fix(responses): encode deployment id into streamed bridge response ids

The chat-completions -> Responses API streaming bridge minted a bare uuid for response.created / response.in_progress and only encoded provider + deployment id on response.completed, so a client that read the id off response.created and sent it back as previous_response_id lost deployment affinity and had the unknown id forwarded to the upstream provider.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/responses/litellm_completion_transformation/streaming_iterator.py

@@ -310,10 +310,22 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
             )
             self._pending_tool_events.append(item_done_event)
 
+    def _build_encoded_response_id(self) -> str:
+        """Build the provider/deployment-encoded id shared by every event of this stream."""
+        metadata: dict[str, object] = cast(dict[str, object], self.litellm_metadata or {})
+        model_info = metadata.get("model_info")
+        raw_model_id = cast(dict[str, object], model_info).get("id") if isinstance(model_info, dict) else None
+        model_id = raw_model_id if isinstance(raw_model_id, str) else None
+        return ResponsesAPIRequestUtils.build_responses_api_response_id(
+            custom_llm_provider=self.custom_llm_provider,
+            model_id=model_id,
+            response_id=f"resp_{uuid.uuid4()}",
+        )
+
     def _default_response_created_event_data(self) -> dict:
         # Use cached response ID if available, otherwise generate a new one
         if self._cached_response_id is None:
-            self._cached_response_id = f"resp_{str(uuid.uuid4())}"
+            self._cached_response_id = self._build_encoded_response_id()
 
         response_created_event_data = {
             "id": self._cached_response_id,

litellm/responses/utils.py

@@ -207,7 +207,7 @@ class ResponsesAPIRequestUtils:
         if ResponsesAPIRequestUtils._is_litellm_encoded_response_id(response_id):
             return responses_api_response
 
-        updated_id = ResponsesAPIRequestUtils._build_responses_api_response_id(
+        updated_id = ResponsesAPIRequestUtils.build_responses_api_response_id(
             model_id=model_id,
             custom_llm_provider=custom_llm_provider,
             response_id=response_id,
@@ -408,7 +408,7 @@ class ResponsesAPIRequestUtils:
         return request_input
 
     @staticmethod
-    def _build_responses_api_response_id(
+    def build_responses_api_response_id(
         custom_llm_provider: Optional[str],
         model_id: Optional[str],
         response_id: str,

tests/test_litellm/responses/litellm_completion_transformation/test_streaming_iterator_response_id.py

@@ -0,0 +1,54 @@
+"""
+Tests for the response id emitted by the chat-completions -> Responses API streaming bridge.
+
+Every streaming event has to carry the same litellm-encoded id (provider + deployment id), so a
+client that reads the id off `response.created` can send it back as `previous_response_id` and
+still get routed to the deployment that served the session.
+"""
+
+from unittest.mock import AsyncMock
+
+from litellm.responses.litellm_completion_transformation.streaming_iterator import (
+    LiteLLMCompletionStreamingIterator,
+)
+from litellm.responses.utils import ResponsesAPIRequestUtils
+from litellm.types.utils import Choices, Message, ModelResponse
+
+
+def _build_iterator() -> LiteLLMCompletionStreamingIterator:
+    return LiteLLMCompletionStreamingIterator(
+        model="test-model",
+        litellm_custom_stream_wrapper=AsyncMock(),
+        request_input="Test input",
+        responses_api_request={},
+        custom_llm_provider="anthropic",
+        litellm_metadata={"model_info": {"id": "deployment-123"}},
+    )
+
+
+def test_response_created_event_id_is_encoded_with_the_deployment_id():
+    iterator = _build_iterator()
+
+    created_event = iterator.create_response_created_event()
+
+    decoded = ResponsesAPIRequestUtils._decode_responses_api_response_id(created_event.response.id)
+    assert decoded["model_id"] == "deployment-123"
+    assert decoded["custom_llm_provider"] == "anthropic"
+
+
+def test_streaming_events_all_share_the_completed_event_id():
+    iterator = _build_iterator()
+
+    created_event = iterator.create_response_created_event()
+    in_progress_event = iterator.create_response_in_progress_event()
+    completed_event = iterator._emit_response_completed_event(
+        ModelResponse(
+            id="chatcmpl-1",
+            choices=[Choices(finish_reason="stop", index=0, message=Message(role="assistant", content="hi"))],
+            model="test-model",
+        )
+    )
+
+    assert completed_event is not None
+    assert created_event.response.id == completed_event.response.id
+    assert in_progress_event.response.id == completed_event.response.id

tests/test_litellm/responses/test_responses_utils.py

@@ -102,7 +102,7 @@ class TestResponsesAPIRequestUtils:
         original_response_id = "resp_abc123"
 
         # Use the helper method to build an encoded response ID
-        encoded_id = ResponsesAPIRequestUtils._build_responses_api_response_id(
+        encoded_id = ResponsesAPIRequestUtils.build_responses_api_response_id(
             custom_llm_provider=test_provider,
             model_id=test_model_id,
             response_id=original_response_id,