fix: properly handle thinking mode switching in qwenThinkingMiddleware

Fix issue #11612 where Ollama couldn't disable Qwen3 thinking mode.
The middleware now properly removes existing /think or /no_think suffixes
before adding the correct suffix based on current reasoning_effort setting.

Previously, messages with existing suffixes were skipped, causing thinking
mode to persist even when user switched back to non-thinking mode.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

src/renderer/src/aiCore/legacy/clients/BaseApiClient.ts

@@ -9,7 +9,6 @@ import {
 import { REFERENCE_PROMPT } from '@renderer/config/prompts'
 import { getLMStudioKeepAliveTime } from '@renderer/hooks/useLMStudio'
 import { getAssistantSettings } from '@renderer/services/AssistantService'
-import type { RootState } from '@renderer/store'
 import type {
   Assistant,
   GenerateImageParams,
@@ -246,20 +245,23 @@ export abstract class BaseApiClient<
 
   protected getVerbosity(model?: Model): OpenAIVerbosity {
     try {
-      const state = window.store?.getState() as RootState
+      const state = window.store?.getState()
       const verbosity = state?.settings?.openAI?.verbosity
 
-      // If model is provided, check if the verbosity is supported by the model
-      if (model) {
-        const supportedVerbosity = getModelSupportedVerbosity(model)
-        // Use user's verbosity if supported, otherwise use the first supported option
-        return supportedVerbosity.includes(verbosity) ? verbosity : supportedVerbosity[0]
+      if (verbosity && ['low', 'medium', 'high'].includes(verbosity)) {
+        // If model is provided, check if the verbosity is supported by the model
+        if (model) {
+          const supportedVerbosity = getModelSupportedVerbosity(model)
+          // Use user's verbosity if supported, otherwise use the first supported option
+          return supportedVerbosity.includes(verbosity) ? verbosity : supportedVerbosity[0]
+        }
+        return verbosity
       }
-      return verbosity
     } catch (error) {
-      logger.warn('Failed to get verbosity from state. Fallback to undefined.', error as Error)
-      return undefined
+      logger.warn('Failed to get verbosity from state:', error as Error)
     }
+
+    return 'medium'
   }
 
   protected getTimeout(model: Model) {

src/renderer/src/aiCore/legacy/clients/openai/OpenAIApiClient.ts

@@ -32,6 +32,7 @@ import {
   isSupportedThinkingTokenModel,
   isSupportedThinkingTokenQwenModel,
   isSupportedThinkingTokenZhipuModel,
+  isSupportVerbosityModel,
   isVisionModel,
   MODEL_SUPPORTED_REASONING_EFFORT,
   ZHIPU_RESULT_TOKENS
@@ -713,8 +714,13 @@ export class OpenAIAPIClient extends OpenAIBaseClient<
           ...modalities,
           // groq 有不同的 service tier 配置，不符合 openai 接口类型
           service_tier: this.getServiceTier(model) as OpenAIServiceTier,
-          // verbosity. getVerbosity ensure the returned value is valid.
-          verbosity: this.getVerbosity(model),
+          ...(isSupportVerbosityModel(model)
+            ? {
+                text: {
+                  verbosity: this.getVerbosity(model)
+                }
+              }
+            : {}),
           ...this.getProviderSpecificParameters(assistant, model),
           ...reasoningEffort,
           // ...getOpenAIWebSearchParams(model, enableWebSearch),

src/renderer/src/aiCore/middleware/qwenThinkingMiddleware.ts

@@ -23,8 +23,10 @@ export function qwenThinkingMiddleware(enableThinking: boolean): LanguageModelMi
             // Process content array
             if (Array.isArray(message.content)) {
               for (const part of message.content) {
-                if (part.type === 'text' && !part.text.endsWith('/think') && !part.text.endsWith('/no_think')) {
-                  part.text += suffix
+                if (part.type === 'text') {
+                  // Remove any existing thinking suffixes first, then add the correct one
+                  const cleanText = part.text.replace(/\s*\/think\s*$/, '').replace(/\s*\/no_think\s*$/, '')
+                  part.text = cleanText + suffix
                 }
               }
             }

src/renderer/src/config/models/__tests__/utils.test.ts

@@ -222,22 +222,18 @@ describe('model utils', () => {
 
     describe('getModelSupportedVerbosity', () => {
       it('returns only "high" for GPT-5 Pro models', () => {
-        expect(getModelSupportedVerbosity(createModel({ id: 'gpt-5-pro' }))).toEqual([undefined, null, 'high'])
-        expect(getModelSupportedVerbosity(createModel({ id: 'gpt-5-pro-2025-10-06' }))).toEqual([
-          undefined,
-          null,
-          'high'
-        ])
+        expect(getModelSupportedVerbosity(createModel({ id: 'gpt-5-pro' }))).toEqual([undefined, 'high'])
+        expect(getModelSupportedVerbosity(createModel({ id: 'gpt-5-pro-2025-10-06' }))).toEqual([undefined, 'high'])
       })
 
       it('returns all levels for non-Pro GPT-5 models', () => {
         const previewModel = createModel({ id: 'gpt-5-preview' })
-        expect(getModelSupportedVerbosity(previewModel)).toEqual([undefined, null, 'low', 'medium', 'high'])
+        expect(getModelSupportedVerbosity(previewModel)).toEqual([undefined, 'low', 'medium', 'high'])
       })
 
       it('returns all levels for GPT-5.1 models', () => {
         const gpt51Model = createModel({ id: 'gpt-5.1-preview' })
-        expect(getModelSupportedVerbosity(gpt51Model)).toEqual([undefined, null, 'low', 'medium', 'high'])
+        expect(getModelSupportedVerbosity(gpt51Model)).toEqual([undefined, 'low', 'medium', 'high'])
       })
 
       it('returns only undefined for non-GPT-5 models', () => {

src/renderer/src/config/models/utils.ts

@@ -10,8 +10,7 @@ import {
   isGPT51SeriesModel,
   isOpenAIChatCompletionOnlyModel,
   isOpenAIOpenWeightModel,
-  isOpenAIReasoningModel,
-  isSupportVerbosityModel
+  isOpenAIReasoningModel
 } from './openai'
 import { isQwenMTModel } from './qwen'
 import { isGenerateImageModel, isTextToImageModel, isVisionModel } from './vision'
@@ -155,10 +154,10 @@ const MODEL_SUPPORTED_VERBOSITY: readonly {
  * For GPT-5-pro, only 'high' is supported; for other GPT-5 models, 'low', 'medium', and 'high' are supported.
  * For GPT-5.1 series models, 'low', 'medium', and 'high' are supported.
  * @param model - The model to check
- * @returns An array of supported verbosity levels, always including `undefined` as the first element and `null` when applicable
+ * @returns An array of supported verbosity levels, always including `undefined` as the first element
  */
 export const getModelSupportedVerbosity = (model: Model | undefined | null): OpenAIVerbosity[] => {
-  if (!model || !isSupportVerbosityModel(model)) {
+  if (!model) {
     return [undefined]
   }
 
@@ -166,7 +165,7 @@ export const getModelSupportedVerbosity = (model: Model | undefined | null): Ope
 
   for (const { validator, values } of MODEL_SUPPORTED_VERBOSITY) {
     if (validator(model)) {
-      supportedValues = [null, ...values]
+      supportedValues = [...values]
       break
     }
   }

src/renderer/src/pages/home/Tabs/components/OpenAISettingsGroup.tsx

@@ -24,12 +24,12 @@ import { useTranslation } from 'react-i18next'
 import { useSelector } from 'react-redux'
 
 type VerbosityOption = {
-  value: NonNullable<OpenAIVerbosity> | 'undefined' | 'null'
+  value: NonNullable<OpenAIVerbosity> | 'undefined'
   label: string
 }
 
 type SummaryTextOption = {
-  value: NonNullable<OpenAISummaryText> | 'undefined' | 'null'
+  value: NonNullable<OpenAISummaryText> | 'undefined'
   label: string
 }
 
@@ -85,10 +85,6 @@ const OpenAISettingsGroup: FC<Props> = ({ model, providerId, SettingGroup, Setti
       value: 'undefined',
       label: t('common.ignore')
     },
-    {
-      value: 'null',
-      label: t('common.off')
-    },
     {
       value: 'auto',
       label: t('settings.openai.summary_text_mode.auto')
@@ -109,10 +105,6 @@ const OpenAISettingsGroup: FC<Props> = ({ model, providerId, SettingGroup, Setti
         value: 'undefined',
         label: t('common.ignore')
       },
-      {
-        value: 'null',
-        label: t('common.off')
-      },
       {
         value: 'low',
         label: t('settings.openai.verbosity.low')
@@ -211,9 +203,9 @@ const OpenAISettingsGroup: FC<Props> = ({ model, providerId, SettingGroup, Setti
                 </Tooltip>
               </SettingRowTitleSmall>
               <Selector
-                value={toOptionValue(summaryText)}
+                value={summaryText}
                 onChange={(value) => {
-                  setSummaryText(toRealValue(value))
+                  setSummaryText(value as OpenAISummaryText)
                 }}
                 options={summaryTextOptions}
               />
@@ -230,9 +222,9 @@ const OpenAISettingsGroup: FC<Props> = ({ model, providerId, SettingGroup, Setti
               </Tooltip>
             </SettingRowTitleSmall>
             <Selector
-              value={toOptionValue(verbosity)}
+              value={verbosity}
               onChange={(value) => {
-                setVerbosity(toRealValue(value))
+                setVerbosity(value as OpenAIVerbosity)
               }}
               options={verbosityOptions}
             />

src/renderer/src/store/migrate.ts

@@ -2906,23 +2906,6 @@ const migrateConfig = {
       logger.error('migrate 179 error', error as Error)
       return state
     }
-  },
-  '180': (state: RootState) => {
-    try {
-      // @ts-expect-error
-      if (state.settings.openAI.summaryText === 'undefined') {
-        state.settings.openAI.summaryText = undefined
-      }
-      // @ts-expect-error
-      if (state.settings.openAI.verbosity === 'undefined') {
-        state.settings.openAI.summaryText = undefined
-      }
-      logger.info('migrate 180 success')
-      return state
-    } catch (error) {
-      logger.error('migrate 180 error', error as Error)
-      return state
-    }
   }
 }
 

src/renderer/src/types/aiCoreTypes.ts

@@ -1,5 +1,5 @@
 import type OpenAI from '@cherrystudio/openai'
-import type { NotUndefined } from '@types'
+import type { NotNull, NotUndefined } from '@types'
 import type { ImageModel, LanguageModel } from 'ai'
 import type { generateObject, generateText, ModelMessage, streamObject, streamText } from 'ai'
 import * as z from 'zod'
@@ -32,15 +32,17 @@ export type GenerateObjectParams = Omit<Parameters<typeof generateObject>[0], 'm
 export type AiSdkModel = LanguageModel | ImageModel
 
 // The original type unite both undefined and null.
+// I pick undefined as the unique falsy type since they seem like share the same meaning according to OpenAI API docs.
 // Parameter would not be passed into request if it's undefined.
-export type OpenAIVerbosity = OpenAI.Responses.ResponseTextConfig['verbosity']
+export type OpenAIVerbosity = NotNull<OpenAI.Responses.ResponseTextConfig['verbosity']>
 export type ValidOpenAIVerbosity = NotUndefined<OpenAIVerbosity>
 
 export type OpenAIReasoningEffort = OpenAI.ReasoningEffort
 
 // The original type unite both undefined and null.
+// I pick undefined as the unique falsy type since they seem like share the same meaning according to OpenAI API docs.
 // Parameter would not be passed into request if it's undefined.
-export type OpenAISummaryText = OpenAI.Reasoning['summary']
+export type OpenAISummaryText = NotNull<OpenAI.Reasoning['summary']>
 
 const AiSdkParamsSchema = z.enum([
   'maxOutputTokens',

src/renderer/src/types/sdk.ts

@@ -128,6 +128,10 @@ export type OpenAIExtraBody = {
     source_lang: 'auto'
     target_lang: string
   }
+  // for gpt-5 series models verbosity control
+  text?: {
+    verbosity?: 'low' | 'medium' | 'high'
+  }
 }
 // image is for openrouter. audio is ignored for now
 export type OpenAIModality = OpenAI.ChatCompletionModality | 'image'