Playwright

• Self-contained. The script must run with only BROWSERBASE_API_KEY in the environment. No reliance on autobrowse state, no reading from workspace files.
• CDP attach, never chromium.launch(). Follow the Playwright ↔ Browserbase bridge reference verbatim for the create-session / connectOverCDP / release dance.
• No browser.close(). Release the session via browse cloud sessions update <id> --status REQUEST_RELEASE in finally.
• Final stdout line is JSON. {"success":true,"data":...} on success or {"success":false,"error":"..."} on failure. The runner parses this line — don’t emit any other JSON-looking lines after it.
• Snap on errors. Wrap main() in try { … } catch (err) { await snap(page, '99-error'); throw err; }. Honor process.env.SCREENSHOT_DIR for snap output.
• Locator preferences in order: data-testid attribute → role + name → id → text → xpath. Prefer Playwright’s auto-waiting (locator.click(), locator.fill()) over explicit waits when possible.
• Use the descriptor data when available. Each descriptors.ndjson entry describes the actual DOM target the agent interacted with — pick locators from those attributes / role / accessibleName fields rather than inventing them.
• Use the trace’s network signals. Where the unified events show a slow XHR after an action, insert page.waitForResponse(...) rather than arbitrary sleeps.

Output schema

The script must define a Zod schema that mirrors the # Output section of the task.md provided in context, and validate the extracted data through that schema before printing the final success: true line.

Imports / runtime

playwright and zod are already in the scaffolded package.json. Do not add other dependencies.

What to emit

Output the complete .ts file content. Start with imports, end with a call to main(). Nothing before the first import, nothing after the last closing brace. No markdown fences.