feat: add Playwright support to browser-use extension

[maks-ivanov]

Note

Medium Risk
Large behavioral and API change (steps→actions, different ref semantics) plus a new runtime dependency (playwright) and persistent browser lifecycle, which could affect existing integrations and resource cleanup.

Overview
Replaces the browser-use extension’s agent-browser/CLI-based implementation with an in-process Playwright (Chromium) driver that returns a text-only CDP accessibility tree and executes interactions via regenerated ref IDs.

The browser tool API changes from steps to a bounded actions list (1–7) and narrows supported operations to goto, snapshot, click, fill, press, scroll, and wait, updating output formatting to include URL/title/ref count plus the rendered accessibility tree. Lifecycle handling is added to keep a persistent browser across calls and cleanly close it on session_shutdown, and repo dependencies/docs are updated to add playwright and new setup instructions.

^{Reviewed by Cursor Bugbot for commit 30e4f50. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Partial init failure in ensureDriver leaks Chromium process
  • Wrapped browser initialization in a cleanup catch so partially created Playwright resources are closed before the error is rethrown.

Or push these changes by commenting:

@cursor push 337502ca66

Preview (337502ca66)

diff --git a/extensions/browser-use/index.ts b/extensions/browser-use/index.ts
--- a/extensions/browser-use/index.ts
+++ b/extensions/browser-use/index.ts
@@ -284,11 +284,16 @@
 		return { page: state.page, cdp: state.cdp };
 	}
 	state.headed = headed ?? false;
-	state.browser = await chromium.launch({ headless: !state.headed });
-	state.context = await state.browser.newContext();
-	state.page = await state.context.newPage();
-	state.cdp = await state.context.newCDPSession(state.page);
-	return { page: state.page, cdp: state.cdp };
+	try {
+		state.browser = await chromium.launch({ headless: !state.headed });
+		state.context = await state.browser.newContext();
+		state.page = await state.context.newPage();
+		state.cdp = await state.context.newCDPSession(state.page);
+		return { page: state.page, cdp: state.cdp };
+	} catch (error) {
+		await closeDriver(state);
+		throw error;
+	}
 }
 
 async function closeDriver(state: DriverState): Promise<void> {

_{You can send follow-ups to the cloud agent here.}

^{Reviewed by Cursor Bugbot for commit 30e4f50. Configure here.}

[cursor]

Partial init failure in ensureDriver leaks Chromium process

Low Severity

If chromium.launch succeeds but a subsequent step (newContext, newPage, or newCDPSession) throws, state.browser is set while state.page/state.cdp remain undefined. On the next call, ensureDriver sees the state.page && state.cdp guard as falsy and launches a brand-new browser, overwriting state.browser and orphaning the previous Chromium process. Neither closeDriver nor the session_shutdown handler can ever reach the leaked instance.

 

^{Reviewed by Cursor Bugbot for commit 30e4f50. Configure here.}