fix(gateway): 修复 SSE 流式响应 usage 统计错误

message_delta 应完全覆盖 message_start 的 usage 数据,
而非仅在值为 0 时才更新。
This commit is contained in:
shaw
2026-01-27 09:16:34 +08:00
parent 45676fdc8d
commit 56a1e29cdd

View File

@@ -3372,21 +3372,14 @@ func (s *GatewayService) parseSSEUsage(data string, usage *ClaudeUsage) {
} `json:"usage"` } `json:"usage"`
} }
if json.Unmarshal([]byte(data), &msgDelta) == nil && msgDelta.Type == "message_delta" { if json.Unmarshal([]byte(data), &msgDelta) == nil && msgDelta.Type == "message_delta" {
// output_tokens 总是从 message_delta 获取 // message_delta 是推理结束后的最终统计,应完全覆盖 message_start 的数据
usage.OutputTokens = msgDelta.Usage.OutputTokens // 这对于 Claude API 和 GLM 等兼容 API 都是正确的行为
// 如果 message_start 中没有值,则从 message_delta 获取兼容GLM等API
if usage.InputTokens == 0 {
usage.InputTokens = msgDelta.Usage.InputTokens usage.InputTokens = msgDelta.Usage.InputTokens
} usage.OutputTokens = msgDelta.Usage.OutputTokens
if usage.CacheCreationInputTokens == 0 {
usage.CacheCreationInputTokens = msgDelta.Usage.CacheCreationInputTokens usage.CacheCreationInputTokens = msgDelta.Usage.CacheCreationInputTokens
}
if usage.CacheReadInputTokens == 0 {
usage.CacheReadInputTokens = msgDelta.Usage.CacheReadInputTokens usage.CacheReadInputTokens = msgDelta.Usage.CacheReadInputTokens
} }
} }
}
func (s *GatewayService) handleNonStreamingResponse(ctx context.Context, resp *http.Response, c *gin.Context, account *Account, originalModel, mappedModel string) (*ClaudeUsage, error) { func (s *GatewayService) handleNonStreamingResponse(ctx context.Context, resp *http.Response, c *gin.Context, account *Account, originalModel, mappedModel string) (*ClaudeUsage, error) {
// 更新5h窗口状态 // 更新5h窗口状态