fix(bedrock): preserve native reason through conversions

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -464,6 +464,14 @@ class ChunkProcessor:
             None,
         )
 
+    @staticmethod
+    def _choice_provider_specific_fields(chunk: "_BaseChunk") -> Mapping[str, object]:
+        choices: Final = chunk.get("choices")
+        if not choices:
+            return MappingProxyType({})
+        fields: Final = choices[0].get("provider_specific_fields")
+        return fields if isinstance(fields, dict) else MappingProxyType({})
+
     def build_base_response(
         self,
         chunks: Sequence["_BaseChunk"],
@@ -482,14 +490,9 @@ class ChunkProcessor:
         role: Final = ChunkProcessor._get_role_from_chunks(chunks)
         finish_reason: Final = ChunkProcessor._finish_reason_of_chunks(chunks)
 
-        choice_provider_specific_fields: Final[dict[str, object]] = {  # mutable-ok: response field requires a dict
-            key: value
-            for chunk in chunks
-            if chunk.get("choices")
-            for fields in (chunk["choices"][0].get("provider_specific_fields"),)
-            if isinstance(fields, dict)
-            for key, value in fields.items()
-        }
+        choice_provider_specific_fields: Final[dict[str, object]] = dict(  # mutable-ok: response field requires a dict
+            chain.from_iterable(ChunkProcessor._choice_provider_specific_fields(chunk).items() for chunk in chunks)
+        )
 
         # Initialize the response dictionary
         response = ModelResponse(

litellm/llms/bedrock/chat/converse_transformation.py

@@ -3093,12 +3093,13 @@ class AmazonConverseConfig(BaseConfig):
 
         ## HANDLE TOOL CALLS
         _message: Final = Message(**chat_completion_message)
-        mapped_finish_reason: Final = map_finish_reason(completion_response["stopReason"])
+        raw_finish_reason: Final = completion_response["stopReason"]
+        mapped_finish_reason: Final = map_finish_reason(raw_finish_reason)
 
         initial_finish_reason: Final = (
             "stop"
             if resolved_json_mode and not filtered_tools and tools and mapped_finish_reason == "tool_calls"
-            else completion_response["stopReason"]
+            else raw_finish_reason
         )
 
         (
@@ -3109,11 +3110,17 @@ class AmazonConverseConfig(BaseConfig):
             tools=optional_params.get("tools"),
             initial_finish_reason=initial_finish_reason,
         )
+        choice_provider_specific_fields: Final = (
+            {"native_finish_reason": raw_finish_reason}  # mutable-ok: Choices requires a dict
+            if returned_finish_reason != raw_finish_reason
+            else None
+        )
         model_response.choices = [
             litellm.Choices(
                 finish_reason=returned_finish_reason,
                 index=0,
                 message=returned_message,
+                provider_specific_fields=choice_provider_specific_fields,
             )
         ]
         model_response.created = int(time.time())

tests/unit/llms/bedrock/chat/test_converse_transformation.py

@@ -6010,6 +6010,69 @@ def test_transform_response_preserves_raw_bedrock_stop_reason(stop_reason: str):
     }
 
 
+def test_transform_response_preserves_raw_stop_reason_when_content_becomes_tool_call():
+    response_json = {
+        "metrics": {"latencyMs": 1},
+        "output": {
+            "message": {
+                "role": "assistant",
+                "content": [
+                    {
+                        "text": json.dumps(
+                            {
+                                "type": "function",
+                                "name": "lookup_weather",
+                                "parameters": {"city": "Seoul"},
+                            }
+                        )
+                    }
+                ],
+            }
+        },
+        "stopReason": "stop_sequence",
+        "usage": {
+            "inputTokens": 10,
+            "outputTokens": 1,
+            "totalTokens": 11,
+        },
+    }
+    raw_response = httpx.Response(
+        200,
+        json=response_json,
+        request=httpx.Request("POST", "https://bedrock.test/converse"),
+    )
+
+    result = AmazonConverseConfig().transform_response(
+        model="bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0",
+        raw_response=raw_response,
+        model_response=ModelResponse(),
+        logging_obj=None,
+        request_data={},
+        messages=[],
+        optional_params={
+            "tools": [
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "lookup_weather",
+                        "parameters": {
+                            "type": "object",
+                            "properties": {"city": {"type": "string"}},
+                        },
+                    },
+                }
+            ]
+        },
+        litellm_params={},
+        encoding=None,
+    )
+
+    assert result.choices[0].finish_reason == "tool_calls"
+    assert result.choices[0].provider_specific_fields == {
+        "native_finish_reason": "stop_sequence"
+    }
+
+
 # AWS Bedrock Converse stopReason values, accessed 2026-09-16:
 # https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_Converse.html
 @pytest.mark.parametrize("stop_reason", ("stop_sequence", "end_turn"))