Repeated cache_read_tokens=0 cache resets in Antigravity 2.0 burning quota / AI Credits

Environment

  • Client: Antigravity 2.0 Desktop (Windows)
  • Timeframe: 2026-11-18 01:51:41Z – 03:18:48Z (09:51–11:18 UTC+8)
  • Model / Tier: [请补充您当时使用的模型档位,例如 Gemini 4 Carbon (High)]

Symptom
Across 11 sessions (588 model turns total) in ~87 minutes, local transcript logs show 34 follow-up turns where cache_read_tokens suddenly dropped to 0, causing full-context uncached token recalculation (14,665,266 uncached input tokens vs 128,502,270 cache read tokens) and rapidly exhausting my quota / AI Credits.

Concrete Example (Session 769bc708-a5f2-4afa-84c4-c6934a7ce97d)
Even with <60 seconds between consecutive turns, cache reads repeatedly reset to 0:

  • Turn 4 (02:29:40Z): input=25,867, cache_read=81,474
  • Turn 5 (02:30:48Z): input=143,731, cache_read=0
  • Turn 6 (02:31:39Z): input=28,712, cache_read=143,732
  • Turn 7 (02:32:37Z): input=221,757, cache_read=0
  • Turn 8 (02:34:49Z): input=249,140, cache_read=0
  • Turn 9 (02:36:04Z): input=313,028, cache_read=0
  • Turn 10 (02:37:21Z): input=372,171, cache_read=0
  • Turn 13 (02:40:08Z): input=453,876, cache_read=0

Please investigate why prompt caching is dropping to 0 on rapid consecutive turns and how this impacts quota/AI Credits deduction. (I have also submitted in-app feedback with full telemetry logs.)