Minimal Electron reproduction for Cloudflare Turnstile error 600010 looping
inside an embedded <webview> (the failure seen in T3 Code preview,
pingdotgg/t3code#5002).
Turnstile loops with error 600010 whenever the guest's User-Agent string is
rewritten to anything other than the native Electron UA. A plain Electron
<webview> with its untouched UA passes. This is a controlled single-variable
experiment: instrumentation (see below) shows that across all modes the
JS-visible navigator.userAgentData is identical and no Sec-CH-UA request
headers are sent at all, so the UA string is the only thing that changes —
and it alone flips the outcome.
Measured on Ubuntu, Electron 41.5.0 / Chromium 146, each mode in a fresh
--user-data-dir, against https://dash.cloudflare.com/login:
| Mode | UA header = navigator.userAgent (always agree) |
Turnstile |
|---|---|---|
| (none) native | … <app>/0.0.0 Chrome/146.0.7680.216 Electron/41.5.0 Safari/537.36 |
passes |
REPRO_UA=noop re-set unchanged via setUserAgent() |
same string as native | passes |
REPRO_UA=appstrip app token stripped, Electron kept |
… Chrome/146.0.7680.216 Electron/41.5.0 Safari/537.36 |
fails (600010 loop) |
REPRO_UA=modified Electron token stripped |
… <app>/0.0.0 Chrome/146.0.7680.216 Safari/537.36 |
fails (600010 loop) |
Constant across all four modes (dumped from the guest and from
https://httpbin.org/headers):
navigator.userAgentData.brands:Not-A.Brand 24,Chromium 146(noGoogle Chrome, no Electron), platformLinux, fullVersionListChromium 146.0.7680.216.- Request headers contain no
Sec-CH-UA*headers at all (Electron 41<webview>default), so the server sees no Client-Hints signal either way. - The
User-Agentrequest header andnavigator.userAgentalways agree.
noop passing while appstrip/modified fail rules out the
setUserAgent() call itself; only the string content matters — and no single
token is the trigger, since keeping the honest Electron/… token
(appstrip) still fails.
Turnstile is a closed system, so the exact discriminator is not observable
from outside. A plausible reading is that it validates the UA string against
its own fingerprint of the real browser environment and accepts the exact
native pattern while rejecting hand-edited variants — but the repro only
proves the what, not the why. The actionable conclusion is: do not
rewrite the UA string (or, if an identity change is truly needed, override
userAgentMetadata together with the UA via CDP
Network.setUserAgentOverride, which keeps every layer consistent).
Also ruled out as causes (toggles below): pre-granted permission handlers,
BrowserWindow vs <webview>, and — in the T3 Code investigation — CDP
debugger attachment, the preview preload / contextIsolation, and WebGL/GPU
state.
npm install # installs [email protected] (Chromium 146)
# baseline: plain webview, native UA -> Turnstile PASSES
npm start
# strip the Electron token -> FAILS (600010 loop)
REPRO_UA=modified npm start
# strip only the app token, keep Electron/… -> FAILS (600010 loop)
REPRO_UA=appstrip npm start
# re-set the UNCHANGED native string -> PASSES
REPRO_UA=noop npm startOther toggles: REPRO_PERMS=grant (pre-grant notifications/geolocation/
clipboard — passes), REPRO_MODE=browser (BrowserWindow instead of webview —
passes), REPRO_URL=… (point the guest elsewhere, e.g.
https://httpbin.org/headers to see exactly which headers are sent). The
host page shows a live dump of the guest's navigator.userAgent /
navigator.userAgentData under the webview.
Stop rewriting the preview User-Agent in
apps/desktop/src/preview/BrowserSession.ts (the getUserAgent() →
replace(/Electron\/…/) → setUserAgent() block). The native Electron UA
passes Turnstile; every rewritten variant tested does not.