Recently, it seems that any type of CLI prompt from the agent is rejected, even simple NPM runs for test cases and isolating scripts. This is effectively crippling my development workflow since the AI is essentially flying blind beyond one or two small changes.
Even if I whitelist the commands or explicitly put them in a task, workflow, Markdown, or plain text file, it does not work. This applies to any model used within the Gemini chat.
I downloaded the previous version of Antigravity 1.21.6, which I knew was working just fine earlier this week, but no luck.
I could manually enter them, but it would not only exhaust all my tokens due to the back and forth with the agent sustaining context; to get any meaningful work done, I’d be out of tokens fast.
Reading through recent posts, it seems there have been a lot of downgrades and MAJOR scale-backs in the capabilities of each tier (personally, I’m on Pro) im not sure if this is partly to blame.
It’s also worth mentioning I did download Claude recently and added it to antigravity. I dont see why/how this could impact Antigravity’s native ability to run CLI but funny enough Claude CLI and the chatbot feature both can run terminal commands but are not as ‘smart’ I use to be able to plan sprints in 3-5 tasks with 1-2 large feature edits. now i have to dumb it down to 1-3 tasks with a lot of hand holding (@’s and direct references to the explicit function or file) for it to be effective.
For me this is especially discourging but a lesson learned, I was attempting a Google heavy stack ( Firebase, Go, antigravity IDE, and gemini AI) but with this huge bug, MASSIVE downgrade, and some ancedotes( a lot of devs seem to say Google will leave you high and dry) along the way I think it’s best I tout Google less and probably learn more of the Opensource tooling.