From 4c0495f76a79d965c39bad11879666b03b3bc48e Mon Sep 17 00:00:00 2001
From: chelsealong <chelsealong@126.com>
Date: Wed, 30 Sep 2026 13:03:01 +0000
Subject: [PATCH] fix(litellm): forward thinking_config to litellm reasoning
 params

---
 src/google/adk/models/lite_llm.py      | 13 +++++++++
 tests/unittests/models/test_litellm.py | 38 ++++++++++++++++++++++++++
 2 files changed, 51 insertions(+)

diff --git a/src/google/adk/models/lite_llm.py b/src/google/adk/models/lite_llm.py
index 1630c3c983..4b106d58ee 100644
--- a/src/google/adk/models/lite_llm.py
+++ b/src/google/adk/models/lite_llm.py
@@ -3039,6 +3039,19 @@ async def _get_completion_inputs(
         mapped_key = param_mapping.get(key, key)
         generation_params[mapped_key] = config_dict[key]
 
+    thinking_config = config_dict.get("thinking_config") or {}
+    thinking_level = thinking_config.get("thinking_level")
+    thinking_budget = thinking_config.get("thinking_budget")
+    if thinking_level and thinking_level != "THINKING_LEVEL_UNSPECIFIED":
+      generation_params["reasoning_effort"] = str(
+          getattr(thinking_level, "value", thinking_level)
+      ).lower()
+    elif thinking_budget and thinking_budget > 0:
+      generation_params["thinking"] = {
+          "type": "enabled",
+          "budget_tokens": thinking_budget,
+      }
+
     if not generation_params:
       generation_params = None
 
diff --git a/tests/unittests/models/test_litellm.py b/tests/unittests/models/test_litellm.py
index caf2ceeb97..b27452f463 100644
--- a/tests/unittests/models/test_litellm.py
+++ b/tests/unittests/models/test_litellm.py
@@ -6407,6 +6407,44 @@ async def test_get_completion_inputs_generation_params():
   assert "stop_sequences" not in generation_params
 
 
+@pytest.mark.asyncio
+async def test_get_completion_inputs_maps_thinking_level_to_reasoning_effort():
+  req = LlmRequest(
+      contents=[
+          types.Content(role="user", parts=[types.Part.from_text(text="hi")]),
+      ],
+      config=types.GenerateContentConfig(
+          thinking_config=types.ThinkingConfig(
+              thinking_level=types.ThinkingLevel.LOW
+          ),
+      ),
+  )
+
+  _, _, _, generation_params, _ = await _get_completion_inputs(
+      req, model="gpt-4o-mini"
+  )
+  assert generation_params == {"reasoning_effort": "low"}
+
+
+@pytest.mark.asyncio
+async def test_get_completion_inputs_maps_thinking_budget():
+  req = LlmRequest(
+      contents=[
+          types.Content(role="user", parts=[types.Part.from_text(text="hi")]),
+      ],
+      config=types.GenerateContentConfig(
+          thinking_config=types.ThinkingConfig(thinking_budget=2048),
+      ),
+  )
+
+  _, _, _, generation_params, _ = await _get_completion_inputs(
+      req, model="gpt-4o-mini"
+  )
+  assert generation_params == {
+      "thinking": {"type": "enabled", "budget_tokens": 2048}
+  }
+
+
 @pytest.mark.asyncio
 async def test_get_completion_inputs_empty_generation_params():
   # Test that generation_params is None when no generation parameters are set