Merge 57f7564f84 into 527d882508

litellm/proxy/common_request_processing.py

@@ -1035,10 +1035,38 @@ async def _parse_event_data_for_error(event_line: str | bytes) -> int | None:
                         )
                         # Not a valid integer string, treat as if no valid code was found for this check
 
-                # Ensure error_code is a valid HTTP status code
-                if error_code is not None and 100 <= error_code <= 599:
-                    return error_code
-                elif error_code_raw is not None:  # Log if original code was present but not valid
+                if error_code is not None:
+                    if 100 <= error_code <= 599:
+                        # Standard HTTP status code
+                        return error_code
+                    elif error_code >= 600:
+                        # Vendor-specific error codes outside the standard HTTP
+                        # range (e.g., ZAI/ZhipuAI 1302 rate limit, DashScope
+                        # throttling codes). Previously these were silently
+                        # dropped (returned None), causing the proxy to treat
+                        # the SSE chunk as normal content — the error was
+                        # forwarded to the client as a 200 response, and router
+                        # fallback/cooldown never triggered.
+                        #
+                        # Map to 502 (Bad Gateway) so that failure handling
+                        # (retry, fallback, cooldown) fires correctly. The
+                        # original vendor code is preserved in the warning log
+                        # for debugging.
+                        verbose_proxy_logger.warning(
+                            "SSE error code %s is outside standard HTTP range (100-599); mapping to 502 for failure handling",
+                            error_code,
+                        )
+                        return 502
+                    else:
+                        # Integer code below 100 (or negative) — likely corrupt data
+                        # but keep the original behaviour of returning None while
+                        # emitting the diagnostic warning (preserved from the
+                        # original elif error_code_raw is not None: branch).
+                        verbose_proxy_logger.warning("Error has invalid or non-convertible code: %s", error_code)
+                elif error_code_raw is not None:
+                    # error_code could not be coerced to int from the raw value
+                    # (e.g., a non-numeric string). Log the original raw value
+                    # for debugging and fall through to return None.
                     verbose_proxy_logger.warning("Error has invalid or non-convertible code: %s", error_code_raw)
         except (orjson.JSONDecodeError, json.JSONDecodeError):
             # not a known error chunk

tests/unit/proxy/test_common_request_processing.py

@@ -2283,11 +2283,15 @@ class TestCommonRequestProcessingHelpers:
             (
                 'data: {"error": {"code": 99, "message": "too low"}}',
                 None,
-            ),  # Integer code too low
+            ),  # Integer code too low (< 100, likely invalid data)
             (
-                'data: {"error": {"code": 600, "message": "too high"}}',
-                None,
-            ),  # Integer code too high
+                'data: {"error": {"code": 600, "message": "vendor error"}}',
+                502,
+            ),  # Non-standard code >= 600 → mapped to 502 for failure handling
+            (
+                'data: {"error": {"code": 1302, "message": "rate limit"}}',
+                502,
+            ),  # Vendor-specific rate limit code (ZAI/ZhipuAI) → mapped to 502
             (
                 'data: {"id": "123", "content": "hello"}',
                 None,