Repository navigation
[Bug]: MCP browser_navigate blocks for the whole navigation timeout when the page opens a dialog during load #42817
Description
Version
1.64.0-next (main @ 07f1a61)
Steps to reproduce
Serve a page that opens a dialog while loading:
<title>T</title><script>alert('hi')</script><h1>Hi</h1>Then, with a default MCP server (Chromium):
browser_navigate { url: "http://localhost:PORT/" }browser_snapshot {}
Repro with the repo's MCP fixtures and timeouts: { navigation: 3000 } to keep it short:
NAV after 4600 ms: ### Error
TimeoutError: browserBackend.callTool: Timeout 3000ms exceeded.
Call log:
- navigating to "http://localhost:PORT/", waiting until "domcontentloaded"
SNAP= ### Error
Error: Tool "browser_snapshot" does not handle the modal state.
### Modal state
- ["alert" dialog with message "hi"]: can be handled by browser_handle_dialog
Expected behavior
browser_navigate returns as soon as the dialog opens and reports it in the modal state, the same way browser_click does when a click opens a dialog (see alert dialog w/ race in tests/mcp/dialogs.spec.ts). The agent can then call browser_handle_dialog and continue.
Actual behavior
page.goto never reaches domcontentloaded while the dialog is open, so the call blocks for the full navigation timeout (60s by default) and then fails with a TimeoutError. Only after that does the modal state become visible. With the default timeout this is long enough to trip client-side tool timeouts, and one hostile or buggy page stalls the whole server for a minute per navigation.
Additional context
Tab.navigate() (packages/playwright-core/src/tools/backend/tab.ts) awaits page.goto directly, while other actions go through waitForCompletion, which races the action against TabEvents.modalState (_raceAgainstModalStates). Navigation should use the same race, returning early when a modal state appears.
I intend to work on this and will send a PR.
Environment
- Operating System: macOS (Darwin 25.6.0)
- Node.js: 24.8.0
- Browser: Chromium
- Playwright: main @ 07f1a6154
Activity
Hronom commentedon Sep 21, 2026
The repro maps cleanly to the existing modal-state race. A few acceptance details may help keep the fix deterministic:
I maintain Hronaut, a local visible Browser/MCP workspace, and use this action-versus-modal/read-back boundary in evaluator material. I have not run Playwright’s fix, so I’m not claiming a result. AI-assisted note.
chore: mark v0.0.83 (#1779)
chore: mark v0.1.22 (#473)
chore(deps): update all non-major dependencies (#199)
chore(deps): update dependency @playwright/mcp to v0.0.83 (#1108)
chore: mark v0.0.83 (microsoft#1779)
Update agent CLIs