fix(core): disable qwen thinking via chat_template_kwargs on non-DashScope servers

packages/core/src/core/openaiContentGenerator/pipeline.test.ts

@@ -1108,6 +1108,141 @@ describe('ContentGenerationPipeline', () => {
       expect(apiCall.enable_thinking).toBeUndefined();
     });
 
+    it('disables qwen thinking via chat_template_kwargs on a non-DashScope endpoint (vLLM/SGLang)', async () => {
+      // Self-hosted OpenAI-compatible servers render the chat template
+      // server-side and read the thinking switch from `chat_template_kwargs`,
+      // silently ignoring a top-level `enable_thinking`. A qwen model on such
+      // an endpoint must therefore get the switch nested, not top-level — and
+      // any top-level `enable_thinking: true` a provider preset injected via
+      // extra_body must be stripped so it can't contradict the opt-out.
+      mockContentGeneratorConfig = {
+        ...mockContentGeneratorConfig,
+        baseUrl: 'https://llm.example.com/v1',
+        model: 'Qwen3.6-27B',
+      } as ContentGeneratorConfig;
+      mockConfig = {
+        ...mockConfig,
+        contentGeneratorConfig: mockContentGeneratorConfig,
+      };
+      pipeline = new ContentGenerationPipeline(mockConfig);
+
+      (mockProvider.buildRequest as Mock).mockImplementation((req) => ({
+        ...req,
+        enable_thinking: true, // Simulates extra_body injection
+      }));
+
+      const request: GenerateContentParameters = {
+        model: 'Qwen3.6-27B',
+        contents: [{ parts: [{ text: 'Suggest' }], role: 'user' }],
+        config: { thinkingConfig: { includeThoughts: false } },
+      };
+
+      (mockConverter.convertGeminiRequestToOpenAI as Mock).mockReturnValue([
+        { role: 'user', content: 'Suggest' },
+      ]);
+      (mockConverter.convertOpenAIResponseToGemini as Mock).mockReturnValue(
+        new GenerateContentResponse(),
+      );
+      (mockClient.chat.completions.create as Mock).mockResolvedValue({
+        id: 'r',
+        choices: [{ message: { content: 'ok' }, finish_reason: 'stop' }],
+      } as OpenAI.Chat.ChatCompletion);
+
+      await pipeline.execute(request, 'forked_query');
+
+      const apiCall = (mockClient.chat.completions.create as Mock).mock
+        .calls[0][0];
+      expect(apiCall.chat_template_kwargs).toEqual({ enable_thinking: false });
+      expect(apiCall.enable_thinking).toBeUndefined();
+    });
+
+    it('disables coder-model thinking via chat_template_kwargs on a non-DashScope endpoint', async () => {
+      // `coder-model` is the QWEN_OAUTH default, but a user can point it at a
+      // self-hosted endpoint. The `model === 'coder-model'` arm must reach the
+      // non-DashScope chat_template_kwargs path just like a `qwen*` model.
+      mockContentGeneratorConfig = {
+        ...mockContentGeneratorConfig,
+        baseUrl: 'https://llm.example.com/v1',
+        model: 'coder-model',
+      } as ContentGeneratorConfig;
+      mockConfig = {
+        ...mockConfig,
+        contentGeneratorConfig: mockContentGeneratorConfig,
+      };
+      pipeline = new ContentGenerationPipeline(mockConfig);
+
+      const request: GenerateContentParameters = {
+        model: 'coder-model',
+        contents: [{ parts: [{ text: 'Suggest' }], role: 'user' }],
+        config: { thinkingConfig: { includeThoughts: false } },
+      };
+
+      (mockConverter.convertGeminiRequestToOpenAI as Mock).mockReturnValue([
+        { role: 'user', content: 'Suggest' },
+      ]);
+      (mockConverter.convertOpenAIResponseToGemini as Mock).mockReturnValue(
+        new GenerateContentResponse(),
+      );
+      (mockClient.chat.completions.create as Mock).mockResolvedValue({
+        id: 'r',
+        choices: [{ message: { content: 'ok' }, finish_reason: 'stop' }],
+      } as OpenAI.Chat.ChatCompletion);
+
+      await pipeline.execute(request, 'forked_query');
+
+      const apiCall = (mockClient.chat.completions.create as Mock).mock
+        .calls[0][0];
+      expect(apiCall.chat_template_kwargs).toEqual({ enable_thinking: false });
+      expect(apiCall.enable_thinking).toBeUndefined();
+    });
+
+    it('merges enable_thinking into pre-existing chat_template_kwargs on a non-DashScope endpoint', async () => {
+      // The non-DashScope path spreads any existing `chat_template_kwargs`
+      // before appending `enable_thinking: false`. Guard the merge so a future
+      // refactor can't silently drop user-configured kwargs.
+      mockContentGeneratorConfig = {
+        ...mockContentGeneratorConfig,
+        baseUrl: 'https://llm.example.com/v1',
+        model: 'Qwen3.6-27B',
+      } as ContentGeneratorConfig;
+      mockConfig = {
+        ...mockConfig,
+        contentGeneratorConfig: mockContentGeneratorConfig,
+      };
+      pipeline = new ContentGenerationPipeline(mockConfig);
+
+      (mockProvider.buildRequest as Mock).mockImplementation((req) => ({
+        ...req,
+        chat_template_kwargs: { apply_chat_template: true },
+      }));
+
+      const request: GenerateContentParameters = {
+        model: 'Qwen3.6-27B',
+        contents: [{ parts: [{ text: 'Suggest' }], role: 'user' }],
+        config: { thinkingConfig: { includeThoughts: false } },
+      };
+
+      (mockConverter.convertGeminiRequestToOpenAI as Mock).mockReturnValue([
+        { role: 'user', content: 'Suggest' },
+      ]);
+      (mockConverter.convertOpenAIResponseToGemini as Mock).mockReturnValue(
+        new GenerateContentResponse(),
+      );
+      (mockClient.chat.completions.create as Mock).mockResolvedValue({
+        id: 'r',
+        choices: [{ message: { content: 'ok' }, finish_reason: 'stop' }],
+      } as OpenAI.Chat.ChatCompletion);
+
+      await pipeline.execute(request, 'forked_query');
+
+      const apiCall = (mockClient.chat.completions.create as Mock).mock
+        .calls[0][0];
+      expect(apiCall.chat_template_kwargs).toEqual({
+        apply_chat_template: true,
+        enable_thinking: false,
+      });
+    });
+
     it('does NOT emit enable_thinking on a non-qwen model routed through DashScope', async () => {
       // DashScope's compatible-mode endpoint routes multiple model families
       // (qwen3, GLM, DeepSeek). Hostname alone is not enough — GLM uses

packages/core/src/core/openaiContentGenerator/pipeline.ts

@@ -584,13 +584,39 @@ export class ContentGenerationPipeline {
       // start with `qwen` but is the most common hybrid-thinking model
       // for first-time users, so it must be covered.
       const model = (context.model ?? '').toLowerCase();
-      if (
-        DashScopeOpenAICompatibleProvider.isDashScopeProvider(
-          this.contentGeneratorConfig,
-        ) &&
-        (model.startsWith('qwen') || model === 'coder-model')
-      ) {
-        typed['enable_thinking'] = false;
+      if (model.startsWith('qwen') || model === 'coder-model') {
+        if (
+          DashScopeOpenAICompatibleProvider.isDashScopeProvider(
+            this.contentGeneratorConfig,
+          )
+        ) {
+          typed['enable_thinking'] = false;
+        } else {
+          // Non-DashScope OpenAI-compatible servers (vLLM, SGLang, ...) render
+          // the model's chat template server-side and read the thinking switch
+          // from `chat_template_kwargs`, not a top-level `enable_thinking`
+          // (which they silently ignore). Send it there so hybrid qwen models
+          // actually stop emitting <think> when reasoning is disabled — e.g.
+          // the auto-mode permission classifier's short structured-output
+          // calls, which otherwise spend their small token budget on thinking
+          // and fail closed. Servers that don't recognise `chat_template_kwargs`
+          // ignore the unknown field, so the switch is a harmless no-op there.
+          //
+          // Drop any top-level `enable_thinking` a provider preset injected via
+          // extra_body (provider-config.ts emits it for models configured with
+          // `enableThinking: true`): leaving it would contradict the
+          // `chat_template_kwargs` opt-out on servers that honour both, and
+          // keeps this path from leaking the qwen-specific field top-level.
+          delete typed['enable_thinking'];
+          const existing = (typed['chat_template_kwargs'] ?? {}) as Record<
+            string,
+            unknown
+          >;
+          typed['chat_template_kwargs'] = {
+            ...existing,
+            enable_thinking: false,
+          };
+        }
       }
       // Strip reasoning config — extra_body could inject it, overriding
       // buildReasoningConfig's decision to return {} for disabled thinking.