ui-test

Run the diff-driven workflow: the agent analyzes your git diff, maps changed files to routes, starts a local browser with browse open --local, and executes happy-path plus adversarial tests. Results are reported as STEP_PASS/STEP_FAIL markers with screenshots for failures.

Playwright verifies intentions you wrote tests for, while this approach explores blind spots like empty states, XSS payloads, rapid clicks, and keyboard-only flows. It also runs axe-core audits and visual judgment checks that scripted suites typically omit.

Yes, use the cookie-sync script to sync cookies from local Chrome into a Browserbase context, then create a session with that context ID and connect via its CDP URL. This lets tests run against authenticated staging dashboards without manual login.

Yes, localhost testing uses browse open <url> --local, which launches a clean isolated local browser with no API key required. The BROWSERBASE_API_KEY is only needed for remote testing of deployed sites.

The browse eval command does not support top-level await, so async operations must use .then() chains instead. For example, run axe.run().then(r => JSON.stringify(...)) rather than awaiting the result directly.

Assign each test group a unique BROWSE_SESSION name so every session gets an independent browser, then fan out groups concurrently via the Agent tool. Each agent must prefix every browse command with its session name and stop its session when finished.