From d181176fb7220d42ff38106bc63947e1710e889b Mon Sep 17 00:00:00 2001
From: Thibault Serot <thibault@linktr.ee>
Date: Thu, 9 Jul 2026 15:09:21 +1000
Subject: [PATCH] fix(responses): monotonic sequence number for terminal error
 event; drop orphaned tool events on batch failure

Review feedback (Greptile on #32579):

- The terminal error event was numbered sequence_number=1, out of order
  after tool-execution events. __anext__ now tracks the highest
  sequence_number that passed through the stream and the error event is
  numbered after it.
- A batch tool-execution failure queued mcp_call.in_progress events that
  never received a terminal per-item event. Those queued events are now
  dropped; the terminal error event carries the failure instead.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 .../responses/mcp/mcp_streaming_iterator.py    | 18 +++++++++++++++++-
 .../mcp/test_mcp_streaming_iterator.py         | 10 ++++++++++
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/litellm/responses/mcp/mcp_streaming_iterator.py b/litellm/responses/mcp/mcp_streaming_iterator.py
index dae7bcbd085..5759960fa13 100644
--- a/litellm/responses/mcp/mcp_streaming_iterator.py
+++ b/litellm/responses/mcp/mcp_streaming_iterator.py
@@ -313,6 +313,10 @@ class MCPEnhancedStreamingIterator(BaseResponsesAPIStreamingIterator):
         self._initial_creation_error: Optional[Exception] = None
         self._stream_error: Optional[Exception] = None
         self._error_event_emitted = False
+        # Highest sequence_number emitted so far; the terminal `error` event
+        # must be numbered after it to keep the stream monotonic for strict
+        # clients.
+        self._last_sequence_number = 0
 
     def _extract_mcp_headers_from_params(self) -> None:
         """Extract MCP headers from original request params to pass to tool calls"""
@@ -386,7 +390,7 @@ class MCPEnhancedStreamingIterator(BaseResponsesAPIStreamingIterator):
         status_code = getattr(err, "status_code", None)
         return ErrorEvent(
             type=ResponsesAPIStreamEvents.ERROR,
-            sequence_number=1,
+            sequence_number=self._last_sequence_number + 1,
             error=ErrorEventError(
                 type="mcp_gateway_error",
                 code=str(status_code) if status_code is not None else "internal_error",
@@ -399,6 +403,13 @@ class MCPEnhancedStreamingIterator(BaseResponsesAPIStreamingIterator):
         return self
 
     async def __anext__(self) -> ResponsesAPIStreamingResponse:
+        chunk = await self._anext_impl()
+        sequence_number = getattr(chunk, "sequence_number", None)
+        if isinstance(sequence_number, int) and sequence_number > self._last_sequence_number:
+            self._last_sequence_number = sequence_number
+        return chunk
+
+    async def _anext_impl(self) -> ResponsesAPIStreamingResponse:
         """
         Phase-based streaming:
         1. initial_response - Stream the first LLM response (includes response.created, response.in_progress, response.output_item.added)
@@ -733,6 +744,11 @@ class MCPEnhancedStreamingIterator(BaseResponsesAPIStreamingIterator):
 
             traceback.print_exc()
             self.tool_results = []
+            # Drop the queued per-tool events: emitting mcp_call.in_progress
+            # items that never receive a completed/failed terminal event is a
+            # protocol deviation. The terminal `error` event carries the
+            # failure instead.
+            self.tool_execution_events = []
             # Remember the failure. Without this, the follow-up call is made
             # with function_call items but no function_call_output items and
             # the provider rejects it with "No tool output found for function
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 d6cefaed1b6..b1c4b2d2844 100644
--- a/tests/test_litellm/responses/mcp/test_mcp_streaming_iterator.py
+++ b/tests/test_litellm/responses/mcp/test_mcp_streaming_iterator.py
@@ -229,6 +229,9 @@ async def test_tool_execution_failure_emits_error_event_and_skips_follow_up(monk
     assert "mcp server exploded" in error_events[0].error.message
     # The doomed follow-up call was never made.
     aresponses_mock.assert_not_called()
+    # No orphaned per-tool events: a batch failure must not emit
+    # mcp_call.in_progress items that never receive a terminal event.
+    assert all(getattr(c, "type", None) != ResponsesAPIStreamEvents.MCP_CALL_IN_PROGRESS for c in chunks)
 
 
 @pytest.mark.asyncio
@@ -257,6 +260,13 @@ async def test_follow_up_failure_emits_error_event(monkeypatch):
     assert "No tool output found" in error_events[0].error.message
     # Tool-execution events were still streamed before the error surfaced.
     assert any(getattr(c, "type", None) == ResponsesAPIStreamEvents.MCP_CALL_COMPLETED for c in chunks)
+    # The terminal error event keeps sequence numbers monotonic for strict clients.
+    prior_sequence_numbers = [
+        c.sequence_number
+        for c in chunks
+        if isinstance(getattr(c, "sequence_number", None), int) and c is not error_events[0]
+    ]
+    assert error_events[0].sequence_number > max(prior_sequence_numbers)
 
 
 @pytest.mark.asyncio