Gemini 4 Carbon/Argon 5-minute stream idle timeout with hidden retry state in UI

Hello Antigravity team,

I am experiencing long stalls with Gemini 4 Carbon in the Windows desktop app, occurring predominantly after 10:00 PM SGT (UTC+8). The conversation remains on “Working” with little or no visible progress, and the UI does not clearly distinguish normal thinking from a stalled request or an automatic retry.

During troubleshooting, the backend recorded this error:

stream idle timeout (5m0s): context deadline exceeded

Environment

  • OS: Windows x64
  • Antigravity App: 2.31.2
  • Subscription tier: Ultra
  • Primary affected model: Gemini 4 Carbon
  • Carbon variants observed: Medium and High
  • Other models tested: Gemini 4 Argon (similar stalls/timeouts), Gemini 4 Hydrogen (performs slightly better)
  • Time zone: Asia/Singapore (UTC+8)
  • Issue occurrence window: After 10:00 PM SGT (UTC+8 / after 14:00 UTC)
  • Latest troubleshooting runs: Direct connection (no VPN or proxy)

Observed behavior

  • Gemini 4 Carbon tasks (Medium and High) frequently remain on “Working” for extended periods without visible progress, consistently occurring after 10:00 PM SGT (UTC+8).
  • Backend failures and automatic retries are not clearly surfaced in the UI, making it difficult to tell whether the model is thinking, the stream has stalled, or a retry is in progress.
  • In recent runs, Carbon High’s time to first token varied substantially (approx. 52s, 83s, and 13s) before eventually completing.
  • An earlier Carbon Medium request hit a 5-minute stream idle timeout and succeeded after an automatic retry.
  • Gemini 4 Argon exhibits similar intermittent stalls and 5-minute timeouts during the same late-night window. Gemini 4 Hydrogen performs slightly better, though my tasks require the higher-tier reasoning models (Carbon/Argon).

Exact error

The recorded failure was:

request failed: stream idle timeout (5m0s): context deadline exceeded

The backend treats this as retryable, but the conversation can continue displaying only “Working.”

To distinguish the evidence accurately: the five-minute timeout was also independently reproduced during a Gemini 4 Argon High run. On November 14, 2026, at approximately 01:27:37 Singapore time (November 13, 2026, 17:27:37 UTC), the backend logged:

Run: attempt 1 failed (request failed: stream idle timeout (5m0s): context deadline exceeded), retrying in 4s

Comparison results

  • During that five-minute stall, a separate Carbon High request successfully completed through the same running backend.
  • Independent HTTPS checks to the Google API endpoints succeeded.

Questions

  1. Is there a known serving or desktop streaming issue affecting Gemini 4 Carbon and Argon?
  2. Why does the UI remain on “Working” without clearly showing the timeout and retry state?
  3. Is there a supported Windows desktop setting for the five-minute stream idle timeout, or a fix available?
  4. Are these extended stalls and timeouts—particularly after 10:00 PM SGT (14:00 UTC)—caused by backend compute capacity constraints, serving bottlenecks, or resource contention on higher-tier models (Carbon and Argon)?

This behavior is severely disrupting my daily work, as a retrying request looks identical to an actively progressing task and leaves no clear indication of what is happening. I hope this issue can be investigated and improved as soon as possible.

Thank you.

Hi @CialloChiao,

Thank you for the detailed report and backend logs.

Here is how each behavior on 2.31.2 works and what changed in our latest release:

  • 5-minute thinking timeout & "Working" state (Q1–Q3): On 2.31.2, streams cancelled after 5 minutes (5m0s) without incoming chunks and retried silently while displaying "Working". Antigravity App v2.33.0 directly fixed 5-minute long-reasoning timeouts and stale pooled connection retries—no manual timeout setting is needed once you update via antigravity.google/download and restart. We’ve also noted your feedback on showing retry states in the UI.
  • Stalls after 10:00 PM SGT / 14:00 UTC (Q4): Longer time-to-first-token during global peak hours reflects real-time serving load on Gemini 4 Carbon and Argon rather than your Ultra plan quota. Using Orchestrator (Carbon) or Claude Opus 5.5 (thinking) (Choose a model) during peak windows helps avoid queueing.

Let us know if v2.33.0 resolves the timeouts.

Thank you for all your hard work—Antigravity is genuinely getting better and better. That said, please lift the ban and restrictions on Antigravity; the entire developer forum is in shambles.