From ed1bc3f0d3ba14c8f7c50bd0c7e4cc961d0d7d80 Mon Sep 17 00:00:00 2001
From: nileshpatil6 <technil6436@gmail.com>
Date: Sun, 3 May 2026 12:42:21 +0530
Subject: [PATCH] fix(exceptions): handle non-integer status_code in
 MidStreamFallbackError

When a CustomLLM handler raises RateLimitError inside astreaming(),
the error propagates through the streaming pipeline and hits
MidStreamFallbackError.__init__. In some code paths the original
exception's status_code attribute is a non-integer string (e.g.
'litellm_error'), causing int() to raise ValueError and crash with
HTTP 500 instead of triggering the fallback chain.

Wrap both int() calls with try/except (ValueError, TypeError) and
fall back to 503 so the fallback mechanism can engage normally.

Adds regression test: test_midstream_fallback_error_non_integer_status_code
verifies status_code and response.status_code both resolve to 503.

Fixes #26913
---
 litellm/exceptions.py                         | 14 +++++++++++--
 .../test_exception_header_preservation.py     | 21 +++++++++++++++++++
 2 files changed, 33 insertions(+), 2 deletions(-)

diff --git a/litellm/exceptions.py b/litellm/exceptions.py
index 8b005291556..825ae2528d8 100644
--- a/litellm/exceptions.py
+++ b/litellm/exceptions.py
@@ -955,7 +955,12 @@ class MidStreamFallbackError(ServiceUnavailableError):  # type: ignore
         is_pre_first_chunk: bool = False,
     ):
         original_status = getattr(original_exception, "status_code", None)
-        self.status_code = int(original_status) if original_status is not None else 503
+        try:
+            self.status_code = (
+                int(original_status) if original_status is not None else 503
+            )
+        except (ValueError, TypeError):
+            self.status_code = 503
         self.message = f"litellm.MidStreamFallbackError: {message}"
         self.model = model
         self.llm_provider = llm_provider
@@ -997,7 +1002,12 @@ class MidStreamFallbackError(ServiceUnavailableError):  # type: ignore
         )
 
         # Restore the propagated status and original response/request objects
-        self.status_code = int(original_status) if original_status is not None else 503
+        try:
+            self.status_code = (
+                int(original_status) if original_status is not None else 503
+            )
+        except (ValueError, TypeError):
+            self.status_code = 503
         self.response = _saved_response
         self.request = _saved_request
         self.message = _saved_message
diff --git a/tests/test_litellm/test_exception_header_preservation.py b/tests/test_litellm/test_exception_header_preservation.py
index 6ea478c633b..bafa1ad38d3 100644
--- a/tests/test_litellm/test_exception_header_preservation.py
+++ b/tests/test_litellm/test_exception_header_preservation.py
@@ -254,6 +254,27 @@ class TestExceptionAttributes:
         assert midstream_fallback.response.status_code == 503
         assert str(midstream_fallback.response.request.url) == "https://openai.com/v1/"
 
+    def test_midstream_fallback_error_non_integer_status_code(self):
+        """
+        MidStreamFallbackError must not crash when original_exception.status_code
+        is a non-integer string (e.g. 'litellm_error'). Before the fix, int()
+        raised ValueError and the error handler returned HTTP 500 instead of
+        engaging the fallback chain.
+        """
+
+        class _FakeException(Exception):
+            status_code = "litellm_error"
+
+        midstream_error = MidStreamFallbackError(
+            message="stream broke",
+            model="gpt-4o-mini",
+            llm_provider="openai",
+            original_exception=_FakeException("boom"),
+        )
+
+        assert midstream_error.status_code == 503
+        assert midstream_error.response.status_code == 503
+
 
 class TestProxyHeaderExtraction:
     """Test that proxy correctly extracts headers from exceptions."""