- 代理发给上游时，仍然会补 stream_options.include_usage=true，所以 token 统计还在。

- 下游如果没有显式传 stream_options.include_usage=true，代理会把“纯 usage 的 SSE chunk”吞掉，不再往客户端发。
- 如果下游本来就显式要求 include_usage=true，那仍然会原样透传 usage chunk。

src/external/response_adapter.py

@@ -21,6 +21,7 @@ def build_proxy_response(
     stream_context_func: Any,
     logger: Logger,
     on_complete: Optional[Callable[[Dict[str, Any]], None]] = None,
+    forward_stream_usage: bool = False,
 ) -> Response:
     """基于上游响应与输出钩子构建 Flask 响应。"""
     hook_func = getattr(hook, "output_body_hook", None) if hook else None
@@ -62,6 +63,14 @@ def build_proxy_response(
             return None
         return stripped[5:].strip()
 
+    def is_usage_only_stream_chunk(payload: Any) -> bool:
+        if not isinstance(payload, dict):
+            return False
+        if not isinstance(payload.get("usage"), dict):
+            return False
+        choices = payload.get("choices")
+        return choices is None or (isinstance(choices, list) and len(choices) == 0)
+
     def process_sse_event(event_text: str) -> Iterator[bytes]:
         # Preserve non-data SSE fields to avoid breaking strict clients.
         normalized_text = event_text.replace("\r\n", "\n").replace("\r", "\n")
@@ -97,6 +106,8 @@ def build_proxy_response(
 
         if isinstance(data, dict):
             update_meta_from_payload(data)
+            if not forward_stream_usage and is_usage_only_stream_chunk(data):
+                return
 
         if hook_func:
             modified = hook_func(ctx, data)

src/presentation/proxy_controller.py

@@ -71,6 +71,9 @@ class ProxyController:
                 )
 
             stream_options = request_data.get("stream_options")
+            client_requested_usage_chunk = isinstance(stream_options, dict) and (
+                stream_options.get("include_usage") is True
+            )
             if not isinstance(stream_options, dict):
                 stream_options = {}
             if stream_options.get("include_usage") is not True:
@@ -109,7 +112,11 @@ class ProxyController:
                 )
 
             result, status_code = self._proxy_service.proxy_request(
-                provider, request_data, headers, on_complete=on_proxy_complete
+                provider,
+                request_data,
+                headers,
+                on_complete=on_proxy_complete,
+                forward_stream_usage=client_requested_usage_chunk,
             )
             if result is None:
                 self._logger.error(f"Proxy failed after retries: model={model_name!r}, ip={client_ip}")

src/services/proxy_service.py

@@ -30,6 +30,7 @@ class ProxyService:
         request_data: Dict[str, Any],
         request_headers: Dict[str, str],
         on_complete: Optional[Callable[[Dict[str, Any]], None]] = None,
+        forward_stream_usage: bool = False,
     ) -> Tuple[Optional[Response], int]:
         """代理请求到目标 provider，并处理重试与输出钩子。"""
         target_url = provider.api
@@ -123,6 +124,7 @@ class ProxyService:
                     stream_with_context,
                     self._logger,
                     on_complete=on_complete,
+                    forward_stream_usage=forward_stream_usage,
                 )
                 self._logger.info(
                     "Upstream request completed: provider=%s status=%s stream=%s",