ProviderGenerationConfig

Generation parameter overrides shared by the remote LLM providers. Each provider consumes the fields it supports and ignores the rest (reasoning_effort is OpenAI-only; thinking_budget is Gemini/Claude; disable_safety_filters is Gemini-only). Sentinels mark "unset, use the provider default": -1 for the numeric fields, "" for reasoning_effort.

Field

Type

Repeated

Description

temperature

double

-1 = provider default; 0 = deterministic

top_p

double

-1 = provider default

thinking_budget

int32

-1 = provider default (thinking tokens)

reasoning_effort

string

"" = provider default (e.g. "low")

disable_safety_filters

bool

Gemini: set all harm categories to BLOCK_NONE

Member of