[Desktop/Windows] Built-in Browser pane loading a Cloudflare challenge page crashes the GPU process and hangs the whole app; reinstall wipes claude-code-sessions

Summary

Claude Desktop (Windows, MSIX) dies silently when the built-in Browser pane (mcp__Claude_Browser__navigate) loads a page that serves a Cloudflare bot-verification challenge ("请稍候…/正在进行安全验证"). main.log shows the Electron GPU process crashing ~2 s after the navigation, after which the main process logs nothing more until the app is relaunched. Reproduced on two app versions; a third occurrence shows the same signature.

After the user reinstalled the app to recover, the reinstall wiped %APPDATA%\Claude (including claude-code-sessions\…\local_*.json), so every previously registered session now shows "Can't reach your computer … Remote Control host unreachable (computer_unreachable)" in the session list, while the CLI transcripts in ~/.claude/projects are intact.

Environment

• Windows 11 Home 10.0.26200, Modern Standby laptop, NVIDIA GPU
• Claude Desktop MSIX Claude_1.34493.1.0_x64__pzs8sxrjxfjjc (today) and Claude_1.30096.5.0 (2026-08-18 occurrence)
• Bundled CLI 2.1.237 (%APPDATA%\Claude\claude-code\2.1.237\claude.exe)

Steps to reproduce (as observed)

1. In a Claude Code desktop session, the agent calls mcp__Claude_Browser__navigate on a site fronted by a Cloudflare challenge (here https://slidescience.co/...; on 2026-08-12 it was https://www.bcix.de/... / https://exainfra.net/...).
2. The page returned is the challenge interstitial (title "请稍候…", body "正在进行安全验证…"); the agent calls get_page_text once.
3. ~2 s later the GPU process crashes and the whole app becomes unresponsive (window stays but nothing works); no Crashpad dump, no Windows Application-Error event.

Evidence (local time UTC+8, 2026-08-22)

%LOCALAPPDATA%\Claude\logs\main.log:

04:51:20 [warn] [PreviewContext] Blocked subresource to private-resolving host { resourceType: 'xhr' }
04:51:22 [info] GPU process gone: { type: 'GPU', reason: 'crashed', exitCode: 101457950, serviceName: 'GPU' }
   (nothing further until 04:56:36 — next app start after reinstall)

unknown-window.log at 04:51:22 (the challenge page running in the pane):

[warn] The powerPreference option is currently ignored when calling requestAdapter() on Windows. See https://crbug.com/369219127

Windows System log, 4 s after the navigate call, both occurrences (and on 08-12 / 08-10 with the same CLSID):

DistributedCOM 10016: … COM Server application with CLSID {2593F8B9-4EAF-457C-B68A-50F6B8EA6B54} (PerAppRuntimeBroker)
… running in the application container Claude_1.34493.1.0_x64__pzs8sxrjxfjjc

Session transcript (~/.claude/projects/<proj>/<session>.jsonl) ends on the get_page_text tool_result at 04:51:20; ~/.claude has zero writes from 04:51:23 until the relaunch at 04:57:55.

The 2026-08-18 07:11 occurrence had the identical shape: last tool call = built-in browser, DCOM 10016 from the Claude container 1 s before the process tree vanished, 7 h gap until the user reset the app.

Impact

• Hard hang/death of the whole desktop app (all concurrent sessions) triggered by a single page load in the Browser pane.
• Recovery by reinstall deletes %APPDATA%\Claude, orphaning every session's local record: the cloud session list keeps the Remote-Control entries, so each shows "Can't reach your computer" forever, even though the transcripts still exist on disk.

Expected

• GPU-process crash inside the Browser pane should not take down / hang the main process (pane should show a crashed-tab state instead).
• Reinstall/repair should not wipe claude-code-sessions (or the app should re-adopt local transcripts whose cliSessionId exists under ~/.claude/projects).

Happy to provide the full main.log / transcript privately.