Regression: every tool call hangs 30s (browserBackend.callTool: Timeout 30000ms) on a classic <frameset> page whose load never fires — 0.0.78 vs 0.0.76 #41784

Activity

1. Dmitry Gozman (dgozman)

   added

   needs-triage

   on Jul 15, 2026
2. github-actions commented on Jul 15, 2026

   

   github-actionsBot

   on Jul 15, 2026 – with GitHub Actions

   Contributor

   More actions

   Hi, I'm the Playwright bot and I took a first look at this.

   Reproduced on tip-of-tree (npm 1.62.0-next, sha 3c71159). This is a real bug, and your diagnosis is right: on a classic <frameset> where a child frame holds a connection open so the top load never fires, the AI-mode aria snapshot hangs until the timeout — while raw page.evaluate returns instantly. That snapshot runs on every MCP tool response, which is why every call times out at 30s.

   Root cause is in playwright-core, not the MCP layer specifically. The mode: 'ai' snapshot recurses into child iframes (ariaSnapshotForFrame in packages/playwright-core/src/server/page.ts) and polls each child frame's body with retryWithProgressAndTimeouts([1000, 2000, 4000, 8000]). A frame whose document never commits/settles never resolves, so the recursion keeps polling until the overall deadline. The default/locator snapshot doesn't descend into iframes, which is why 0.0.76 (bundled 1.61) stopped at the frameset shell and was fast — the iframe distillation in 1.62 is what walks into the never-settling frame. --snapshot-mode none doesn't help because the tool backend still captures a snapshot for its own bookkeeping.

   Minimal repro (tests/page, npm run ctest)

   import { test, expect } from './pageTest';
   
   test('ai snapshot hangs on frameset with never-loading child frame', async ({ page, server }) => {
     // Child frame that never finishes loading (holds the connection open) ->
     // the top frameset load never fires.
     server.setRoute('/hang', (req, res) => {
       res.writeHead(200, { 'content-type': 'text/html' });
       res.write('<html><body>hanging');
       // never res.end()
     });
     server.setRoute('/frameset', (req, res) => {
       res.writeHead(200, { 'content-type': 'text/html' });
       res.end('<frameset cols="50%,50%"><frame src="/hang"><frame src="about:blank"></frameset>');
     });
   
     await page.goto(server.PREFIX + '/frameset', { waitUntil: 'commit' });
   
     // Raw evaluate is instant.
     const t0 = Date.now();
     await page.evaluate(() => location.href);
     expect(Date.now() - t0).toBeLessThan(3000); // passes (~a few ms)
   
     // AI-mode snapshot hangs until the timeout.
     const t1 = Date.now();
     await page.ariaSnapshot({ mode: 'ai', timeout: 20000 });
     expect(Date.now() - t1).toBeLessThan(5000); // fails: TimeoutError 20000ms exceeded
   });

   page.evaluate returns in a few ms; page.ariaSnapshot({ mode: 'ai' }) times out. Chromium, headless — matches your setup.

   I didn't dig into a fix, and I couldn't A/B every detail of your app, but your 0.0.76-vs-0.0.78 evidence lines up exactly with what I see: the regression is the AI snapshot descending into a frame that never reaches load.

   _{Triaged by the Playwright bot - agent run}
3. Dmitry Gozman (dgozman)

   added

   v1.62

   on Jul 16, 2026
4. Dmitry Gozman (dgozman)

   self-assigned this

   on Jul 16, 2026
5. Dmitry Gozman (dgozman)

   closed this as completedin #41816on Jul 16, 2026