Fix Qwen3 thinking mode control for Ollama using aiCore middleware (#10947)

* Initial plan

* Fix Qwen3 thinking mode control for Ollama by using reasoning_effort instead of qwenThinkMode

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

* Refactor: Move Qwen thinking mode control from legacy to aiCore middleware

- Revert changes to legacy OpenAIApiClient.ts
- Create new qwenThinkingMiddleware for handling /think and /no_think suffixes
- Update AiSdkMiddlewareBuilder to add assistant field and apply Qwen middleware
- Middleware automatically handles Qwen models on Ollama, LM Studio, and NVIDIA providers
- This follows the modern aiCore architecture instead of deprecated legacy approach

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

* Fix: Explicitly pass assistant field to buildAiSdkMiddlewares

Ensure that the assistant field from ModernAiProviderConfig is explicitly passed
to buildAiSdkMiddlewares so that Qwen thinking middleware can access
assistant.settings.reasoning_effort correctly.

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

* Fix: Remove unnecessary whitespace in processPostsuffixQwen3Model tests

* Refactor: Simplify user message suffix handling in qwenThinkingMiddleware

* Refactor: Remove processPostsuffixQwen3Model tests to streamline ModelMessageService tests

* refactor: remove logger and debug statement from qwenThinkingMiddleware

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>
Co-authored-by: suyao <sy20010504@gmail.com>

src/renderer/src/aiCore/index_new.ts

@@ -97,7 +97,8 @@ export default class ModernAiProvider {
     // 提前构建中间件
     const middlewares = buildAiSdkMiddlewares({
       ...config,
-      provider: this.actualProvider
+      provider: this.actualProvider,
+      assistant: config.assistant
     })
     logger.debug('Built middlewares in completions', {
       middlewareCount: middlewares.length,

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

@@ -1,12 +1,15 @@
 import { WebSearchPluginConfig } from '@cherrystudio/ai-core/built-in/plugins'
 import { loggerService } from '@logger'
-import { type MCPTool, type Message, type Model, type Provider } from '@renderer/types'
+import { isSupportedThinkingTokenQwenModel } from '@renderer/config/models'
+import { isSupportEnableThinkingProvider } from '@renderer/config/providers'
+import { type Assistant, MCPTool, type Message, type Model, type Provider } from '@renderer/types'
 import type { Chunk } from '@renderer/types/chunk'
 import { extractReasoningMiddleware, LanguageModelMiddleware, simulateStreamingMiddleware } from 'ai'
 
 import { isOpenRouterGeminiGenerateImageModel } from '../utils/image'
 import { noThinkMiddleware } from './noThinkMiddleware'
 import { openrouterGenerateImageMiddleware } from './openrouterGenerateImageMiddleware'
+import { qwenThinkingMiddleware } from './qwenThinkingMiddleware'
 import { toolChoiceMiddleware } from './toolChoiceMiddleware'
 
 const logger = loggerService.withContext('AiSdkMiddlewareBuilder')
@@ -19,6 +22,7 @@ export interface AiSdkMiddlewareConfig {
   onChunk?: (chunk: Chunk) => void
   model?: Model
   provider?: Provider
+  assistant?: Assistant
   enableReasoning: boolean
   // 是否开启提示词工具调用
   isPromptToolUse: boolean
@@ -218,6 +222,21 @@ function addProviderSpecificMiddlewares(builder: AiSdkMiddlewareBuilder, config:
 function addModelSpecificMiddlewares(builder: AiSdkMiddlewareBuilder, config: AiSdkMiddlewareConfig): void {
   if (!config.model || !config.provider) return
 
+  // Qwen models on providers that don't support enable_thinking parameter (like Ollama, LM Studio, NVIDIA)
+  // Use /think or /no_think suffix to control thinking mode
+  if (
+    config.provider &&
+    isSupportedThinkingTokenQwenModel(config.model) &&
+    !isSupportEnableThinkingProvider(config.provider)
+  ) {
+    const enableThinking = config.assistant?.settings?.reasoning_effort !== undefined
+    builder.add({
+      name: 'qwen-thinking-control',
+      middleware: qwenThinkingMiddleware(enableThinking)
+    })
+    logger.debug(`Added Qwen thinking middleware with thinking ${enableThinking ? 'enabled' : 'disabled'}`)
+  }
+
   // 可以根据模型ID或特性添加特定中间件
   // 例如：图像生成模型、多模态模型等
   if (isOpenRouterGeminiGenerateImageModel(config.model, config.provider)) {

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

@@ -0,0 +1,39 @@
+import { LanguageModelMiddleware } from 'ai'
+
+/**
+ * Qwen Thinking Middleware
+ * Controls thinking mode for Qwen models on providers that don't support enable_thinking parameter (like Ollama)
+ * Appends '/think' or '/no_think' suffix to user messages based on reasoning_effort setting
+ * @param enableThinking - Whether thinking mode is enabled (based on reasoning_effort !== undefined)
+ * @returns LanguageModelMiddleware
+ */
+export function qwenThinkingMiddleware(enableThinking: boolean): LanguageModelMiddleware {
+  const suffix = enableThinking ? ' /think' : ' /no_think'
+
+  return {
+    middlewareVersion: 'v2',
+
+    transformParams: async ({ params }) => {
+      const transformedParams = { ...params }
+      // Process messages in prompt
+      if (transformedParams.prompt && Array.isArray(transformedParams.prompt)) {
+        transformedParams.prompt = transformedParams.prompt.map((message) => {
+          // Only process user messages
+          if (message.role === 'user') {
+            // 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
+                }
+              }
+            }
+          }
+          return message
+        })
+      }
+
+      return transformedParams
+    }
+  }
+}