refactor: update AiSdkToChunkAdapter and middleware for improved chunk handling

- Modified convertAndEmitChunk method to handle new chunk types and streamline processing.
- Adjusted thinkingTimeMiddleware to remove unnecessary parameters and enhance clarity.
- Updated middleware integration in AiSdkMiddlewareBuilder for better middleware management.

src/renderer/src/aiCore/AiSdkToChunkAdapter.ts

@@ -68,7 +68,7 @@ export class AiSdkToChunkAdapter {
    * 转换 AI SDK chunk 为 Cherry Studio chunk 并调用回调
    * @param chunk AI SDK 的 chunk 数据
    */
-  private convertAndEmitChunk(chunk: TextStreamPart<any>, final: { text: string; reasoning_content: string }) {
+  private convertAndEmitChunk(chunk: any, final: { text: string; reasoning_content: string }) {
     console.log('AI SDK chunk type:', chunk.type, chunk)
     switch (chunk.type) {
       // === 文本相关事件 ===
@@ -80,12 +80,24 @@ export class AiSdkToChunkAdapter {
         })
         break
 
-      case 'reasoning-signature':
-        // 不再需要处理，中间件会发出 THINKING_COMPLETE
+      case 'reasoning':
+        this.onChunk({
+          type: ChunkType.THINKING_DELTA,
+          text: chunk.textDelta || ''
+        })
         break
-
       case 'redacted-reasoning':
-        // 不再需要处理
+        this.onChunk({
+          type: ChunkType.THINKING_DELTA,
+          text: chunk.data || ''
+        })
+        break
+      case 'reasoning-signature':
+        this.onChunk({
+          type: ChunkType.THINKING_COMPLETE,
+          text: chunk.text || '',
+          thinking_millsec: chunk.thinking_millsec || 0
+        })
         break
 
       // === 工具调用相关事件 ===

src/renderer/src/aiCore/index_new.ts

@@ -19,7 +19,6 @@ import {
 } from '@cherrystudio/ai-core'
 import { isDedicatedImageGenerationModel } from '@renderer/config/models'
 import type { GenerateImageParams, Model, Provider } from '@renderer/types'
-import { Chunk } from '@renderer/types/chunk'
 
 // 引入适配器
 import AiSdkToChunkAdapter from './AiSdkToChunkAdapter'
@@ -121,15 +120,13 @@ export default class ModernAiProvider {
   public async completions(
     modelId: string,
     params: StreamTextParams,
-    middlewareConfig: AiSdkMiddlewareConfig,
-    onChunk?: (chunk: Chunk) => void
+    middlewareConfig: AiSdkMiddlewareConfig
   ): Promise<CompletionsResult> {
     // const model = params.assistant.model
 
     // 检查是否应该使用现代化客户端
     // if (this.modernClient && model && isModernSdkSupported(this.provider, model)) {
     // try {
-    console.log('completions', modelId, params, onChunk)
     return await this.modernCompletions(modelId, params, middlewareConfig)
     // } catch (error) {
     // console.warn('Modern client failed, falling back to legacy:', error)

src/renderer/src/aiCore/middleware/aisdk/AiSdkMiddlewareBuilder.ts

@@ -100,7 +100,7 @@ export function buildAiSdkMiddlewares(config: AiSdkMiddlewareConfig): NamedAiSdk
   if (config.onChunk && config.model && isReasoningModel(config.model)) {
     builder.add({
       name: 'thinking-time',
-      aiSdkMiddlewares: [thinkingTimeMiddleware(config.onChunk)]
+      aiSdkMiddlewares: [thinkingTimeMiddleware()]
     })
   }
 

src/renderer/src/aiCore/middleware/aisdk/ThinkingTimeMiddleware.ts

@@ -1,5 +1,5 @@
 import { LanguageModelV1Middleware, LanguageModelV1StreamPart } from '@cherrystudio/ai-core'
-import { Chunk, ChunkType, ThinkingCompleteChunk } from '@renderer/types/chunk'
+import { ChunkType, ThinkingCompleteChunk } from '@renderer/types/chunk'
 
 /**
  * 一个用于统计 LLM "思考时间"（Time to First Token）的 AI SDK 中间件。
@@ -9,15 +9,16 @@ import { Chunk, ChunkType, ThinkingCompleteChunk } from '@renderer/types/chunk'
  * 2. 它会创建一个新的 `TransformStream` 来代理原始的流。
  * 3. 当第一个数据块 (chunk) 从原始流中到达时，记录结束时间。
  * 4. 计算两者之差，即为 "思考时间"
+ * 这里只处理了thinking_complete
  */
-export default function thinkingTimeMiddleware(onChunkReceived: (chunk: Chunk) => void): LanguageModelV1Middleware {
+export default function thinkingTimeMiddleware(): LanguageModelV1Middleware {
   return {
     wrapStream: async ({ doStream }) => {
       let hasThinkingContent = false
       let thinkingStartTime = 0
       let accumulatedThinkingContent = ''
       const { stream, ...reset } = await doStream()
-      const transformStream = new TransformStream<LanguageModelV1StreamPart, LanguageModelV1StreamPart>({
+      const transformStream = new TransformStream<LanguageModelV1StreamPart, any>({
         transform(chunk, controller) {
           if (chunk.type === 'reasoning' || chunk.type === 'redacted-reasoning') {
             if (!hasThinkingContent) {
@@ -25,19 +26,15 @@ export default function thinkingTimeMiddleware(onChunkReceived: (chunk: Chunk) =
               thinkingStartTime = Date.now()
             }
             accumulatedThinkingContent += chunk.textDelta || ''
-            onChunkReceived({
-              type: ChunkType.THINKING_DELTA,
-              text: chunk.textDelta || ''
-            })
           } else {
             if (hasThinkingContent && thinkingStartTime > 0) {
               const thinkingTime = Date.now() - thinkingStartTime
-              const thinkingCompleteChunk: ThinkingCompleteChunk = {
-                type: ChunkType.THINKING_COMPLETE,
+              const thinkingCompleteChunk = {
+                type: 'reasoning-signature',
                 text: accumulatedThinkingContent,
                 thinking_millsec: thinkingTime
               }
-              onChunkReceived(thinkingCompleteChunk)
+              controller.enqueue(thinkingCompleteChunk)
               hasThinkingContent = false
               thinkingStartTime = 0
               accumulatedThinkingContent = ''
@@ -55,7 +52,7 @@ export default function thinkingTimeMiddleware(onChunkReceived: (chunk: Chunk) =
               text: accumulatedThinkingContent,
               thinking_millsec: thinkingTime
             }
-            onChunkReceived(thinkingCompleteChunk)
+            controller.enqueue(thinkingCompleteChunk)
           }
           controller.terminate()
         }