From a634cb9b62ecfe7e37de8715bf2e2cd763b2a8b2 Mon Sep 17 00:00:00 2001
From: shivam <shivam@berri.ai>
Date: Wed, 29 Jul 2026 00:30:33 +0000
Subject: [PATCH] fix(responses): initialize completed_response and
 _hidden_params on the completion bridge iterator

LiteLLMCompletionStreamingIterator subclasses ResponsesAPIStreamingIterator but never runs the base constructor, so it lacks completed_response and _hidden_params. When a provider errors before the first content chunk, the router reads completed_response directly in _extract_partial_responses_usage and raises AttributeError, and the missing _hidden_params makes the proxy pre-wrap the iterator for header attachment so the Responses mid-stream fallback path is skipped entirely. Initialize both on the bridge iterator so provider errors surface and fallbacks run.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../streaming_iterator.py                     |  7 +++-
 ...st_router_aresponses_streaming_fallback.py | 40 +++++++++++++++++++
 2 files changed, 46 insertions(+), 1 deletion(-)

diff --git a/litellm/responses/litellm_completion_transformation/streaming_iterator.py b/litellm/responses/litellm_completion_transformation/streaming_iterator.py
index cf69654d15d..2aded792f73 100644
--- a/litellm/responses/litellm_completion_transformation/streaming_iterator.py
+++ b/litellm/responses/litellm_completion_transformation/streaming_iterator.py
@@ -67,7 +67,12 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
         self.request_input: str | ResponseInputParam = request_input
         self.responses_api_request: ResponsesAPIOptionalRequestParams = responses_api_request
         self.custom_llm_provider: str | None = custom_llm_provider
-        self.litellm_metadata: dict | None = litellm_metadata or {}
+        self.litellm_metadata = litellm_metadata or {}
+        self.completed_response: Any | None = None
+        _wrapper_hidden_params = getattr(litellm_custom_stream_wrapper, "_hidden_params", None)
+        self._hidden_params: dict[str, Any] = (
+            dict(_wrapper_hidden_params) if isinstance(_wrapper_hidden_params, dict) else {}
+        )
         # Store lightweight dict snapshots for stream_chunk_builder to reduce
         # repeated Pydantic attribute access in end-of-stream assembly.
         self.collected_chat_completion_chunks: list[dict[str, Any]] = []
diff --git a/tests/router_unit_tests/test_router_aresponses_streaming_fallback.py b/tests/router_unit_tests/test_router_aresponses_streaming_fallback.py
index 2fb7bdfceb5..a1ac020e209 100644
--- a/tests/router_unit_tests/test_router_aresponses_streaming_fallback.py
+++ b/tests/router_unit_tests/test_router_aresponses_streaming_fallback.py
@@ -483,3 +483,43 @@ async def test_aresponses_client_error_event_skips_fallback():
 
     assert exc_info.value.status_code == 400
     mock_fallback.assert_not_awaited()
+
+
+# -------- regression: real bridge iterator honors the base contract (LIT-4912) --------
+
+
+def test_extract_partial_responses_usage_real_bridge_iterator_pre_first_chunk():
+    """
+    Regression for LIT-4912.
+
+    When a provider errors before the first content chunk, the completion-bridge
+    iterator reaches _extract_partial_responses_usage with no collected chunks.
+    A real LiteLLMCompletionStreamingIterator (not a MagicMock standing in for
+    it) must expose the base-class contract so usage extraction returns None
+    instead of raising AttributeError on completed_response, and must carry
+    _hidden_params so the router does not pre-wrap it for header attachment and
+    thereby skip the mid-stream fallback path entirely.
+    """
+    from litellm.responses.litellm_completion_transformation.streaming_iterator import (
+        LiteLLMCompletionStreamingIterator,
+    )
+    from litellm.responses.streaming_iterator import BaseResponsesAPIStreamingIterator
+
+    class _FakeStreamWrapper:
+        def __init__(self) -> None:
+            self.logging_obj = MagicMock()
+            self._hidden_params = {"model_id": "deployment-123"}
+
+    iterator = LiteLLMCompletionStreamingIterator(
+        model="anthropic/claude-3-5-sonnet-latest",
+        litellm_custom_stream_wrapper=_FakeStreamWrapper(),
+        request_input="hi",
+        responses_api_request={},
+    )
+
+    assert isinstance(iterator, BaseResponsesAPIStreamingIterator)
+    assert iterator.completed_response is None
+    assert iterator._hidden_params == {"model_id": "deployment-123"}
+    assert not iterator.collected_chat_completion_chunks
+
+    assert Router._extract_partial_responses_usage(iterator) is None