I’ve been using Antigravity CLI (agy) heavily with Gemini 4 models and kept seeing sudden weekly quota drops. Instead of guessing, I pulled exact token telemetry (input_tokens, cache_read_tokens, output_tokens) from 3,060 consecutive call pairs across our Antigravity CLI transcripts (Nov 17–19, 2026).
There are two distinct caching issues that together multiply quota burn by 5×–10× in real interactive work.
Issue 1 (Bug): Mid-turn implicit cache collapse on gemini-4-carbon-high
In one uninterrupted 46-call agent turn on gemini-4-carbon-high (2026-11-19 12:21–13:01 UTC, 0 network errors and 0 retries in cli.log):
- 30 of 46 calls returned
cache_read_tokens = 0. - 9 consecutive calls at
436K–501Ktokens got0%cache hit. - Cache dropped to
0%even after 6s, 7s, and 15s gaps between sequential tool calls (e.g. Call #20: 7s gap → 0% hit on 245K; Call #26: 15s gap → 0% hit on 436K; Call #39: 6s gap → 0% hit on 59K).
API-equivalent cost of this single 40-min turn:
- Actual:
$16.69(intro$2/$10) /$33.39(standard$4/$20) —7.79Muncached input vs1.74Mcached,94Koutput. - Counterfactual (same 46 calls if cache had held):
$3.03/$6.07→ 5.5× cost blowup purely from cache misses. - Reference: Identical task run in parallel on GPT-6.2 Astra cost
$3.10.
Why this is NOT a client/prefix mutation bug:
- The entire turn ran inside a single
agybinary process with zero mid-turn intervention. - Prompt length grew strictly monotonically across every
HIT → MISStransition (only standardread_file/run_commandtool results in between; no trimming). Whenever a call did hit,cache_read_tokensequaledprev_total + 1(100% prefix match). - On the exact same
agybinary and tool setup (≤60s intra-turn gaps):gemini-4-argon-medium: 99.5% hit rate (580/583)gemini-4-hydrogen-high: 99.5% hit rate (568/571)gemini-4-neon-medium: 96.6% hit rate (569/589)gemini-4-carbon-high: 87.6% avg, collapsing to 34.8% (16/46) in this long run.
This strongly points to server-side replica/shard routing losing session stickiness on gemini-4-carbon-high.
(If a Google engineer wants to inspect server-side traces: conversation ID ca718c60-1375-49c5-b6ab-e78bc20816..., 2026-11-19 12:21–13:01 UTC, agy CLI.)
Issue 2 (Feature Request): No usable cache TTL for human-in-the-loop work
Across all 3,060 consecutive call pairs (same model, no compaction in between):
| Pause since previous call | Hit / Miss | Hit rate |
|---|---|---|
| 0–1 min | 2,447 / 48 | 98.1% |
| 1–3 min | 139 / 29 | 82.7% |
| 3–5 min | 39 / 14 | 73.6% |
| 5–8 min | 18 / 28 | 39.1% |
| 8–12 min | 9 / 21 | 30.0% |
| 12–15 min | 40 / 194 | 17.1% |
| >15 min | 0 / 34 | 0.0% |
Implicit caching works well for tight automated tool loops on Argon/Hydrogen, but if a user pauses 5–15 minutes to read the model’s analysis/proposals and think, they pay for a full cold rebuild 61%–100% of the time.
Concrete asks
- Fix session/replica stickiness on
gemini-4-carbon-highso intra-turn cache stops dropping to0%after 6–15s. - Add a guaranteed 1-hour cache TTL for interactive Antigravity sessions (this would make subscriptions effectively 5×–10× more usable for real human-in-the-loop workflows and would make me move my entire workflow to Gemini full-time).
- Make mid-turn auto-compaction threshold configurable in
agyCLI (currently built-in compaction waits until ~500K tokens).
(Charts attached from the same telemetry dataset.)