- 上游 chat 流 finish_reason 块先于 usage 块, 原实现在 finish 时即发 response.completed 导致 usage 为空; 改为等 usage 块到达(或 [DONE] 兜底)再发, completed 携带 input/output tokens Co-Authored-By: Claude <noreply@anthropic.com>
- 上游 chat 流 finish_reason 块先于 usage 块, 原实现在 finish 时即发 response.completed 导致 usage 为空; 改为等 usage 块到达(或 [DONE] 兜底)再发, completed 携带 input/output tokens Co-Authored-By: Claude <noreply@anthropic.com>