[Bug] playwright-cli: one discarded (Memory Saver) tab makes every command time out — tab header rendering awaits page.title() without timeout

[Emericen]

Version

@playwright/cli 0.1.15 (bundled playwright-core 1.62.0-alpha-2026-06-29)

Steps to reproduce

Environment: Windows 11, Chrome 150.0.7871.101, attaching to a long-running real Chrome profile.

1. Run Chrome with --remote-debugging-port=9222 --user-data-dir=<dedicated profile> and use it normally for a few days with several tabs open (Stripe/Azure/Vercel dashboards in my case).
2. Chrome's Memory Saver discards/freezes one of the idle background tabs. (Likely forcible on demand via chrome://discards → "Urgent discard", though my occurrence was organic.)
3. playwright-cli attach --cdp=http://localhost:9222 → succeeds.
4. Run any command: playwright-cli tab-list, eval "location.href", snapshot, …

Expected behavior

Commands work, with the discarded tab either listed as-is (title from target info) or marked unresponsive.

Actual behavior

Every command fails after 30s with:

### Error
TimeoutError: Timeout 30000ms exceeded.

The daemon is fully wedged — even commands that don't touch the bad tab (e.g. eval on the current tab) fail, because every command's response renders the "Open tabs" / "Page" header sections.

Diagnosis

The discarded tab's renderer never answers renderer-bound CDP calls. Since #41128 (fix for #41093), connectOverCDP no longer hangs at connect time on such a tab — it reports the page "as-is". The result is a half-initialized Page whose url() returns "" and whose title() / evaluate() never resolve (they don't recover even on fresh connections, or after the renderer is revived via Target.activateTarget / Page.navigate; only closing the tab clears it).

The tools backend then hangs on it for every command: Tab.headerSnapshot() does

await this._raceAgainstModalStates(async () => {
  title = await this.page.title();   // no timeout, no catch
});

(packages/playwright-core/src/tools/backend/tab.ts, visible at lib/coreBundle.js:63606 in the published bundle) and response rendering awaits headerSnapshot() for all tabs on every command. One zombie tab ⇒ every command times out.

Probing each page with Promise.race([page.title(), timeout]) over a raw connectOverCDP connection pinpoints the culprit: five healthy tabs answer in ms; the discarded one (empty url()) hangs on both title() and evaluate('1+1'), while browser-side CDP on the same target (Target.getTargetInfo) still answers fine and even knows the real title/URL.

Suggested fix

Guard the per-tab header with a small timeout and/or .catch() (like collectConsoleMessages already does), falling back to TargetInfo.title/url from the browser session, and render the tab as e.g. (unresponsive). Related report of the same underlying condition in another CDP-attaching agent tool: vercel-labs/agent-browser#1036.

Workaround

Close and recreate the discarded tab via a browser-level CDP session (Target.closeTarget + Target.createTarget) — the CLI heals instantly.

[github-actions]

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

Confirmed by reading the source on tip-of-tree (sha 9fb3602) — your diagnosis is spot on. This is a real bug in the @playwright/cli tools backend (which lives in this repo, so it's ours to fix). A single tab whose renderer never answers renderer-bound CDP calls wedges every command.

The two pieces line up exactly as you described:

• Tab.headerSnapshot() awaits page.title() with no timeout and no .catch() — unlike the neighbouring collectConsoleMessages(), which already guards with .catch(() => []):

  // packages/playwright-core/src/tools/backend/tab.ts
  await this._raceAgainstModalStates(async () => {
    title = await this.page.title();   // no timeout, no catch
  });

• Response rendering fans headerSnapshot() out across all tabs on every command:

  // packages/playwright-core/src/tools/backend/response.ts:274
  const tabHeaders = await Promise.all(this._context.tabs().map(tab => tab.headerSnapshot()));

  (also in backend/tabs.ts:62)

So one discarded tab whose title() never resolves blocks the Promise.all, and every command — even ones that don't touch the bad tab — dies at the 30s timeout. That matches your report precisely.

What I checked

• Read packages/playwright-core/src/tools/backend/tab.ts, response.ts, and tabs.ts on tip-of-tree.
• Verified headerSnapshot() → page.title() has no timeout/catch, while collectConsoleMessages() two methods up does use .catch().
• Verified both render callers (response.ts:274, tabs.ts:62) Promise.all over every tab, confirming that one hung tab stalls all output.

I didn't stand up a synthetic frozen/discarded Chrome tab to reproduce the CDP hang end-to-end — that condition is hard to force deterministically — but the code paths you point at are exactly as described, so the failure mode is unambiguous. Your suggested fix (bound the per-tab header with a small timeout/.catch() and fall back to TargetInfo.title/url, rendering the tab as unresponsive) is the right shape.

Triaged by the Playwright bot · run

[yury-s]

Related bug #41093 and the fix that we ended up not landing #41128