From 91a3d4ea5eb2c5bf73f940687d4ece701104ce64 Mon Sep 17 00:00:00 2001
From: Ambuj Upadhyay <34904987+lets-order-some-fries@users.noreply.github.com>
Date: Tue, 18 Aug 2026 22:44:52 +0530
Subject: [PATCH] fix(openai): async streaming no-response branch now uses the
 computed status

Review follow-up: the helper computed 400 for a bare construction-time
OpenAIError, but async_streaming's no-response branch hard-coded 500 and
discarded it -- so with stream=True a missing API key still surfaced as a
retryable server error, the exact class of misreport this PR exists to
fix. Verified at runtime: acompletion(stream=True) with no key raised
InternalServerError(500) before, BadRequestError(400) after.

Adds an end-to-end async-streaming regression test (red without the fix,
green with it); the existing tests covered the helper only, which is why
this slipped through. Transient classes are unaffected:
_status_code_for_openai_sdk_error still returns 500 for anything without
a status_code that is not a bare OpenAIError, and the ReadTimeout->408
branch is untouched.
---
 litellm/llms/openai/openai.py                 |  8 ++++++-
 .../test_openai_sdk_error_status_code.py      | 24 +++++++++++++++++++
 2 files changed, 31 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/openai/openai.py b/litellm/llms/openai/openai.py
index 49465ade3e1..94706e19ed9 100644
--- a/litellm/llms/openai/openai.py
+++ b/litellm/llms/openai/openai.py
@@ -1202,8 +1202,14 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
                             body=exception_body,
                         )
                     else:
+                        # status_code was computed above by
+                        # _status_code_for_openai_sdk_error; hard-coding 500
+                        # here made the no-response branch discard it, so an
+                        # async streaming call with no configured API key
+                        # surfaced as a retryable 500 instead of the 400 the
+                        # non-streaming path reports.
                         raise OpenAIError(
-                            status_code=500,
+                            status_code=status_code,
                             message=f"{e}",
                             headers=error_headers,
                             body=exception_body,
diff --git a/tests/unit/llms/openai/test_openai_sdk_error_status_code.py b/tests/unit/llms/openai/test_openai_sdk_error_status_code.py
index 795dc90c93f..c0857fe6e0f 100644
--- a/tests/unit/llms/openai/test_openai_sdk_error_status_code.py
+++ b/tests/unit/llms/openai/test_openai_sdk_error_status_code.py
@@ -48,3 +48,27 @@ def test_status_code_zero_is_preserved():
         status_code = 0
 
     assert _status_code_for_openai_sdk_error(_ZeroStatus()) == 0
+
+
+@pytest.mark.asyncio
+async def test_async_streaming_missing_credentials_is_not_a_server_error(monkeypatch):
+    """
+    The helper alone is not enough: async_streaming's no-response branch used
+    to hard-code 500, discarding the status computed above it, so the same
+    missing-key error that maps to 400 on the non-streaming path surfaced as a
+    retryable 500 when stream=True. This exercises the full call path.
+    """
+    import litellm
+
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    monkeypatch.setattr(litellm, "api_key", None, raising=False)
+    monkeypatch.setattr(litellm, "openai_key", None, raising=False)
+
+    with pytest.raises(Exception) as exc_info:
+        await litellm.acompletion(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": "hi"}],
+            stream=True,
+        )
+
+    assert getattr(exc_info.value, "status_code", None) == 400