Internal IDE Communication Bottlenecks & Upstream AI Capacity Issues since new versions of Anti Gravity

Google Antigravity
1

1. Upstream AI Capacity Issues (503 Errors)

In the logs from yesterday (February 26th), there are massive chains of errors that look like this: UNAVAILABLE (code 503): No capacity available for model gemini-3.1-pro-high on the server

What this means: The Google DeepMind servers hosting the AI models were severely overloaded. Because the IDE heavily relies on this connection to generate plans and context, it kept retrying in a continuous loop every few seconds. This aggressive retrying locks up network resources and eventually causes the connection to drop entirely (UnifiedStateSyncClient disconnected).

2. Internal IDE Communication Bottlenecks

In other log files, I see endless streams of this error happening every single minute: SendActionToChatPanel (unknown): can't send message because channel is full

What this means: The background process of Antigravity (the “daemon”) is trying to communicate with your visual IDE window (the chat panel or UI), but the internal channel is completely blocked or flooded. This is a classic software bug in this specific version (v1.19.4)—likely a memory leak or a blocked thread. When the chat interface can’t receive updates, the entire state goes out of sync, forcing the IDE to “reset” or crash to recover.

Summary

The problems from the last 2-3 days are a combination of external server outages (the model being unreachable) interacting poorly with an internal bug in version 1.19.4 (which fails to handle these disconnects gracefully and clogs up the internal chat channels).

Title: IDE crashes/resets (v1.19.4) due to unhandled 503 capacity errors causing channel overflow

Description: Since updating to version 1.19.4 of the Antigravity Desktop app, the IDE frequently resets, freezes, or crashes entirely.

Upon inspecting the internal daemon logs in ~/.gemini/antigravity/daemon/, I found that the crashes are triggered by upstream AI server capacity issues that are not being handled gracefully by the local daemon.

Steps to Reproduce (Inferred):

  1. Use the Antigravity IDE (v1.19.4) during periods of high upstream load or unavailability.

  2. The IDE daemon hits an API limit/unavailable state: UNAVAILABLE (code 503): No capacity available for model gemini-3.1-pro-high on the server.

  3. The daemon aggressively retries handling this state.

  4. Shortly after, the internal messaging channel between the daemon and the UI gets completely blocked, flooding the logs with: SendActionToChatPanel (unknown): can't send message because channel is full.

  5. Eventually, the synchronization service drops entirely: UnifiedStateSyncClient disconnected, causing the IDE to crash or forcefully reset the active session.

Expected Behavior: The IDE should gracefully handle 503 HTTP unavailable errors. If the model server is temporarily down, the IDE should back off and pause operations, rather than flooding the internal Go channels and crashing the UnifiedStateSyncClient.

Logs / Evidence: Included below are snippet patterns seen consistently across multiple log files before a crash:

E0226 17:51:24.825629 18464 log.go:380] UNAVAILABLE (code 503): No capacity available for model gemini-3.1-pro-high on the server
E0226 14:14:26.521164 38980 interceptor.go:74] /exa.language_server_pb.LanguageServerService/SendActionToChatPanel (unknown): can't send message because channel is full
I0226 18:04:08.046003 27812 client.go:416] UnifiedStateSyncClient disconnected.

E0226 17:51:24.825629 18464 log.go:380] UNAVAILABLE (code 503): No capacity available for model gemini-3.1-pro-high on the server

E0226 14:14:26.521164 38980 interceptor.go:74] /exa.language_server_pb.LanguageServerService/SendActionToChatPanel (unknown): can’t send message because channel is full

I0226 18:04:08.046003 27812 client.go:416] UnifiedStateSyncClient disconnected.

Environment:

  • OS: Windows

  • App Version: Antigravity Desktop v1.19.4

1 Like