Browser Use Version

0.13.11 (main at c75e847)

Bug Description, Steps to Reproduce, Screenshots

Clicking a link to an attachment (Content-Disposition: attachment) produces two FileDownloadedEvents for one file in the first tab DownloadsWatchdog attaches to, and one in every other tab:

These are counts from real downloads in headless Chrome. The repro below gives the same result 3 runs out of 3.

Why: both download paths report the same file. Chrome's own download goes through Browser.downloadWillBegin/downloadProgress (auto_download=False). The network path, Network.responseReceived → download_file_from_url, also treats the attachment as downloadable (auto_download=True). Nothing links the two. The network path only dedupes by URL, against itself.

Why only the first tab: attach_to_target returns early once the browser-level download listener exists, before _setup_network_monitoring. So only the first tab ever gets network monitoring. That early return is also why per-tab network monitoring can't simply be turned on yet: every tab would start firing twice. It came up in review on #6022, which leaves the early return alone for this reason. Open PR #6027 makes exactly that change.

Not a duplicate of #5657 or #5519: those fix a different double dispatch, between Chrome's completion event and the filesystem poller, plus partial .crdownload files. Here the second event comes from the network-detection path. #2842 reported the same symptom on 0.7.0 and was closed as fixed in 2025-09. In this case the second event carries auto_download=True, so it comes from the network path.

Happy to open a PR making one path the source of truth if that's wanted.

Failing Python Code

# tests/ci/browser/test_attachment_download_events.py  (uses the repo's tests/ci conftest)
import asyncio

import pytest
from pytest_httpserver import HTTPServer

from browser_use.browser import BrowserProfile, BrowserSession
from browser_use.browser.events import FileDownloadedEvent, NavigateToUrlEvent


@pytest.fixture
async def session(tmp_path):
	s = BrowserSession(
		browser_profile=BrowserProfile(headless=True, user_data_dir=None, keep_alive=True, downloads_path=str(tmp_path))
	)
	await s.start()
	yield s
	await s.kill()


@pytest.mark.parametrize('tab', ['first', 'second'])
async def test_one_event_per_attachment(session: BrowserSession, httpserver: HTTPServer, tab: str):
	httpserver.expect_request('/f.zip').respond_with_data(
		b'PK\x03\x04' + b'\0' * 100,
		content_type='application/zip',
		headers={'Content-Disposition': 'attachment; filename="f.zip"'},
	)
	httpserver.expect_request('/page').respond_with_data('<a id="dl" href="/f.zip">f.zip</a>', content_type='text/html')
	await session.event_bus.dispatch(NavigateToUrlEvent(url=httpserver.url_for('/page')))
	if tab == 'second':
		await session.event_bus.dispatch(NavigateToUrlEvent(url=httpserver.url_for('/page'), new_tab=True))
	events: list[FileDownloadedEvent] = []
	session.event_bus.on(FileDownloadedEvent, lambda event: events.append(event))
	cdp = await session.get_or_create_cdp_session()
	await cdp.cdp_client.send.Runtime.evaluate(
		params={'expression': "document.getElementById('dl').click()"}, session_id=cdp.session_id
	)
	await asyncio.sleep(4)
	print(f'\n{tab} tab: {[(e.file_name, e.auto_download) for e in events]}')
	assert len(events) == 1

first fails with 2 events, second passes with 1.

LLM Model

N/A (no agent; BrowserSession only)

Operating System & Browser Versions

macOS 26.6.2, Google Chrome 154.0.8037.98 (headless)