Antigravity Login Loop – OAuth Redirect Fails (Windows Protocol Handler Issue)

I am currently unable to log into Google Antigravity due to a persistent authentication loop.

After signing in via the browser, the authentication process appears to complete successfully. However, the redirect back to the IDE using the custom protocol (antigravity://oauth-success) fails.

As a result, the IDE never receives the authentication token, and I remain stuck on the login screen.


Observed Behavior:

  • Browser login completes successfully

  • Redirect to antigravity://oauth-success is triggered

  • The Antigravity IDE does not capture the token

  • User remains unauthenticated in the IDE


Expected Behavior:

After successful browser authentication, the IDE should receive the token via the protocol handler and complete the login process.


Additional Issue:

The alternative login option “Use a GCP project instead” also becomes unresponsive/freezes, making it impossible to access the tool through any available method.


Environment Details:

Same here plus no i could see any models too.
i swithced accounts tried login with them same. authentication is successfull but not logged in ide.

Antigravity v2.19.1 (Windows): OAuth loopback callback receives the auth code, but the language server panics processing it — sign-in never completes

Related thread: Antigravity Login Loop – OAuth Redirect Fails (Windows Protocol Handler Issue)

Environment

  • Windows 10 22H2 (build 10.0.19045), ru-RU locale
  • Antigravity 2.19.1 (fresh install; also reproduced after a full state reset)
  • Browsers tested: Edge (default) and Chrome — identical behavior
  • Network: home connection; tested with and without a Tailscale exit-node — no difference

Verified flow trace

Traced via browser history (SQLite visits table) and a process-spawn watcher:

  1. IDE “Continue with Google” → the language server starts a loopback listener on a random port and opens the Google auth URL:
    redirect_uri=http://localhost:<random_port>/auth/callback, response_type=code, PKCE, client_id=10710060605..-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com
  2. Google authentication completes successfully every time.
  3. The browser commits the navigation to http://localhost:<port>/auth/callback?state=...&iss=https://accounts.google.com&code=4/0AXl... — the authorization code IS delivered to the local listener. Verified timestamps (local): 17:07:28 (port 58994), 17:09:25 (59321), 17:21:12 (54522). Chromium does not record failed navigations in history, so these committed entries mean the loopback connections were accepted.
  4. At this point the language server panics and dies. The Electron app detects it (CRASH_TRIGGER_PHRASES in languageServer.ts: panic:, fatal error:, runtime: …) and restarts the LS. Crash log files are created in %USERPROFILE%\.gemini\antigravity\crashes\ at exactly these moments — but they are 0 bytes.
  5. After the restart the new LS logs Auth succeeded, refreshing features and managers — but with empty user fields (SetProjectID "", SetUserTier "", SetLocation "") and no token is ever persisted (no token files appear in ~/.gemini/antigravity/).
  6. The UI stays on the sign-in screen forever. ~12 attempts across several hours, 100% reproducible, including on a completely fresh state.

Ruled out (all tested)

  • Windows protocol handler / deep link plumbing (antigravity://): verified end-to-end — launching the handler with a synthetic URL delivers the full URL including query params into argv (Launched with deep link: antigravity://oauth-success/?code=... in main.log).
  • Firewall/ports: callback ports are free and not inside Hyper-V excluded port ranges; loopback connections commit (see above).
  • System proxy and HTTP(S)_PROXY env vars: not set.
  • Third-party cookies: allowed; [*.]google.com and [*.]google.dev explicitly allowed in Edge.
  • Tailscale exit-node on/off: no difference.
  • Full state reset: %APPDATA%\Antigravity, %LOCALAPPDATA%\antigravity, %USERPROFILE%\.gemini all removed and regenerated — crash still reproduces.
  • Run as administrator: same crash.
  • Windows Event Log / WER: no entries (the LS crash handler swallows the panic; crash log files are 0 bytes).

Extra observations

  • The panic text goes to the LS stderr, which the Electron app captures (MAX_STDERR_BUFFER = 100000 in languageServer.ts) but never logs anywhere — so the actual panic message cannot be obtained from the outside. A debug flag (or logging the captured stderr to main.log when a CRASH_TRIGGER_PHRASES match occurs) would make this class of bug actually reportable.
  • Periodic calls to daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist keep returning 401 — the endpoint is reachable, only auth is broken.
  • Unrelated noise: WSL detection errors in main.log (wsl.exe --list --verbose failed) appear on every startup.

Request

A fix, or a workaround (e.g., a manual code-entry fallback like the old loopback page). Also please consider logging the captured LS stderr to main.log when a CRASH_TRIGGER_PHRASES match occurs, so future reports can include the actual panic message.