — use the API for arrange, the browser for act/assert.

Concept adapted from Webwright (MIT, © Microsoft Corporation).