fix(streaming): preserve provider status for string error codes

litellm/exceptions.py

@@ -1086,6 +1086,15 @@ class BlockedPiiEntityError(Exception):
         super().__init__(self.message)
 
 
+def _numeric_status_code(value: object) -> int | None:
+    if value is None:
+        return None
+    try:
+        return int(value)
+    except (TypeError, ValueError):
+        return None
+
+
 class MidStreamFallbackError(ServiceUnavailableError):
     def __init__(
         self,
@@ -1100,8 +1109,13 @@ class MidStreamFallbackError(ServiceUnavailableError):
         generated_content: str = "",
         is_pre_first_chunk: bool = False,
     ):
-        original_status: Final = getattr(original_exception, "status_code", None)
-        self.status_code = int(original_status) if original_status is not None else 503
+        original_status: Final = _numeric_status_code(getattr(original_exception, "status_code", None))
+        original_response: Final = getattr(original_exception, "response", None)
+        response_status: Final = _numeric_status_code(
+            getattr(response, "status_code", getattr(original_response, "status_code", None))
+        )
+        status_code: Final = original_status or response_status or 503
+        self.status_code = status_code
         self.message = f"litellm.MidStreamFallbackError: {message}"
         self.model = model
         self.llm_provider = llm_provider
@@ -1143,7 +1157,7 @@ class MidStreamFallbackError(ServiceUnavailableError):
         )
 
         # Restore the propagated status and original response/request objects
-        self.status_code = int(original_status) if original_status is not None else 503
+        self.status_code = status_code
         self.response = _saved_response
         self.request = _saved_request
         self.message = _saved_message

tests/test_litellm/test_exception_header_preservation.py

@@ -22,6 +22,13 @@ from litellm.exceptions import (
 )
 
 
+class ProviderToolSchemaError(Exception):
+    def __init__(self, response: httpx.Response) -> None:
+        self.status_code = "tool_use_failed"
+        self.response = response
+        super().__init__("tool call validation failed")
+
+
 class TestExceptionHeaderPreservation:
     """Test that exception classes preserve headers from provider responses."""
 
@@ -255,6 +262,25 @@ 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_accepts_non_numeric_provider_status(self):
+        """Provider error codes can be strings while the response still has an HTTP status."""
+        original_response = httpx.Response(
+            status_code=400,
+            request=httpx.Request("POST", "https://api.groq.com/openai/v1/chat/completions"),
+        )
+        provider_error = ProviderToolSchemaError(original_response)
+
+        midstream_error = MidStreamFallbackError(
+            message=str(provider_error),
+            model="openai/gpt-oss-120b",
+            llm_provider="groq",
+            original_exception=provider_error,
+        )
+
+        assert midstream_error.status_code == 400
+        assert midstream_error.response.status_code == 400
+        assert "tool call validation failed" in midstream_error.message
+
 
 class TestProxyHeaderExtraction:
     """Test that proxy correctly extracts headers from exceptions."""