Repository navigation
LiteLlm drops GenerateContentConfig.thinking_config; map it to litellm's reasoning_effort #7360
Description
google.adk.models.lite_llm.LiteLlm._get_completion_inputs builds the litellm kwargs from a fixed allow-list of GenerateContentConfig fields (temperature, max_output_tokens, top_p, top_k, seed, stop_sequences, presence_penalty, frequency_penalty). thinking_config is present in config_dict but never forwarded, so an agent that sets e.g. ThinkingConfig(thinking_level=ThinkingLevel.LOW) gets the backend's default reasoning effort when run through LiteLlm, with no warning.
Verified on 2.8.0 and 2.10.0 (src/google/adk/models/lite_llm.py, the param_mapping / for key in (...) block, around line 3024 in v2.10.0). The only thinking handling in the file is the Anthropic thinking-block parsing on the response side.
litellm already normalises this across providers: reasoning_effort (low / medium / high, translated per provider, e.g. to Gemini thinkingConfig), and thinking={"type": "enabled", "budget_tokens": N} for budget-style providers.
Proposal: in _get_completion_inputs, map
thinking_config.thinking_level→reasoning_effortthinking_config.thinking_budget→thinking={"type": "enabled", "budget_tokens": ...}thinking_config.include_thoughts→ the provider's equivalent where litellm exposes one
Related but different: #5712 (thinking blocks dropped when parsing responses), #5805 (thinking_config ignored in run_live).
Activity
chelsealong commentedon Sep 30, 2026
Picking this one up now — opening a PR shortly. Flagging it here so nobody duplicates the work; if someone is already on it, say so and I will drop mine.
[-]LiteLlm drops GenerateContentConfig.thinking_config; map it to litellm's reasoning_effort[/-][+]`LiteLlm` drops `GenerateContentConfig.thinking_config`; map it to litellm's `reasoning_effort`[/+]surajksharma07 commentedon Oct 1, 2026
Reproduced this on 2.10.0 @chriswithers-fuse. temperature reaches litellm but neither thinking_level=LOW nor thinking_budget=1024 does so no release fixes it yet.
Until #7361 lands a workaround worth trying: pass the litellm param straight to the model, e.g. LiteLlm(model="...", reasoning_effort="low") or thinking={"type": "enabled", "budget_tokens": N} for budget-style providers. Could you check that this actually changes the reasoning behaviour with your backend? Only checked the kwargs going into litellm not a real provider call.
@chelsealong the mapping looks right. Before review could you also decide how thinking_budget=0/-1 should be handled and add a test for when a constructor reasoning_effort and thinking_config are both set (config currently wins)?
fix: forward thinking_config to litellm reasoning params