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
{{ message }}
Repository navigation
[BUG] Temperature handling gaps: non-streaming path ignores modelTemperature, other providers still default to 0 #454
PR #247 correctly fixes the OpenAI-Compatible provider's streaming path to omit temperature when no custom value is set (instead of forcing temperature: 0). However, several related gaps remain across the codebase.
The non-streaming createMessage branch builds requestOptions without any temperature field. A user who sets modelTemperature: 0.9 and disables streaming will silently get the server's default instead of their chosen value. The streaming path (fixed in #247) correctly handles this — the non-streaming path should mirror that logic.
2. Other providers still default to temperature: 0
The same ?? 0 pattern that #247 fixed in OpenAiHandler exists in several other providers:
These providers will override the server's default temperature with 0 when the user hasn't set a custom value — the same bug that #247 fixed for the OpenAI-Compatible streaming path.
3. getModel() dead defaultTemperature: 0 (cleanup)
In src/api/providers/openai.ts (line ~298), getModel() calls getModelParams({ ..., defaultTemperature: 0 }). The returned temperature field is never used in createMessage (only info and reasoning are destructured). After #247's fix, defaultTemperature: 0 is misleading since the streaming path no longer defaults to 0. This should either be removed or the temperature computation should be centralized through getModelParams.
Expected behavior
All code paths (streaming and non-streaming) should respect modelTemperature when set
When modelTemperature is unset, temperature should be omitted (not forced to 0) so the server's default applies
This behavior should be consistent across all providers
Activity
ainthek commentedon Jun 11, 2026
+1 please fix for LiteLLM, it is trivial change and takes more than week ?