Hi there, I have a question about antigravity 2.0, particularly as it pertains to coding in Rust. I have a small crate that implements an AI for the game oort.rs. One of the things I’ve noticed with Gemini 3.5 Flash is that there appears to be a bias toward reading entire files, even fairly large files, for small requested changes.
In a recent example, I asked the AI to remove a certain class that was responsible for tracking the location of an enemy and replace instances of it with a numerical ID that would reference an enemy database. My natural expectation would be for the AI to look for instances of the given class and read a little surrounding context. However, I found that in most cases it would read entire files, even files with 1000+ lines of code. In many cases, it was reading the contents of test modules that never reference the class I was removing, because in idiomatic rust tests are defined inline.
I’m a little puzzled by this behavior. My expectation would be that if there’s a language server involved, we don’t need to read files that contain thousands or tens of thousands of tokens just to find usages of a class. But maybe my expectations are wrong? It is my recollection that Antigravity IDE pre Gemini 3.5 Flash did this less often.
There does seem to be some sort of binary called “language_server” running with my antigravity instance. However, it has flags that don’t obviously indicate that it’s being used to index my current project. Even if it is, I’m not sure how to tell whether the agent is actually using it properly or not.