fix(mcp): identify downloads explicitly

[Skn0tt]

This PR stops treating every Chromium net::ERR_ABORTED navigation as a possible download. Since Chromium 140, released in September 2025, Page.navigate reports downloads explicitly, which Playwright surfaces as Download is starting.

Previously, MCP waited up to three seconds for a download event before returning net::ERR_ABORTED. This didn’t change the result, but happened to give the renderer’s crash event time to arrive. The error is now returned immediately. Crash recovery stays the same: once the event arrives, a subsequent MCP call resets the page to about:blank and reports Page crashed and was reset to about:blank.

The crash tests now poll for recovery instead of relying on the download wait, and explicitly await the crash event when asserting tab state.

[github-actions]

Hi, I'm the Playwright bot and I took a first look at the CI failures here.

🟢 The one failure is a pre-existing flake — this PR is clear

The single red test, [chrome] › mcp/cli-navigation.spec.ts:56 › goto chrome:// page that closes the tab does not crash the response, has been flaking on mcp-windows-latest-chrome for weeks, on SHAs this PR can't be responsible for. Nothing here looks caused by your change.

Details

Pre-existing flake / infra

• [chrome] › mcp/cli-navigation.spec.ts:56 › goto chrome:// page that closes the tab does not crash the response @mcp-windows-latest-chrome — pre-existing flake. In the aggregated CI results this test fails 41 of 348 runs (~12%) on mcp-windows-latest-chrome, and never once on the other five MCP bots (0/356 chromium-ubuntu, 0/348 chromium-windows, 0/351 chromium-macos, 0/356 chrome-ubuntu, 0/351 chrome-macos). The failures span 24 unrelated PRs plus many push runs on main — e.g. PR test: fixme partitioned cookie API test on WebKit macOS < 26 #41691 (sha 22ff0ef), PR fix(mcp): await init page before screenshot #41646 (fa1c703), push 39a418a — all predating this branch, all with the identical signature: expect(received).toContain(expected) looking for the "No open tabs. Navigate to a URL to create one." substring. It's isolated to the real-Chrome channel on Windows, where the chrome://extensions/ page occasionally doesn't tear the tab down in time. Your run (#29913780666, PR fix(mcp): identify downloads explicitly #41933) is just one more hit of the same flake.

On the diff itself

Worth noting your change does touch this exact code path — Tab navigation now races a crash event against the download event on net::ERR_ABORTED. But the failure predates the PR with an unchanged rate and message, so it isn't a regression from your rewrite. If anything, waiting on the crash event is the kind of thing that could eventually reduce this flake; it doesn't introduce it.

I'm a first pass, so treat the green as "no PR-caused failures found," not a guarantee the branch is bug-free.

_{Triaged by the Playwright bot - agent run}

[github-actions]

Test results for "MCP"

7786 passed, 1262 skipped

---

Merge workflow run.