​Safety filters overly aggressive and truncating outputs for text-based RP (Discrepancy with CLI)

Google Antigravity
1

Hi team,

​I am working on some text-based AI roleplaying (RP) games. Recently, I’ve noticed that the safety filters have become incredibly aggressive. The model frequently truncates outputs or triggers safety blocks when the narrative includes slightly mature, edgy, or explicit themes, even though the context is entirely fictional and harmless.

​Previously, when I used the Gemini CLI, I didn’t experience this level of restriction. The workflow was much smoother, and I was able to generate creative fictional content without constant interruptions. The current interface seems to have a much higher false-positive rate for safety triggers compared to my past CLI experience.

​Are there any plans to expose safety setting adjustments (similar to setting thresholds like BLOCK_NONE or BLOCK_ONLY_HIGH in the API) for this environment? Alternatively, could the team look into relaxing the filter sensitivity for fictional RP contexts to reduce these false positives?

​Any advice or workarounds would be greatly appreciated. Thanks!

2

With the recent announcement that Google is cancelling the Gemini CLI, fixing this issue has become much more critical.

The development workflow on the CLI was great, and it seems counterintuitive to remove it when the experience on Antigravity is noticeably worse for text generation. The constant false-positive safety blocks disrupt the workflow entirely. If we are expected to transition fully to this environment, we really need the ability to adjust the safety filter sensitivities ASAP so it’s on par with what the CLI offered.