Bug: The server cleared a prefix of the conversation as it grew too large.

Google Antigravity
,
1

When a long-running agent conversation exceeds the maximum conversation context / trajectory step limit (~15,000+ trajectory steps), the Antigravity IDE backend enters a degraded failure state:

  1. Silent Trajectory Execution Freeze (status: CLEARED): All subsequent planner actions and tool executions are marked status: CLEARED.
  2. Infinite Tool Retry Loop (“Zombie Agent”): Because tool executions return empty/cleared results, the agent repeatedly attempts to inspect workspace state (e.g., executing git status or the previous command over and over). Each successive invocation is immediately marked status: CLEARED, locking up the session in an infinite execution loop.
  3. Catastrophic Message Loss (Silent Purge of Queued User Prompts): While the session is looping in this cleared state, any messages typed and queued by the user in the IDE chat UI are ingested by the language server, assigned status: CLEARED, and their text payloads are permanently stripped (content: None / cleared blob in SQLite steps table). The user is never alerted that their queued messages were discarded, nor are drafts preserved in the UI.
2

Hi @dedasema,

To help them look into it, could you share:

  1. Your Antigravity version?
  2. The conversation ID or an Error ID from that session?
  3. Roughly how long the conversation was when the problem started?
  4. If it is possible, could you please send the screenshot of the issue you are facing?

For now: when a conversation gets very long, start a new one and paste in a short summary of where you left off. That should stop the loop. If you’re about to send something important, keep a copy of it until the agent picks it up.