I assume this is deliberate action, because all of these features were available in gemini-cli and giving up on them must have been intentional. But what for? To have fewer people use the new product?
-
The Gemini model since version 1.5 is famous for ignoring user instructions (even in terms of text formatting! Gemma does this 10x better). Instead of letting the user scrap the default system instructions, you only allow adding gemini.md and agents.md files? Up to 1000+ characters? Is this a joke? Even claude-code allows completely throwing out their instructions using a flag, so that the user has their own and can test them. With your instructions, the model goes even wilder, it won’t even load files, because it prefers to do something else.
-
Working on context has been completely killed. In gemini-cli, I could load 200k or 800k tokens at the start and perform several actions (even 1000!), do what I needed, and move on to something else. In antigravity, I am unable to load even a single file! Just one, a simple 10,000 lines exceeds the capabilities of Antigravity, and there are another 7 files waiting in the queue. Why do you defend yourselves so much against a dedicated folder where the context would be kept? For example, if the user wants to throw in 900k tokens, they throw them in, perform a few actions, and go work on another tool.
Really, how do you want to compete with Anthropic when they had no problem with me loading 0.5M tokens and using FABLE for 60-90 minutes. In Antigravity, the very attempt to load a file consumes the limit, in 4 minutes! You can’t compete this way when BASIC features are hard-locked or pushed towards a nightmare from the user’s perspective. And they were basic in gemini-cli. It was so convenient. If you want, you overwrite system instructions; if you want, you load 950K tokens to the brim. From a convenient tool, something was created that actually hurts.