Headless UI preview (Playwright + Chrome for Testing)

Designed for the remote-execution container where:

• cdn.playwright.dev / dl.google.com are blocked — use Chrome for Testing via storage.googleapis.com.
• Ubuntu's chromium-browser is a snap stub and won't run.
• MCP servers can't be registered mid-session — use Playwright's Node API directly.

Setup (run once per fresh container)

> ⚠️ This project uses pnpm (frontend/pnpm-lock.yaml is the source of truth). > Do NOT run npm i here — npm ignores the pnpm lockfile and re-resolves > transitive deps, which pulls broken versions (e.g. es-toolkit@1.47.0 vs the > pinned 1.46.1) that crash the whole SPA at load. See Troubleshooting below.

playwright is already a committed devDependency in frontend/package.json, so a plain install provides it — no ad-hoc add/revert needed. @emotion/react / @emotion/styled are likewise already declared.

bash
# 1. Install deps from the lockfile (also fixes a node_modules previously
#    polluted by npm — wipe it first with `rm -rf node_modules` if you suspect that).
cd frontend
pnpm install --frozen-lockfile

# 2. Download Chrome for Testing (Playwright's own Chromium download is blocked)
mkdir -p /tmp/chrome && cd /tmp/chrome
curl -fsSL -o chrome.zip \
  "https://storage.googleapis.com/chrome-for-testing-public/148.0.7778.96/linux64/chrome-linux64.zip"
unzip -q chrome.zip          # → /tmp/chrome/chrome-linux64/chrome

Do NOT modify frontend/package.json / frontend/pnpm-lock.yaml for tooling anymore — everything needed is already committed. If a previous session left them dirty (or npm polluted the tree), reset with:

bash
cd <repo> && git checkout -- frontend/package.json frontend/pnpm-lock.yaml
pnpm -C frontend install --frozen-lockfile

Dev server

bash
cd frontend
USE_MOCK=true node_modules/.bin/vite --mode mock --port 3000 &
until curl -fs http://localhost:3000 >/dev/null; do sleep 1; done

USE_MOCK=true must be a shell env var (read by vite.config.ts before .env.mock).

Auth seeding

Every script must seed localStorage.user_auth_token via addInitScript or the app redirects to the login screen:

js
await page.addInitScript(() => {
    localStorage.setItem('user_auth_token', 'mock-token-for-screenshots');
});

First-run onboarding guides

Mock mode defaults these off (src/mocks/mockConfig.ts), so a fresh browser context no longer pops the "Direct Routing Guide" dialog and steals focus/clicks — no Escape-key dismissal needed before interacting with the page. If a script specifically needs to screenshot or assert the first-run guide itself, force it on per-navigation with a query param (no code change, no shared state between runs):

js
await page.goto('http://localhost:3000/agent/openai?mockOnboarding=on', { waitUntil: 'networkidle' });

?mockOnboarding=off is the (redundant) explicit default, useful when a test wants to assert the off state regardless of what a prior run left in localStorage.

Scripts

All scripts live here and are run from frontend/. They use createRequire(cwd) to resolve playwright from frontend/node_modules regardless of the script's location:

js
import { createRequire } from 'module';
const { chromium } = createRequire('file://' + process.cwd() + '/')('playwright');

| Script | Mode | Purpose | |--------|------|---------| | screenshot.mjs | mock | Ad-hoc template — copy here and customise; do NOT commit | | docs-screenshots.mjs | mock | All 9 docs/images/ product screenshots + theme previews | | regression-credentials.mjs | mock | Assertion-based regression for /credentials Add API Key flow | | scenario-routing-graph.mjs | real backend | Codex routing graph screenshots (requires running Go server) |

Each script is self-documenting — see its file header for usage, outputs, and known issues.

Shrinking screenshots (optional)

Screenshots are captured with Playwright's plain page.screenshot() — the skill deliberately does not compress images internally, because optimizers depend on host tooling that is not guaranteed in the run environment. If a PNG is large and you want to trim it before committing, run any optimizer you have available on the captured file afterwards. None of these are required:

• pngquant --quality 80-95 --strip --force --out foo.png foo.png (lossy, smallest PNGs)
• oxipng -o 4 --strip safe foo.png (lossless)
• Node: sharp(file).png({ quality: 80, palette: true }).toFile(out)
• Python: Pillow → img.save(out, optimize=True)

Prefer keeping the original capture when quality matters more than size.

After capturing

1. SendUserFile the PNGs so the user can review them.
2. For ad-hoc screenshot.mjs: delete it before committing — the stop hook will flag it.
3. docs/ is gitignored; force-add images: git add -f docs/images/.
4. Free the port: fuser -k 3000/tcp or pkill -f "vite --mode mock".

Troubleshooting

Blank white page / root is empty, console shows TypeError: require_isUnsafeProperty is not a function (stack points at es-toolkit/dist/compat/object/get.js ← recharts).

Cause: node_modules was installed/resolved by npm instead of pnpm, so a transitive dep (es-toolkit) drifted to a version whose CJS→ESM interop esbuild mis-bundles. It breaks the React vendor chunk, so every route renders blank — not just the page you're testing. Tweaking optimizeDeps (include/exclude/ keepNames) does not fix it.

Fix — reinstall the pinned tree with pnpm:

bash
cd frontend
rm -rf node_modules node_modules/.vite        # drop the npm-polluted tree + dep cache
pnpm install --frozen-lockfile
# verify the pinned (working) version is what's on disk:
ls node_modules/.pnpm | grep es-toolkit       # expect es-toolkit@1.46.1, NOT 1.47.0
pnpm add -D playwright                         # re-add the tool

Then restart the dev server. If you edited vite.config.ts while chasing this, revert it — the config is not the problem.

Dev-server process exits immediately (e.g. exit 144) on --force: usually a follow-on of the crash above (the unhandled client error). Once the pnpm tree is correct it starts cleanly; --force is only needed once to drop a stale node_modules/.vite optimize cache.