Browser

Open a named session with playwright-cli -s=name open <url> --persistent, run snapshot to get element refs, then use click, fill, or select with those refs. Always close the session with playwright-cli -s=name close when finished.

Run bunx playwright screenshot "<url>" /tmp/shot.png for a one-shot capture without managing a session. For PDFs, use bunx playwright pdf, and for raw HTML use Chrome with --headless=new --dump-dom.

Use playwright-cli for deterministic multi-step tasks like navigation, clicking, and form filling since it costs zero tokens. Use BrowserAgent only when AI reasoning is needed to decide what to do next, such as diagnosing console errors or handling dynamic content.

Yes, each playwright-cli -s=<name> creates an isolated browser instance, so multiple named sessions run simultaneously. The ReviewStories workflow also fans out YAML user stories to up to 8 parallel UIReviewer agents.

Use headed Chrome via claude --chrome, which connects to your real browser with existing cookies, sessions, and extensions. This tier handles authenticated flows that headless sessions cannot replicate, but it is limited to a single non-parallel session.

Sessions opened with --persistent keep the browser process alive between commands. Failing to run playwright-cli -s=name close leaves a zombie browser process consuming system resources.