@baishuyun/coze-provider 2.0.45 → 2.0.47

package/src/utils/accumulate-sse-response.ts

@@ -5,6 +5,7 @@ import {
 } from '@ai-sdk/provider';
 import { ServerSentEventTransformStream } from 'parse-sse';
 import { TextDecoderStream } from 'stream/web';
+import { mapCozeUsage } from './mcp-to-llm-chunk';
 
 export interface AccumulatedSSEResponse {
   content: LanguageModelV2Content[];
@@ -57,11 +58,7 @@ export async function accumulateSSEResponse(
       } else if (event === 'conversation.chat.completed' || event === 'done') {
         finishReason = 'stop';
         if (evtData.usage) {
-          usage = {
-            inputTokens: evtData.usage.input_count || 0,
-            outputTokens: evtData.usage.output_count || 0,
-            totalTokens: evtData.usage.total_count || 0,
-          };
+          usage = mapCozeUsage(evtData.usage);
         }
       } else if (event === 'error') {
         finishReason = 'error';

package/src/utils/mcp-to-llm-chunk.ts

@@ -1,6 +1,34 @@
 import { LanguageModelV2StreamPart, SharedV2ProviderMetadata } from '@ai-sdk/provider';
 import { ChatEventType } from '@coze/api';
 
+/**
+ * 上游 usage 字段兼容：官方 Coze v3 的 completed 事件为
+ * token_count/input_count/output_count（无 total_count），
+ * 总额按 total_count → token_count → input+output 的顺序取
+ */
+export const mapCozeUsage = (usage: any) => {
+  const inputTokens = usage?.input_count || 0;
+  const outputTokens = usage?.output_count || 0;
+  const totalTokens = usage?.total_count ?? usage?.token_count ?? inputTokens + outputTokens;
+  return { inputTokens, outputTokens, totalTokens };
+};
+
+/**
+ * 从消息 meta_data 提取 token 用量。上游会把本轮模型用量写入 answer 消息
+ * （含 workflow 用量消息）的 meta_data：token / input_tokens / output_tokens，
+ * 是 chat.completed 事件不带 usage 时（旧版上游）的兜底来源
+ */
+export const extractUsageFromMsgMeta = (meta: any) => {
+  if (!meta) return null;
+  const totalTokens = Number(meta.token);
+  if (!totalTokens) return null;
+  return {
+    inputTokens: Number(meta.input_tokens) || 0,
+    outputTokens: Number(meta.output_tokens) || 0,
+    totalTokens,
+  };
+};
+
 export const mcp2llmChunk = (part: any): LanguageModelV2StreamPart | null => {
   const fullData = JSON.stringify(part);
 
@@ -64,11 +92,7 @@ export const mcp2llmChunk = (part: any): LanguageModelV2StreamPart | null => {
       return {
         type: 'finish',
         finishReason: 'stop',
-        usage: {
-          inputTokens: part.data.usage?.input_count || 0,
-          outputTokens: part.data.usage?.output_count || 0,
-          totalTokens: part.data.usage?.total_count || 0,
-        },
+        usage: mapCozeUsage(part.data.usage),
       };
       break;
 
@@ -77,11 +101,7 @@ export const mcp2llmChunk = (part: any): LanguageModelV2StreamPart | null => {
       return {
         type: 'finish',
         finishReason: 'stop',
-        usage: {
-          inputTokens: part.data.usage?.input_count || 0,
-          outputTokens: part.data.usage?.output_count || 0,
-          totalTokens: part.data.usage?.total_count || 0,
-        },
+        usage: mapCozeUsage(part.data.usage),
       };
       break;