From 5072590c27aba27212be4d86c2daba03c05ea0cf Mon Sep 17 00:00:00 2001
From: Tin Chi Lo <tin@berri.ai>
Date: Fri, 24 Jul 2026 19:31:29 -0700
Subject: [PATCH] fix(anthropic-adapter): dedupe reasoning_effort wrapping to
 close sibling gap

translate_thinking_for_model duplicated the same summary/auto_summary
wrapping logic as _translate_thinking_to_openai without the
disabled-thinking guard, so it could still wrap "none" into an
{effort, summary} dict when reasoning_auto_summary is enabled (caught
by Cursor Bugbot). Extract the wrapping rule into one shared
_apply_reasoning_summary_wrapping helper used by both call sites so
this invariant can't drift apart again.
---
 .../adapters/transformation.py                | 73 ++++++++-----------
 ...erimental_pass_through_messages_handler.py | 20 +++++
 2 files changed, 52 insertions(+), 41 deletions(-)

diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py
index 3d4233e719a..d046ff1eaeb 100644
--- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py
+++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py
@@ -684,25 +684,37 @@ class LiteLLMAnthropicMessagesAdapter:
                 thinking
             )
             if reasoning_effort:
-                summary = thinking.get("summary") if isinstance(thinking, dict) else None
-                auto_summary = is_reasoning_auto_summary_enabled()
-                if summary:
-                    return {
-                        "reasoning_effort": {
-                            "effort": reasoning_effort,
-                            "summary": summary,
-                        }
-                    }
-                elif auto_summary:
-                    return {
-                        "reasoning_effort": {
-                            "effort": reasoning_effort,
-                            "summary": "detailed",
-                        }
-                    }
-                return {"reasoning_effort": reasoning_effort}
+                return {
+                    "reasoning_effort": LiteLLMAnthropicMessagesAdapter._apply_reasoning_summary_wrapping(
+                        reasoning_effort, thinking
+                    )
+                }
             return {}
 
+    @staticmethod
+    def _apply_reasoning_summary_wrapping(
+        reasoning_effort: str,
+        thinking: Dict[str, Any],
+    ) -> Any:
+        """
+        Apply the reasoning_effort/summary wrapping rules shared by every
+        thinking->reasoning_effort translation path.
+
+        Disabled thinking always stays a plain string - there's no reasoning
+        trace to summarize, and non-Claude providers (e.g. Fireworks) expect
+        reasoning_effort as a plain string, not a summary dict.
+        """
+        thinking_type = thinking.get("type") if isinstance(thinking, dict) else None
+        if thinking_type == "disabled":
+            return reasoning_effort
+
+        summary = thinking.get("summary") if isinstance(thinking, dict) else None
+        if summary:
+            return cast(Any, {"effort": reasoning_effort, "summary": summary})
+        if is_reasoning_auto_summary_enabled():
+            return cast(Any, {"effort": reasoning_effort, "summary": "detailed"})
+        return reasoning_effort
+
     def translate_anthropic_tool_choice_to_openai(
         self, tool_choice: AnthropicMessagesToolChoice
     ) -> ChatCompletionToolChoiceValues:
@@ -997,30 +1009,9 @@ class LiteLLMAnthropicMessagesAdapter:
             if isinstance(output_config, dict) and output_config.get("effort"):
                 reasoning_effort = output_config["effort"]
 
-        if thinking_type == "disabled":
-            new_kwargs["reasoning_effort"] = reasoning_effort
-            return
-
-        summary = thinking.get("summary") if isinstance(thinking, dict) else None
-        auto_summary = is_reasoning_auto_summary_enabled()
-        if summary:
-            new_kwargs["reasoning_effort"] = cast(
-                Any,
-                {
-                    "effort": reasoning_effort,
-                    "summary": summary,
-                },
-            )
-        elif auto_summary:
-            new_kwargs["reasoning_effort"] = cast(
-                Any,
-                {
-                    "effort": reasoning_effort,
-                    "summary": "detailed",
-                },
-            )
-        else:
-            new_kwargs["reasoning_effort"] = reasoning_effort
+        new_kwargs["reasoning_effort"] = self._apply_reasoning_summary_wrapping(
+            reasoning_effort, cast(Dict[str, Any], thinking)
+        )
 
     def _translate_output_format_to_openai(
         self,
diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py
index 8875a75e86f..df3db3d2c57 100644
--- a/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py
+++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py
@@ -651,6 +651,26 @@ class TestThinkingSummaryPreservation:
             "reasoning_effort": {"effort": "high", "summary": "concise"}
         }
 
+    def test_translate_thinking_for_model_disabled_stays_plain_string_when_auto_summary_enabled(self):
+        """Disabled thinking must stay a plain string even when reasoning_auto_summary is on."""
+        import litellm
+        from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
+            LiteLLMAnthropicMessagesAdapter,
+        )
+
+        original = litellm.reasoning_auto_summary
+        try:
+            litellm.reasoning_auto_summary = True
+            thinking = {"type": "disabled"}
+            result = LiteLLMAnthropicMessagesAdapter.translate_thinking_for_model(
+                thinking=thinking,
+                model="openai/gpt-5.2",
+            )
+        finally:
+            litellm.reasoning_auto_summary = original
+
+        assert result == {"reasoning_effort": "none"}
+
 
 # ---------------------------------------------------------------------------
 # Parity tests: redundant empty-text-block sanitization scan removal.