You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Changed
Gemini 3 models are no longer quietly detuned. Google strongly recommends leaving Gemini 3 models at their default temperature of 1.0 — lower values can cause looping or degraded reasoning — but every feature here sent its own default of 0.2, and since the Gemini defaults moved to gemini-3.5-flash-lite (v3.33.1) that meant every out-of-the-box Gemini install was running detuned on every turn, with nothing in the log to say so. When a Gemini 3-family model is selected and the feature's temperature is still the recommended default, 1.0 is now sent instead and top_p is left to the API default. A temperature you set yourself is still sent exactly as configured, with a warning at setup pointing at Google's guidance. Gemini 2.5 and earlier models — and non-Gemini models serving the same feature, fallback chains included — keep the configured values.