fix: thinking time on stop (#11900)

* fix: preserve thinking time when stopping reply

Fixes #11886

Signed-off-by: Calvin <calvinvwei@gmail.com>

* fix: also preserve thinking time when stopping during thinking

This extends the previous fix to also handle the case when the user
stops the reply while thinking is still in progress (not just after
thinking is complete).

Signed-off-by: Calvin <calvinvwei@gmail.com>

* fix: auto-complete thinking when text output starts

This fixes the issue where the thinking timer continues running after
thinking is complete and text output begins. Some AI providers don't
send a reasoning-end event explicitly, so we now auto-complete thinking
when a text-start event is received with accumulated reasoning content.

Fixes #11796

Signed-off-by: Calvin <calvinvwei@gmail.com>

* refactor: extract emitThinkingCompleteIfNeeded to reduce duplication

Extract the shared logic for emitting THINKING_COMPLETE chunk into a
reusable method. This removes code duplication between text-start and
reasoning-end event handlers as suggested in code review.

Signed-off-by: Calvin <calvinvwei@gmail.com>

---------

Signed-off-by: Calvin <calvinvwei@gmail.com>

src/renderer/src/aiCore/chunk/AiSdkToChunkAdapter.ts

@@ -120,6 +120,21 @@ export class AiSdkToChunkAdapter {
     }
   }
 
+  /**
+   * 如果有累积的思考内容，发送 THINKING_COMPLETE chunk 并清空
+   * @param final 包含 reasoningContent 的状态对象
+   * @returns 是否发送了 THINKING_COMPLETE chunk
+   */
+  private emitThinkingCompleteIfNeeded(final: { reasoningContent: string; [key: string]: any }) {
+    if (final.reasoningContent) {
+      this.onChunk({
+        type: ChunkType.THINKING_COMPLETE,
+        text: final.reasoningContent
+      })
+      final.reasoningContent = ''
+    }
+  }
+
   /**
    * 转换 AI SDK chunk 为 Cherry Studio chunk 并调用回调
    * @param chunk AI SDK 的 chunk 数据
@@ -145,6 +160,9 @@ export class AiSdkToChunkAdapter {
       }
       // === 文本相关事件 ===
       case 'text-start':
+        // 如果有未完成的思考内容，先生成 THINKING_COMPLETE
+        // 这处理了某些提供商不发送 reasoning-end 事件的情况
+        this.emitThinkingCompleteIfNeeded(final)
         this.onChunk({
           type: ChunkType.TEXT_START
         })
@@ -215,11 +233,7 @@ export class AiSdkToChunkAdapter {
         })
         break
       case 'reasoning-end':
-        this.onChunk({
-          type: ChunkType.THINKING_COMPLETE,
-          text: final.reasoningContent || ''
-        })
-        final.reasoningContent = ''
+        this.emitThinkingCompleteIfNeeded(final)
         break
 
       // === 工具调用相关事件（原始 AI SDK 事件，如果没有被中间件处理） ===

src/renderer/src/services/messageStreaming/callbacks/baseCallbacks.ts

@@ -8,7 +8,7 @@ import { updateOneBlock } from '@renderer/store/messageBlock'
 import { selectMessagesForTopic } from '@renderer/store/newMessage'
 import { newMessagesActions } from '@renderer/store/newMessage'
 import type { Assistant } from '@renderer/types'
-import type { PlaceholderMessageBlock, Response } from '@renderer/types/newMessage'
+import type { PlaceholderMessageBlock, Response, ThinkingMessageBlock } from '@renderer/types/newMessage'
 import { AssistantMessageStatus, MessageBlockStatus, MessageBlockType } from '@renderer/types/newMessage'
 import { uuid } from '@renderer/utils'
 import { isAbortError, serializeError } from '@renderer/utils/error'
@@ -29,10 +29,20 @@ interface BaseCallbacksDependencies {
   assistantMsgId: string
   saveUpdatesToDB: any
   assistant: Assistant
+  getCurrentThinkingInfo?: () => { blockId: string | null; millsec: number }
 }
 
 export const createBaseCallbacks = (deps: BaseCallbacksDependencies) => {
-  const { blockManager, dispatch, getState, topicId, assistantMsgId, saveUpdatesToDB, assistant } = deps
+  const {
+    blockManager,
+    dispatch,
+    getState,
+    topicId,
+    assistantMsgId,
+    saveUpdatesToDB,
+    assistant,
+    getCurrentThinkingInfo
+  } = deps
 
   const startTime = Date.now()
   const notificationService = NotificationService.getInstance()
@@ -98,10 +108,17 @@ export const createBaseCallbacks = (deps: BaseCallbacksDependencies) => {
       const possibleBlockId = findBlockIdForCompletion()
 
       if (possibleBlockId) {
-        // 更改上一个block的状态为ERROR
-        const changes = {
+        // 更改上一个block的状态为ERROR/PAUSED
+        const changes: Partial<ThinkingMessageBlock> = {
           status: isErrorTypeAbort ? MessageBlockStatus.PAUSED : MessageBlockStatus.ERROR
         }
+        // 如果是 thinking block，保留实际思考时间
+        if (blockManager.lastBlockType === MessageBlockType.THINKING) {
+          const thinkingInfo = getCurrentThinkingInfo?.()
+          if (thinkingInfo?.blockId === possibleBlockId && thinkingInfo?.millsec && thinkingInfo.millsec > 0) {
+            changes.thinking_millsec = thinkingInfo.millsec
+          }
+        }
         blockManager.smartBlockUpdate(possibleBlockId, changes, blockManager.lastBlockType!, true)
       }
 
@@ -111,13 +128,28 @@ export const createBaseCallbacks = (deps: BaseCallbacksDependencies) => {
       if (currentMessage) {
         const allBlockRefs = findAllBlocks(currentMessage)
         const blockState = getState().messageBlocks
+        // 获取当前思考信息（如果有），用于保留实际思考时间
+        const thinkingInfo = getCurrentThinkingInfo?.()
         for (const blockRef of allBlockRefs) {
           const block = blockState.entities[blockRef.id]
           if (block && block.status === MessageBlockStatus.STREAMING && block.id !== possibleBlockId) {
+            // 构建更新对象
+            const changes: Partial<ThinkingMessageBlock> = {
+              status: isErrorTypeAbort ? MessageBlockStatus.PAUSED : MessageBlockStatus.ERROR
+            }
+            // 如果是 thinking block 且有思考时间信息，保留实际思考时间
+            if (
+              block.type === MessageBlockType.THINKING &&
+              thinkingInfo?.blockId === block.id &&
+              thinkingInfo?.millsec &&
+              thinkingInfo.millsec > 0
+            ) {
+              changes.thinking_millsec = thinkingInfo.millsec
+            }
             dispatch(
               updateOneBlock({
                 id: block.id,
-                changes: { status: isErrorTypeAbort ? MessageBlockStatus.PAUSED : MessageBlockStatus.ERROR }
+                changes
               })
             )
           }

src/renderer/src/services/messageStreaming/callbacks/index.ts

@@ -23,6 +23,12 @@ interface CallbacksDependencies {
 export const createCallbacks = (deps: CallbacksDependencies) => {
   const { blockManager, dispatch, getState, topicId, assistantMsgId, saveUpdatesToDB, assistant } = deps
 
+  // 首先创建 thinkingCallbacks ，以便传递 getCurrentThinkingInfo 给 baseCallbacks
+  const thinkingCallbacks = createThinkingCallbacks({
+    blockManager,
+    assistantMsgId
+  })
+
   // 创建基础回调
   const baseCallbacks = createBaseCallbacks({
     blockManager,
@@ -31,13 +37,8 @@ export const createCallbacks = (deps: CallbacksDependencies) => {
     topicId,
     assistantMsgId,
     saveUpdatesToDB,
-    assistant
-  })
-
-  // 创建各类回调
-  const thinkingCallbacks = createThinkingCallbacks({
-    blockManager,
-    assistantMsgId
+    assistant,
+    getCurrentThinkingInfo: thinkingCallbacks.getCurrentThinkingInfo
   })
 
   const toolCallbacks = createToolCallbacks({

src/renderer/src/services/messageStreaming/callbacks/thinkingCallbacks.ts

@@ -19,6 +19,12 @@ export const createThinkingCallbacks = (deps: ThinkingCallbacksDependencies) =>
   let thinking_millsec_now: number = 0
 
   return {
+    // 获取当前思考时间（用于停止回复时保留思考时间）
+    getCurrentThinkingInfo: () => ({
+      blockId: thinkingBlockId,
+      millsec: thinking_millsec_now > 0 ? performance.now() - thinking_millsec_now : 0
+    }),
+
     onThinkingStart: async () => {
       if (blockManager.hasInitialPlaceholder) {
         const changes: Partial<MessageBlock> = {