test(router): add regression test for FallbackStreamWrapper _hidden_params preservation

tests/test_litellm/test_router.py

@@ -1297,6 +1297,61 @@ async def test_acompletion_streaming_iterator_edge_cases():
     print("✓ Edge case tests passed!")
 
 
+@pytest.mark.asyncio
+async def test_acompletion_streaming_iterator_preserves_hidden_params():
+    """
+    Regression test: FallbackStreamWrapper must copy _hidden_params from the
+    original CustomStreamWrapper so that x-litellm-overhead-duration-ms (and
+    other hidden params) are present in the proxy response headers for streaming.
+    """
+    from unittest.mock import MagicMock
+
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "gpt-4",
+                "litellm_params": {"model": "gpt-4", "api_key": "fake-key"},
+            }
+        ],
+    )
+
+    # Simulate a CustomStreamWrapper that already has timing metadata set by
+    # update_response_metadata (litellm_overhead_time_ms, _response_ms, etc.)
+    mock_response = MagicMock()
+    mock_response.model = "gpt-4"
+    mock_response.custom_llm_provider = "openai"
+    mock_response.logging_obj = MagicMock()
+    mock_response._hidden_params = {
+        "litellm_overhead_time_ms": 12.34,
+        "_response_ms": 500.0,
+        "litellm_call_id": "test-call-id",
+        "api_base": "https://api.openai.com",
+        "additional_headers": {},
+    }
+
+    # Make the mock iterable (yields nothing — we only care about hidden_params)
+    async def _empty():
+        return
+        yield  # make it an async generator
+
+    mock_response.__aiter__ = lambda self: _empty().__aiter__()
+
+    result = await router._acompletion_streaming_iterator(
+        model_response=mock_response,
+        messages=[{"role": "user", "content": "hi"}],
+        initial_kwargs={"model": "gpt-4", "stream": True},
+    )
+
+    # The returned FallbackStreamWrapper must carry the original _hidden_params
+    assert hasattr(result, "_hidden_params"), "result must have _hidden_params"
+    assert result._hidden_params.get("litellm_overhead_time_ms") == 12.34, (
+        "litellm_overhead_time_ms must be preserved — "
+        "this is what drives x-litellm-overhead-duration-ms in streaming responses"
+    )
+    assert result._hidden_params.get("litellm_call_id") == "test-call-id"
+    assert result._hidden_params.get("_response_ms") == 500.0
+
+
 @pytest.mark.asyncio
 async def test_async_function_with_fallbacks_common_utils():
     """Test the async_function_with_fallbacks_common_utils method"""
@@ -1858,7 +1913,7 @@ def test_get_deployment_credentials_with_provider_resolves_credential_name():
     litellm_credential_name to actual credential values (for UI-created models).
     """
     from litellm.types.utils import CredentialItem
-    
+
     # Setup credential list with a test credential
     litellm.credential_list = [
         CredentialItem(