Antigravity IDE — Reproducible freeze/crash on startup (Linux)

Summary

Antigravity IDE consistently becomes unresponsive ~10–17 seconds after Google sign-in completes, every launch, with no error surfaced to the user. Fully reproducible on a completely clean install and profile.

Environment

  • Antigravity build: 1.107.0 (ideVersion 1.23.2)
  • Commit: 15487b3041e65228cae24980a3f796c905ef582c
  • OS: Ubuntu, kernel 6.8.0-136-generic, x86_64
  • Install method: official APT repo (us-central1-apt.pkg.dev, antigravity-debian)
  • GPU: Nvidia (VA-API warns “Should skip nVidia device named: nvidia-drm”)

Steps to reproduce

  1. Fresh install via APT (or fully purged + reinstalled existing install)
  2. Launch antigravity (also reproduces with --disable-gpu)
  3. Sign in with Google account
  4. Within ~10–17 seconds of sign-in completing, window becomes unresponsive (“The window is not responding” dialog)

What’s been ruled out

  • GPU/rendering: reproduces identically with --disable-gpu
  • Corrupted local profile: reproduces on a brand-new --user-data-dir profile with no prior state
  • Corrupted cache: reproduces after wiping CachedData, CachedProfilesData, globalStorage/google.antigravity
  • Stale model preference: antigravityUnifiedStateSync.modelPreferences key in state.vscdb was empty; deleting it made no difference
  • Network reachability: confirmed working (curl to update endpoint succeeds cleanly, TLS handshake fine, HTTP 200)
  • Full purge + clean reinstall: apt remove --purge, deleted ~/.config/Antigravity and ~/.antigravity entirely, reinstalled from scratch — still reproduces identically

Crash signature (from main.log, consistent across every run)

CodeWindow: detected unresponsive
CodeWindow unresponsive samples:
    at PJr (workbench.desktop.main.js:3270:49337)
    at Array.sort (<anonymous>)
    at Z_t (workbench.desktop.main.js:3270:39240)
<1>
    at WHl (workbench.desktop.main.js:3210:35784)
    at hs (workbench.desktop.main.js:3210:34638)
    at o2n.Kb (workbench.desktop.main.js:5143:84827)
    at o2n.setSelectedModel (workbench.desktop.main.js:5143:75156)
    at ... (workbench.desktop.main.js:4956:34820)
    at Etn (workbench.desktop.main.js:3270:52997)
    at Array.forEach (<anonymous>)

Freeze consistently occurs inside setSelectedModel, triggered during/after an Array.sort, right after the model list is fetched.

Relevant timeline (ls-main.log / cloudcode.log, one representative run)

12:11:32.421  Auth state: uninitialized
12:11:32.502  Auth state: validatingLogin
12:11:36.340  POST cloudcode-pa.googleapis.com/v1internal:loadCodeAssist
12:11:37.671  POST daily-cloudcode-pa.googleapis.com/v1internal:fetchAvailableModels
12:11:43.434  Auth state: signedIn
12:11:44.853  POST daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist
12:11:46.002  POST daily-cloudcode-pa.googleapis.com/v1internal:fetchAvailableModels
12:11:49.581  CodeWindow: detected unresponsive   <-- no further log lines before this

Note the pattern: first call hits production cloudcode-pa.googleapis.com, all subsequent calls hit daily-cloudcode-pa.googleapis.com (pre-release/ canary endpoint). No error, quota message, or non-2xx status is ever logged by the client — the language server logs only outgoing request traceIds, no response codes are visible client-side.

Also seen consistently at startup (does not appear related, but noting for completeness):

Failed to set Cloud Code URL on Language Server: Language server client has not been initialized!

On the one run where --disable-gpu was combined with a completely fresh --user-data-dir, the app successfully reached the “Welcome to Antigravity” onboarding screen rather than freezing — the only run so far that got past this point, though onboarding was not completed further to confirm full stability.

Request

This looks like either (a) a client-side bug in how setSelectedModel processes the response from fetchAvailableModels — particularly given the mixed production/daily-canary endpoint routing — or (b) a malformed/oversized response from the daily-cloudcode-pa.googleapis.com endpoint for this account. Would appreciate guidance on which, and whether pinning to the production endpoint only is possible as a workaround.

Indeed, I have encountered a similar issue within the Windows 11 environment (Antigravity 1.23.2). This anomaly appears to have emerged concurrently with the deployment of the gemini-3.6 update. Prompted by a previously reported issue where the agent became persistently unresponsive during execution and resisted all termination attempts, I resolved to restart the IDE. Upon doing so, I immediately encountered the exact symptoms described by the original poster, which manifest as follows:

  1. futile attempts to load the workspace and agent windows;
  2. a timeout error, accompanied by prompts suggesting closure or a restart;
  3. persistence of the issue even after restarting;
  4. identical failures when attempting to open a pristine workspace post-closure;
  5. the same outcome when attempting to access a remote workspace after termination;
  6. the failure persisting even when attempting to open the workspace while disconnected from the network.

In conclusion, an anomaly during workspace initialization prevents a successful launch; however, the system lacks both diagnostic error reports and a safe-boot mode to isolate the issue.

Starting today, the same error. I tried the same steps as the author. Nothing helped.

Antigravity Version: 1.23.2

VSCode OSS Version: 1.107.0

Commit: 15487b3041e65228cae24980a3f796c905ef582c

Date: 2026-04-16T08:28:19.366Z

Electron: 39.2.3

Chromium: 142.0.7444.175

Node.js: 22.21.1

V8: 14.2.231.21-electron.0

OS: Linux x64 6.8.0-134-generic

Language Server CL: 9005663..

Same issue here, and I think I can add some corroborating detail to V.lakshmipathiraja’s findings above.

Environment

  • Two separate Bluefin (Fedora Silverblue-based, Fedora 43) machines, plus a Windows machine — all under the same account, all reproduce identically
  • Antigravity build 1.107.0 (ideVersion 1.23.2), commit 15487b3041e65228cae24980a3f796c905ef582c

Crash signature
Identical Array.sort() stack trace in the render process:

CodeWindow: detected unresponsive
UnresponsiveSampleError: from window with ID 1 belonging to process with pid
at PJr (workbench.desktop.main.js:3270:493..)
at Array.sort ()
at Z_t (workbench.desktop.main.js:3270:392..)
Total Samples: 14 (spanning ~15 seconds)

What I’ve ruled out on my end (matches the OP’s findings)

  • Not GPU/rendering — window fully loads and renders before hanging
  • Not workspace-specific — reproduces on brand-new empty local folders
  • Not extension-related — reproduces with --disable-extensions
  • Not tied to a specific remote host — reproduces with no remote connection at all
  • Not fixed by deleting local Antigravity config/cache directories — freeze returns as soon as I sign back in, on any of my three machines

Cross-device pattern
The part I can add: since this reproduces identically across two Linux machines and Windows under one account, and persists even after full local wipes on each, this really does point to something account-linked being fetched at sign-in — consistent with the OP’s observation that fetchAvailableModels/setSelectedModel calls shift to the daily-cloudcode-pa.googleapis.com canary endpoint partway through sign-in, right before the freeze. A full wipe + fresh sign-in gets me through exactly once before the same freeze returns, which fits a corrupted/oversized response being re-synced or re-fetched from that endpoint on every subsequent launch.

I also inspected globalStorage/state.vscdb locally and found large antigravityUnifiedStateSync.* keys (trajectorySummaries ~130KB, artifactReview ~45KB), but deleting these individually did not fix the freeze for me — so I now think those are a red herring and the OP’s setSelectedModel/model-list theory is the more likely culprit, given how well the timing lines up.

Would be great to get confirmation from the Antigravity team on whether this is indeed the daily/canary model endpoint misbehaving (possibly tied to the gemini-3.6 rollout mentioned above), and whether there’s a way to pin to the production endpoint as a workaround while it’s fixed server-side.

Environment:

  • OS: Ubuntu 24.04.4 LTS (kernel 7.0.0-28-generic)
  • Antigravity: 1.23.2-1776332190 (installed via official apt repo)
  • GPU: NVIDIA GeForce RTX 4060 Laptop, driver 580.159.03, CUDA 13.0
  • Launcher already runs with --disable-gpu flag (pre-existing from an earlier
    troubleshooting attempt) — the freeze still occurs, so GPU/rendering does
    not appear to be the root cause in this case.

Reproducibility: 100% (every launch), consistent with other reports in this thread.

Exact timeline from logs (all timestamps same session, 2026-07-23):

00:57:11.808 auth.log Auth state: uninitialized
00:57:11.864 auth.log Auth state: validatingLogin
00:57:11.864 cloudcode.log [error] “Failed to set Cloud Code URL on
Language Server: Language server client has not been
initialized!”
00:57:13.098 cloudcode.log POST loadCodeAssist (traceId 658c960bc1fedc..)
00:57:13.175 cloudcode.log POST loadCodeAssist (traceId 9605c982bcfca4..)
— second call, 77ms after the first
00:57:13.620 cloudcode.log POST fetchAvailableModels (traceId 5fec6153427521..)
00:57:13.881 cloudcode.log POST onboardUser (traceId 405fa1b1a115e5..)
00:57:14.669 cloudcode.log POST loadCodeAssist (traceId 33d72e52005384..)
— third call
00:57:14.887 cloudcode.log POST fetchAvailableModels (traceId f318d3db5aaf1..)
— second call
00:57:14.890 auth.log Auth state: signedIn
00:57:14.890 auth.log AntigravityAuthMainService initialized

-– 15.1 seconds of apparent normal idle state —

00:57:29.991 main.log [error] CodeWindow: detected unresponsive
00:57:31.885 main.log [error] CodeWindow unresponsive samples:
at PJr (workbench.desktop.main.js:3270:493..)
at Array.sort ()
at Z_t (workbench.desktop.main.js:3270:392..)
Total Samples: 1

Observation (not a confirmed root cause, just flagging for whoever
investigates): loadCodeAssist is called 3 times and fetchAvailableModels
2 times within a 1.3-second window during startup. I don’t have visibility
into whether this call pattern is expected (retries/prefetch) or anomalous.
If it is unexpected, overlapping/duplicate responses arriving out of order
could plausibly produce a malformed or oversized array by the time the
Array.sort() call is reached ~15s later — but I want to be clear this is
a hypothesis based on timing correlation, not something I’ve verified in
the source.

Additional note: I have full logs (cloudcode.log, main.log, auth.log,
renderer.log, ls-main.log, exthost logs) from this session available if
useful for triage — happy to share the complete files rather than excerpts

Please use the latest IDE version. The dev are not supporting 1.x older version

https://edgedl.me.gvt1.com/edgedl/release2/j0qc3/antigravity/stable/2.1.1-6123990880747520/linux-x64/Antigravity%20IDE.tar.gz

Weird. The “latest IDE version” for me is 1.23.2. This is the latest version that’s published in Google’s APT repository for Debian/Ubuntu. The APT repo reliably kept the antigravity package updated all this time, until recently.

I think this whole debacle is quite embarrassing for AG. They should be embarrassed at this state of affairs. You mean to tell me they can’t figure out how to auto-update 1.3 to 2.0, via the APT repo? That’s ridiculous. Canonical had no issues figuring out how to automatically upgrade everyone from native Firefox debs to Firefox snaps. And that’s order of magnitude more complex than what happened with 1.3 → 2.0. Setting aside the issues with the firefox snap itself, the upgrade itself was flawless.

And before 1.23.2 … all over itself it kept nagging everyone that an update was available, click here, then once you click here … if you installed antigravity via the distribution repo, update it there". And there’s nothing to update there.

Do better, Google.

Weird. The “latest IDE version” for me is 1.23.2. This is the latest version that’s published in Google’s APT repository for Debian/Ubuntu.

Don’t use that repo. The current way the Google DeepMind team is distributing their software is via tarballs.

Distributing tarballs is so last century. That’s what someone does
when they’re new to software development. In other words: distributing
via an APT repository, like every other vendor, is just too
complicated for Google. …