feat: add LLM_ENABLE_THINKING env var for thinking-mode LLMs

server/reflector/llm.py

@@ -206,6 +206,12 @@ def _configure_llamaindex(self):
         """Configure llamaindex Settings with OpenAILike LLM"""
         session_id = llm_session_id.get() or f"fallback-{uuid4().hex}"
 
+        extra_body: dict = {"litellm_session_id": session_id}
+        # Only send enable_thinking when explicitly set (not None/unset).
+        # Models that don't support it will ignore the param.
+        if self.settings_obj.LLM_ENABLE_THINKING is not None:
+            extra_body["enable_thinking"] = self.settings_obj.LLM_ENABLE_THINKING
+
         Settings.llm = OpenAILike(
             model=self.model_name,
             api_base=self.url,
@@ -215,7 +221,7 @@ def _configure_llamaindex(self):
             is_function_calling_model=False,
             temperature=self.temperature,
             max_tokens=self.max_tokens,
-            additional_kwargs={"extra_body": {"litellm_session_id": session_id}},
+            additional_kwargs={"extra_body": extra_body},
         )
 
     async def get_response(

server/reflector/settings.py

@@ -75,6 +75,7 @@ class Settings(BaseSettings):
     LLM_URL: str | None = None
     LLM_API_KEY: str | None = None
     LLM_CONTEXT_WINDOW: int = 16000
+    LLM_ENABLE_THINKING: bool | None = None
 
     LLM_PARSE_MAX_RETRIES: int = (
         3  # Max retries for JSON/validation errors (total attempts = retries + 1)

server/tests/test_llm_retry.py

@@ -8,6 +8,7 @@
 from workflows.errors import WorkflowRuntimeError, WorkflowTimeoutError
 
 from reflector.llm import LLM, LLMParseError, StructuredOutputWorkflow
+from reflector.settings import Settings
 from reflector.utils.retry import RetryException
 
 
@@ -26,6 +27,57 @@ def make_completion_response(text: str):
     return response
 
 
+class TestLLMEnableThinking:
+    """Test that LLM_ENABLE_THINKING setting is passed through to OpenAILike"""
+
+    def test_enable_thinking_false_passed_in_extra_body(self):
+        """enable_thinking=False should be in extra_body when LLM_ENABLE_THINKING=False"""
+        settings = Settings(
+            LLM_ENABLE_THINKING=False,
+            LLM_URL="http://fake",
+            LLM_API_KEY="fake",
+        )
+
+        with (
+            patch("reflector.llm.OpenAILike") as mock_openai,
+            patch("reflector.llm.Settings"),
+        ):
+            LLM(settings=settings)
+            extra_body = mock_openai.call_args.kwargs["additional_kwargs"]["extra_body"]
+            assert extra_body["enable_thinking"] is False
+
+    def test_enable_thinking_true_passed_in_extra_body(self):
+        """enable_thinking=True should be in extra_body when LLM_ENABLE_THINKING=True"""
+        settings = Settings(
+            LLM_ENABLE_THINKING=True,
+            LLM_URL="http://fake",
+            LLM_API_KEY="fake",
+        )
+
+        with (
+            patch("reflector.llm.OpenAILike") as mock_openai,
+            patch("reflector.llm.Settings"),
+        ):
+            LLM(settings=settings)
+            extra_body = mock_openai.call_args.kwargs["additional_kwargs"]["extra_body"]
+            assert extra_body["enable_thinking"] is True
+
+    def test_enable_thinking_none_not_in_extra_body(self):
+        """enable_thinking should not be in extra_body when LLM_ENABLE_THINKING is None (default)"""
+        settings = Settings(
+            LLM_URL="http://fake",
+            LLM_API_KEY="fake",
+        )
+
+        with (
+            patch("reflector.llm.OpenAILike") as mock_openai,
+            patch("reflector.llm.Settings"),
+        ):
+            LLM(settings=settings)
+            extra_body = mock_openai.call_args.kwargs["additional_kwargs"]["extra_body"]
+            assert "enable_thinking" not in extra_body
+
+
 class TestLLMParseErrorRecovery:
     """Test parse error recovery with Workflow feedback loop"""