fix(bedrock): preserve native converse stop reason

litellm/llms/bedrock/chat/converse_transformation.py

@@ -22,7 +22,6 @@ from litellm.constants import (
 from litellm.litellm_core_utils.core_helpers import (
     filter_exceptions_from_params,
     filter_internal_params,
-    map_finish_reason,
     safe_deep_copy,
 )
 from litellm.litellm_core_utils.litellm_logging import Logging
@@ -2659,7 +2658,7 @@ class AmazonConverseConfig(BaseConfig):
 
         ## HANDLE TOOL CALLS
         _message: Final = Message(**chat_completion_message)
-        initial_finish_reason = map_finish_reason(completion_response["stopReason"])
+        initial_finish_reason = completion_response["stopReason"]
 
         # When json_mode filtered out all synthetic tool calls the response
         # is plain content, not a pending tool invocation. Fix finish_reason

litellm/llms/bedrock/chat/invoke_handler.py

@@ -619,6 +619,7 @@ class AWSEventStreamDecoder:
             tool_use: ChatCompletionToolCallChunk | None = None
             finish_reason = ""
             usage: Usage | None = None
+            raw_stop_reason: Final[object] = chunk_data.get("stopReason")
             provider_specific_fields: dict = {}
             reasoning_content: str | None = None
             thinking_blocks: list[ChatCompletionThinkingBlock | ChatCompletionRedactedThinkingBlock] | None = None
@@ -642,8 +643,8 @@ class AWSEventStreamDecoder:
                 ) = self._handle_converse_delta_event(delta_obj, content_block_index)
             elif "contentBlockIndex" in chunk_data:  # stop block, no 'start' or 'delta' object
                 tool_use = self._handle_converse_stop_event(content_block_index)
-            elif "stopReason" in chunk_data:
-                finish_reason = map_finish_reason(chunk_data.get("stopReason", "stop"))
+            elif isinstance(raw_stop_reason, str):
+                finish_reason = map_finish_reason(raw_stop_reason)
                 self._provider_reasoning_tokens = AmazonConverseConfig.thinking_tokens_from_additional_fields(
                     chunk_data.get("additionalModelResponseFields")
                 )
@@ -657,6 +658,12 @@ class AWSEventStreamDecoder:
                 self._thinking_ran = True
 
             trace: Final = chunk_data.get("trace")
+            choice_provider_specific_fields: Final[dict[str, str] | None] = (
+                {"native_finish_reason": raw_stop_reason}  # mutable-ok: response field contract requires a dict
+                if isinstance(raw_stop_reason, str) and raw_stop_reason != finish_reason
+                else None
+            )
+
             carries_message_content: Final = bool(trace) or any(
                 key in chunk_data for key in ("start", "delta", "contentBlockIndex", "stopReason")
             )
@@ -667,6 +674,7 @@ class AWSEventStreamDecoder:
                     StreamingChoices(
                         finish_reason=finish_reason,
                         index=0,  # Always 0 - Bedrock never returns multiple choices
+                        provider_specific_fields=choice_provider_specific_fields,
                         delta=Delta(
                             content=text if carries_message_content else None,
                             role="assistant" if carries_message_content else None,

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

@@ -11,6 +11,7 @@ import litellm
 from litellm import ModelResponse
 from litellm.litellm_core_utils.prompt_templates.mid_conversation_system import CONVERTED_SYSTEM_NOTE
 from litellm.llms.bedrock.chat.converse_transformation import AmazonConverseConfig
+from litellm.llms.bedrock.chat.invoke_handler import AWSEventStreamDecoder
 from litellm.types.llms.bedrock import ConverseTokenUsageBlock
 
 
@@ -4628,6 +4629,59 @@ def test_transform_response_native_structured_output():
     assert result.choices[0].finish_reason == "stop"
 
 
+@pytest.mark.parametrize("stop_reason", ("stop_sequence", "end_turn"))
+def test_transform_response_preserves_raw_bedrock_stop_reason(stop_reason: str):
+    response_json = {
+        "metrics": {"latencyMs": 1},
+        "output": {
+            "message": {
+                "role": "assistant",
+                "content": [{"text": "done"}],
+            }
+        },
+        "stopReason": stop_reason,
+        "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={},
+        litellm_params={},
+        encoding=None,
+    )
+
+    assert result.choices[0].finish_reason == "stop"
+    assert result.choices[0].provider_specific_fields == {
+        "native_finish_reason": stop_reason
+    }
+
+
+@pytest.mark.parametrize("stop_reason", ("stop_sequence", "end_turn"))
+def test_converse_chunk_parser_preserves_raw_bedrock_stop_reason(stop_reason: str):
+    result = AWSEventStreamDecoder(model="bedrock/claude").converse_chunk_parser(
+        {"stopReason": stop_reason}
+    )
+
+    assert result.choices[0].finish_reason == "stop"
+    assert result.choices[0].provider_specific_fields == {
+        "native_finish_reason": stop_reason
+    }
+
+
 def test_add_additional_properties_simple_object():
     """Object schemas without additionalProperties get it set to false."""
     schema = {