Using Playwright with Kiln

Two separate things share one browser install:

• playwright-cli — a browser you drive from the shell, for looking at the UI you are changing and taking screenshots of it. This is what you want when building UI.
• npm run tests:e2e — the end-to-end suite, in app/web_ui/tests/e2e.

If neither works, the install is probably missing: bash .config/utils/setup_env.sh --add-playwright adds the browser and playwright-cli (~800 MB together). See that script's --help for what a cloud environment needs.

Looking at the UI

Start a server, open the app, and land on a page with data in it:

.agents/scripts/playwright_server.sh start     # prints http://localhost:6544

start prints the next three commands with the seeded ids already filled in. Run all three, in that order — the app redirects to a task picker without the localStorage state, whatever URL you ask for:

playwright-cli open http://localhost:6544
playwright-cli localstorage-set ui_state \
  '{"current_project_id":"<id>","current_task_id":"<id>","selected_model":null}'
playwright-cli goto http://localhost:6544

Then goto any deep link you like. stop when you are done, status if you are not sure what is running.

The sandbox keeps its data in app/web_ui/.agent_dev_home and is seeded from a committed fixture project, so the screens have content instead of an onboarding wizard. It never touches real Kiln projects.

A provider may be connected — spend it carefully

If OPENROUTER_QA_KEY is set in the environment, seeding writes it into the sandbox's settings and the app comes up with OpenRouter connected, so live model calls work. start reports it when it is there — assume nothing is connected until it says so.

When it is connected, it is a real key on a hard, low limit — a runaway loop hits it and the next person gets nothing. Use it when a live call is the only way to check what you are working on, and:

• Default to GPT-5.6 Luna. The ui_state hint start prints already preselects it in the model dropdown, so running as-is costs you nothing and picks the cheap model. Change it only when the model itself is what you are testing.
• Keep every run minimal — one sample, one eval row, one query.
• Never aim a paid test suite at it (pytest --runpaid / --runprerelease read OPENROUTER_API_KEY, which is deliberately a different variable).
• Treat a 402 or 429 as budget exhausted: stop and report it, don't retry.

references/seeded_project.md has the details.

Three rules that will cost you an afternoon otherwise

Never trust the first frame. A screenshot taken right after open or goto is very often blank white, and nothing says so. snapshot is not a usable gate — the DOM is complete before Chromium paints. Settle the page first:

playwright-cli run-code "async page => await page.waitForLoadState('networkidle')"
playwright-cli screenshot --filename=/tmp/ui.png

Never redirect a playwright-cli failure into /dev/null. It reports failures on stdout and exits 1, so >/dev/null 2>&1 — the thing you add to keep a transcript readable — hides the entire reason a command failed. Discard output if you must, but then test $?.

Never guess a locator's role. Kiln styles links as buttons. Run find first and write the locator from what it reports — and prefer clicking the ref it gives you, which is a single element by construction and cannot go strict-mode ambiguous.

Reference

Load only what the task needs.