Same here, Gemini 4 Neon finally made me stay on native Antigravity.
Only real annoyance left was guessing when the prompt cache goes cold and oversized reads before the first edit.
Wrote a small local Rust plugin to put a live context/cache meter in the sidecar and trim those reads.
Shared it on GitHub if useful: https://t.co/beLGfBLtsT
Feels better on fast loops, yeah.
In my own local logs though, Carbon still cools off way faster than Hydrogen, and one long background run can wipe the prefix.
Ended up building a tiny sidecar plugin to watch each model's actual cache window live instead of guessing.
Open-sourced it here: https://t.co/beLGfBLtsT
Saw the exact same cliff scanning my own local Antigravity transcripts — Carbon drops off way earlier (~90s), Neon/Argon around ~3 min, while Hydrogen stays warm much longer.
Built a small local Rust plugin (BobbyCats/agy-hyper) mostly to survive this: it measures each model's real cache drop-off from your own logs, keeps <60s test runs synchronous so background wakeups don't chill the cache, and shows a live warmth bar in the sidecar. Sharing in case it helps anyone else bleeding Ultra quota here.
@dragon_feiman@shengzheyao Saw the exact same thing in my sessions—half the wait was just watching it grep and read files round after round. Ended up writing a small open-source plugin to bundle those lookups into one step and time each turn. Try it on your repos: https://t.co/PnFzrWnQH5
@shengzheyao Glad to see you digging into local logs! I kept hitting slow lookup loops and guessing if my cache was still warm, so I made a small open-source plugin to bundle reads and show live context/cache in the side panel. Feel free to try it: https://t.co/PnFzrWnQH5