From b627f12b38cb8b4d64f60eaa1017b3e647367647 Mon Sep 17 00:00:00 2001
From: onatozmenn <onatozmen44@gmail.com>
Date: Sat, 1 Aug 2026 14:24:32 +0300
Subject: [PATCH] fix(responses): initialize completed_response on the MCP
 gateway streaming iterator

MCPEnhancedStreamingIterator has the same shape as the completion bridge that
#35413 fixed: it subclasses BaseResponsesAPIStreamingIterator, never calls
super().__init__, and never sets completed_response. It only catches
StopAsyncIteration around the base iterator, so a MidStreamFallbackError
propagates out of __anext__ and the Router reads completed_response off it in
_extract_partial_responses_usage, raising AttributeError and masking the
provider error.

The iterator is reachable there: aresponses dispatches to
aresponses_api_with_mcp when the request carries litellm_proxy MCP tools, which
returns this iterator for stream=True, and the Router wraps any
BaseResponsesAPIStreamingIterator.

Initialize it to None alongside the other iterator state. With #35413 merged
this was the last BaseResponsesAPIStreamingIterator subclass missing it.

Signed-off-by: onatozmenn <onatozmen44@gmail.com>
---
 .../responses/mcp/mcp_streaming_iterator.py   |  1 +
 .../mcp/test_mcp_streaming_iterator.py        | 39 +++++++++++++++++++
 2 files changed, 40 insertions(+)

diff --git a/litellm/responses/mcp/mcp_streaming_iterator.py b/litellm/responses/mcp/mcp_streaming_iterator.py
index 186852f91c2..f9ff335d332 100644
--- a/litellm/responses/mcp/mcp_streaming_iterator.py
+++ b/litellm/responses/mcp/mcp_streaming_iterator.py
@@ -274,6 +274,7 @@ class MCPEnhancedStreamingIterator(BaseResponsesAPIStreamingIterator):
         # Streaming state management
         self.phase = "initial_response"  # initial_response -> mcp_discovery -> (continue_initial_response <-> tool_execution) -> finished
         self.finished = False
+        self.completed_response: Any | None = None
 
         # Event queues and generation flags
         self.mcp_discovery_events: list[ResponsesAPIStreamingResponse] = (
diff --git a/tests/test_litellm/responses/mcp/test_mcp_streaming_iterator.py b/tests/test_litellm/responses/mcp/test_mcp_streaming_iterator.py
index 24edf12fffe..3afaf2d52be 100644
--- a/tests/test_litellm/responses/mcp/test_mcp_streaming_iterator.py
+++ b/tests/test_litellm/responses/mcp/test_mcp_streaming_iterator.py
@@ -257,3 +257,42 @@ async def test_initial_call_failure_is_stashed_for_eager_reraise(monkeypatch):
 
     assert iterator._initial_creation_error is not None
     assert "initial boom" in str(iterator._initial_creation_error)
+
+
+@pytest.mark.asyncio
+async def test_mid_stream_error_leaves_completed_response_readable(monkeypatch):
+    """
+    Regression test: on a mid-stream provider error the Router's fallback path
+    reads `completed_response` off this iterator directly (see
+    Router._extract_partial_responses_usage). This iterator skips
+    super().__init__(), so that read used to raise AttributeError and mask the
+    provider error, stopping the fallback from running.
+    """
+    from litellm import Router
+    from litellm.exceptions import MidStreamFallbackError
+
+    _mock_mcp_environment(monkeypatch)
+
+    class _ErroringStream:
+        def __aiter__(self):
+            return self
+
+        async def __anext__(self):
+            raise MidStreamFallbackError(
+                message="provider overloaded",
+                model="gpt-4",
+                llm_provider="openai",
+            )
+
+    iterator = MCPEnhancedStreamingIterator(
+        base_iterator=_ErroringStream(),
+        mcp_events=[],
+        tool_server_map={},
+        original_request_params={"model": "gpt-4"},
+    )
+
+    with pytest.raises(MidStreamFallbackError):
+        await iterator.__anext__()
+
+    assert iterator.completed_response is None
+    assert Router._extract_partial_responses_usage(iterator) is None