From 1ad29e2a1ec5802707b5a42c68f3d24ebf71f957 Mon Sep 17 00:00:00 2001
From: maxi-maxima <240684166@qq.com>
Date: Thu, 13 Aug 2026 10:19:16 +0800
Subject: [PATCH] fix(llm): pass max reasoning effort to LiteLLM

---
 strix/core/inputs.py | 14 +++++---------
 tests/test_inputs.py | 10 +++++-----
 2 files changed, 10 insertions(+), 14 deletions(-)

diff --git a/strix/core/inputs.py b/strix/core/inputs.py
index a1106ae2d..d2437706d 100644
--- a/strix/core/inputs.py
+++ b/strix/core/inputs.py
@@ -219,7 +219,7 @@ def make_model_settings(
         and model_supports_reasoning(model_name)
     ):
         model_settings = model_settings.resolve(
-            _reasoning_settings(reasoning_effort, model_settings.extra_args),
+            _reasoning_settings(reasoning_effort),
         )
     if force_required_tool_choice and _accepts_required_tool_choice(model_name):
         model_settings = model_settings.resolve(ModelSettings(tool_choice="required"))
@@ -247,18 +247,14 @@ def _request_headers(
 
 def _reasoning_settings(
     effort: ReasoningEffort,
-    extra_args: dict[str, Any] | None,
 ) -> ModelSettings:
-    """``max`` is not in the OpenAI SDK's ``Reasoning.effort`` enum, so send it as
-    a raw body field instead — also keeping it clear of LiteLLM's DeepSeek mapping,
-    which collapses every ``reasoning_effort`` level to plain thinking-enabled.
-    Providers that don't support ``max`` reject the request.
+    """``max`` is not in the OpenAI SDK's ``Reasoning.effort`` enum, so send it
+    through ``extra_args`` for the LiteLLM model to promote to its top-level
+    ``reasoning_effort`` argument. Providers that don't support ``max`` reject it.
     """
     if effort != "max":
         return ModelSettings(reasoning=Reasoning(effort=effort))
-    return ModelSettings(
-        extra_args={**(extra_args or {}), "extra_body": {"reasoning_effort": "max"}},
-    )
+    return ModelSettings(extra_args={"reasoning_effort": "max"})
 
 
 def _prompt_cache_extra_args(model_name: str) -> dict[str, Any] | None:
diff --git a/tests/test_inputs.py b/tests/test_inputs.py
index 2ff9a6035..fac06b720 100644
--- a/tests/test_inputs.py
+++ b/tests/test_inputs.py
@@ -134,15 +134,15 @@ def test_prompt_cache_kept_for_non_bedrock_claude_even_if_unmapped(monkeypatch:
         ]
 
 
-def test_max_reasoning_effort_sent_as_raw_body_field() -> None:
-    # "max" is absent from the OpenAI SDK's Reasoning enum, and LiteLLM's DeepSeek
-    # mapping collapses every effort to thinking-enabled, so it has to ride along
-    # as a raw body field to reach the provider.
+def test_max_reasoning_effort_sent_as_litellm_argument() -> None:
+    # "max" is absent from the OpenAI SDK's Reasoning enum, so it has to use the
+    # LiteLLM model's extra_args escape hatch to reach acompletion as a top-level
+    # reasoning_effort argument.
     settings = make_model_settings(
         "max", model_name="deepseek/deepseek-v4-flash", request_timeout=30
     )
     assert settings.reasoning is None
-    assert settings.extra_args == {"timeout": 30, "extra_body": {"reasoning_effort": "max"}}
+    assert settings.extra_args == {"timeout": 30, "reasoning_effort": "max"}
 
 
 def test_conversation_tail_breakpoint_moves_with_appended_transcript() -> None: