From 6d2730d117d9664942b2045cd325fe833281752a Mon Sep 17 00:00:00 2001
From: ZedingZhang <me@zedingzhang.com>
Date: Fri, 23 Oct 2026 11:50:51 +0800
Subject: [PATCH] fix(openai): preserve Responses stream error metadata

---
 .../agent_framework_openai/_chat_client.py    | 24 ++++++++---
 .../tests/openai/test_openai_chat_client.py   | 43 +++++++++++++++++++
 2 files changed, 61 insertions(+), 6 deletions(-)

diff --git a/python/packages/openai/agent_framework_openai/_chat_client.py b/python/packages/openai/agent_framework_openai/_chat_client.py
index a5c18ebda85..17cbfd4d7d6 100644
--- a/python/packages/openai/agent_framework_openai/_chat_client.py
+++ b/python/packages/openai/agent_framework_openai/_chat_client.py
@@ -87,6 +87,7 @@
 from openai.types.responses import (
     FunctionShellToolParam,
     ResponseCustomToolCall,
+    ResponseErrorEvent,
     ResponseFunctionShellToolCall,
     ResponseFunctionShellToolCallOutput,
     ResponseFunctionToolCallOutputItem,
@@ -163,6 +164,22 @@ class _PromptCacheOptions(TypedDict, total=False):
 _TERMINAL_RESPONSE_STATUSES: Final[frozenset[str]] = frozenset({"completed", "failed", "incomplete", "cancelled"})
 
 
+class _OpenAIResponseStreamError(RuntimeError):
+    """Retain the provider metadata from a flat Responses API error event."""
+
+    def __init__(self, event: ResponseErrorEvent) -> None:
+        self.event = event
+        self.code = event.code
+        self.message = event.message
+        self.param = event.param
+        error_details = event.message
+        if event.code:
+            error_details = f"{event.code}: {error_details}"
+        if event.param:
+            error_details = f"{error_details} (parameter: {event.param})"
+        super().__init__(f"OpenAI streaming error: {error_details}")
+
+
 def _is_refusal_text_content(content: Content) -> bool:
     return (
         content.type == "text"
@@ -3936,12 +3953,7 @@ def parse_annotations(
             # ResponseCustomToolCallInputDeltaEvent,
             # ResponseCustomToolCallInputDoneEvent,
             case "error":
-                error_details = event.message
-                if event.code:
-                    error_details = f"{event.code}: {error_details}"
-                if event.param:
-                    error_details = f"{error_details} (parameter: {event.param})"
-                self._handle_request_error(RuntimeError(f"OpenAI streaming error: {error_details}"))
+                self._handle_request_error(_OpenAIResponseStreamError(event))
             case "response.content_part.added":
                 event_part = event.part
                 match event_part.type:
diff --git a/python/packages/openai/tests/openai/test_openai_chat_client.py b/python/packages/openai/tests/openai/test_openai_chat_client.py
index 15cadc9d9b9..419d7b89634 100644
--- a/python/packages/openai/tests/openai/test_openai_chat_client.py
+++ b/python/packages/openai/tests/openai/test_openai_chat_client.py
@@ -9750,6 +9750,49 @@ def test_streaming_error_event_raises_chat_client_exception() -> None:
     assert "OpenAI streaming error: server_error: The response stream failed. (parameter: input)" in str(exc_info.value)
 
 
+@pytest.mark.parametrize("request_path", ["create", "structured", "resume"])
+@pytest.mark.parametrize("code", ["server_error", "rate_limit_exceeded", None])
+@pytest.mark.parametrize("error_param", ["input", None])
+async def test_streaming_error_event_preserves_provider_details(
+    request_path: str, code: str | None, error_param: str | None
+) -> None:
+    client = OpenAIChatClient(model="test-model", api_key="test-key")
+    error_event = ResponseErrorEvent(
+        code=code,
+        message="The response stream failed.",
+        param=error_param,
+        sequence_number=2,
+        type="error",
+    )
+    events = _FakeAsyncEventStream([error_event])
+    options: OpenAIChatOptions[Any] = {}
+    if request_path == "structured":
+        options["response_format"] = OutputStruct
+        request_patch = patch.object(client.client.responses, "stream", return_value=events)
+    else:
+        method = "retrieve" if request_path == "resume" else "create"
+        request_patch = patch.object(client.client.responses.with_raw_response, method, return_value=events)
+        if request_path == "resume":
+            options["continuation_token"] = {"response_id": "resp_stream_error"}
+
+    with request_patch, pytest.raises(ChatClientException) as exc_info:
+        stream = client._inner_get_response(
+            messages=[Message(role="user", contents=["test"])], stream=True, options=options
+        )
+        assert isinstance(stream, ResponseStream)
+        async for _ in stream:
+            pytest.fail("An error event must not yield a response update.")
+
+    cause = exc_info.value.__cause__
+    assert isinstance(cause, RuntimeError)
+    assert getattr(cause, "code", None) == code
+    assert getattr(cause, "message", None) == error_event.message
+    assert getattr(cause, "param", None) == error_param
+    assert getattr(cause, "event", None) is error_event
+    if request_path == "resume":
+        assert "continuation_token" not in options
+
+
 async def test_resumed_stream_error_clears_continuation_token() -> None:
     """A failed resumed stream should clear its caller-owned continuation token."""
     client = OpenAIChatClient(model="test-model", api_key="test-key")