From 7dd09daa988223ee7b8046da761e8943b971ca5f Mon Sep 17 00:00:00 2001
From: joereyna <joseph.reyna@gmail.com>
Date: Fri, 10 Apr 2026 10:59:27 -0700
Subject: [PATCH] route OpenAI responses WebSocket through managed HTTP path

---
 .../llms/openai/responses/transformation.py   | 11 ++++--
 .../test_responses_websocket_all_providers.py | 34 ++++++++++++-------
 2 files changed, 31 insertions(+), 14 deletions(-)

diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py
index cafb745862d..3631244495e 100644
--- a/litellm/llms/openai/responses/transformation.py
+++ b/litellm/llms/openai/responses/transformation.py
@@ -348,8 +348,15 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
         return False
 
     def supports_native_websocket(self) -> bool:
-        """OpenAI supports native WebSocket for Responses API"""
-        return True
+        """Use managed HTTP-streaming path for Responses API WebSocket mode.
+
+        OpenAI's wss://api.openai.com/v1/responses endpoint exists but does not
+        reliably stream events back in response to response.create messages in all
+        deployment contexts. The ManagedResponsesWebSocketHandler makes a normal
+        HTTP streaming call to /v1/responses and forwards the events over the
+        WebSocket, which is proven to work.
+        """
+        return False
 
     #########################################################
     ########## DELETE RESPONSE API TRANSFORMATION ##############
diff --git a/tests/test_litellm/responses/test_responses_websocket_all_providers.py b/tests/test_litellm/responses/test_responses_websocket_all_providers.py
index efec841cc4d..1f98f43266d 100644
--- a/tests/test_litellm/responses/test_responses_websocket_all_providers.py
+++ b/tests/test_litellm/responses/test_responses_websocket_all_providers.py
@@ -38,19 +38,19 @@ from litellm.llms.xai.responses.transformation import XAIResponsesAPIConfig
 class TestResponsesAPIWebSocketSupport:
     """Test that all providers have websocket support configured correctly"""
 
-    def test_openai_supports_native_websocket(self):
-        """OpenAI should support native websocket"""
+    def test_openai_uses_managed_websocket(self):
+        """OpenAI uses managed WebSocket handler (HTTP streaming path)"""
         config = OpenAIResponsesAPIConfig()
         assert (
-            config.supports_native_websocket() is True
-        ), "OpenAI should support native websocket"
+            config.supports_native_websocket() is False
+        ), "OpenAI should use managed websocket handler"
 
-    def test_azure_supports_native_websocket(self):
-        """Azure should support native websocket (inherits from OpenAI)"""
+    def test_azure_uses_managed_websocket(self):
+        """Azure uses managed WebSocket handler (inherits from OpenAI)"""
         config = AzureOpenAIResponsesAPIConfig()
         assert (
-            config.supports_native_websocket() is True
-        ), "Azure should support native websocket"
+            config.supports_native_websocket() is False
+        ), "Azure should use managed websocket handler"
 
     def test_xai_uses_managed_websocket(self):
         """XAI should use managed websocket handler"""
@@ -1000,7 +1000,9 @@ class TestNativeWebSocketUrlConstruction:
 
         mock_config = MagicMock(spec=OpenAIResponsesAPIConfig)
         mock_config.supports_native_websocket.return_value = True
-        mock_config.get_complete_url.return_value = "https://api.openai.com/v1/responses"
+        mock_config.get_complete_url.return_value = (
+            "https://api.openai.com/v1/responses"
+        )
         mock_config.validate_environment.return_value = {}
 
         mock_logging = MagicMock()
@@ -1024,8 +1026,11 @@ class TestNativeWebSocketUrlConstruction:
 
         assert len(captured_urls) == 1
         from urllib.parse import parse_qs, urlparse
+
         qs = parse_qs(urlparse(captured_urls[0]).query)
-        assert qs.get("model") == ["gpt-4o-mini"], f"Expected model in URL, got: {captured_urls[0]}"
+        assert qs.get("model") == [
+            "gpt-4o-mini"
+        ], f"Expected model in URL, got: {captured_urls[0]}"
 
     @pytest.mark.asyncio
     async def test_ws_url_preserves_existing_params_and_adds_model(self):
@@ -1071,6 +1076,11 @@ class TestNativeWebSocketUrlConstruction:
 
         assert len(captured_urls) == 1
         from urllib.parse import parse_qs, urlparse
+
         qs = parse_qs(urlparse(captured_urls[0]).query)
-        assert qs.get("model") == ["gpt-4o"], f"model missing from URL: {captured_urls[0]}"
-        assert qs.get("api-version") == ["2024-05-01"], f"existing param lost: {captured_urls[0]}"
+        assert qs.get("model") == [
+            "gpt-4o"
+        ], f"model missing from URL: {captured_urls[0]}"
+        assert qs.get("api-version") == [
+            "2024-05-01"
+        ], f"existing param lost: {captured_urls[0]}"