fix(proxy): preserve HTTPException headers (retry-after) when converting to ProxyException

proxy_exception_from_http_exception() replaced the exception's own
headers with only the caller-supplied dict, so a 429 raised by the rate
limiter with retry-after lost that header on routes that convert
HTTPException -> ProxyException (e.g. /v1/messages, unlike
/v1/chat/completions where FastAPI serves HTTPException.headers
directly). Merge exc.headers under the caller-supplied headers.

Fixes #46630

litellm/proxy/common_request_processing.py

@@ -637,13 +637,16 @@ def proxy_exception_from_http_exception(exc: HTTPException, headers: dict[str, s
     existing_fields: Final = getattr(exc, "provider_specific_fields", None) or {}
     merged_fields: Final = {**existing_fields, **structured_fields} if structured_fields else (existing_fields or None)
     error_status: Final = error_status_code(exc, status.HTTP_400_BAD_REQUEST)
+    # Preserve headers carried on the HTTPException itself (e.g. `retry-after`
+    # set by the rate limiter on 429); caller-supplied headers win on conflict.
+    merged_headers: Final = {**(exc.headers or {}), **headers}
     return ProxyException(
         message=message,
         type=openai_error_type(exc, error_status),
         param=openai_error_param(exc),
         code=error_status,
         provider_specific_fields=merged_fields,
-        headers=headers,
+        headers=merged_headers,
     )
 
 

tests/unit/proxy/test_common_request_processing.py

@@ -2355,6 +2355,22 @@ class TestCommonRequestProcessingHelpers:
         assert plain.code == "429"
         assert plain.provider_specific_fields is None
 
+    async def test_proxy_exception_from_http_exception_preserves_exc_headers(self):
+        """Headers carried on the HTTPException itself (e.g. `retry-after` set by
+        the rate limiter on 429) must survive the conversion to ProxyException;
+        caller-supplied headers win on conflict."""
+        from litellm.proxy.common_request_processing import (
+            proxy_exception_from_http_exception,
+        )
+
+        exc = HTTPException(status_code=429, detail="slow down", headers={"retry-after": "60"})
+        result = proxy_exception_from_http_exception(exc, {"x-litellm-call-id": "abc"})
+        assert result.headers == {"retry-after": "60", "x-litellm-call-id": "abc"}
+
+        conflict = HTTPException(status_code=429, detail="slow down", headers={"retry-after": "60"})
+        overridden = proxy_exception_from_http_exception(conflict, {"retry-after": "30"})
+        assert overridden.headers == {"retry-after": "30"}
+
     async def test_proxy_exception_from_route_error_helper(self) -> None:
         """The shared route error -> ProxyException conversion keeps the status an exception
         carries, never the OpenAI SDK's ``code`` field (``None`` on most litellm exceptions)."""