Add preliminary assistant model

package-lock.json

@@ -12,6 +12,7 @@
       ],
       "dependencies": {
         "@lvce-editor/ripgrep": "^1.6.0",
+        "@qwen-code/qwen-code": "^0.0.14",
         "simple-git": "^3.28.0",
         "strip-ansi": "^7.1.0"
       },
@@ -19,6 +20,7 @@
         "qwen": "bundle/gemini.js"
       },
       "devDependencies": {
+        "@google/genai": "^1.22.0",
         "@types/marked": "^5.0.2",
         "@types/mime-types": "^3.0.1",
         "@types/minimatch": "^5.1.2",
@@ -998,9 +1000,10 @@
       "link": true
     },
     "node_modules/@google/genai": {
-      "version": "1.13.0",
+      "version": "1.22.0",
-      "resolved": "https://registry.npmjs.org/@google/genai/-/genai-1.13.0.tgz",
+      "resolved": "https://registry.npmjs.org/@google/genai/-/genai-1.22.0.tgz",
-      "integrity": "sha512-BxilXzE8cJ0zt5/lXk6KwuBcIT9P2Lbi2WXhwWMbxf1RNeC68/8DmYQqMrzQP333CieRMdbDXs0eNCphLoScWg==",
+      "integrity": "sha512-siETS3zTm3EGpTT4+BFc1z20xXBYfueD3gCYfxkOjuAKRk8lt8TJevDHi3zepn1oSI6NhG/LZvy0i+Q3qheObg==",
+      "dev": true,
       "license": "Apache-2.0",
       "dependencies": {
         "google-auth-library": "^9.14.2",
@@ -1010,7 +1013,7 @@
         "node": ">=20.0.0"
       },
       "peerDependencies": {
-        "@modelcontextprotocol/sdk": "^1.11.0"
+        "@modelcontextprotocol/sdk": "^1.11.4"
       },
       "peerDependenciesMeta": {
         "@modelcontextprotocol/sdk": {
@@ -13731,6 +13734,27 @@
         "node-pty": "^1.0.0"
       }
     },
+    "packages/core/node_modules/@google/genai": {
+      "version": "1.13.0",
+      "resolved": "https://registry.npmjs.org/@google/genai/-/genai-1.13.0.tgz",
+      "integrity": "sha512-BxilXzE8cJ0zt5/lXk6KwuBcIT9P2Lbi2WXhwWMbxf1RNeC68/8DmYQqMrzQP333CieRMdbDXs0eNCphLoScWg==",
+      "license": "Apache-2.0",
+      "dependencies": {
+        "google-auth-library": "^9.14.2",
+        "ws": "^8.18.0"
+      },
+      "engines": {
+        "node": ">=20.0.0"
+      },
+      "peerDependencies": {
+        "@modelcontextprotocol/sdk": "^1.11.0"
+      },
+      "peerDependenciesMeta": {
+        "@modelcontextprotocol/sdk": {
+          "optional": true
+        }
+      }
+    },
     "packages/core/node_modules/ajv": {
       "version": "8.17.1",
       "resolved": "https://registry.npmjs.org/ajv/-/ajv-8.17.1.tgz",

package.json

@@ -61,6 +61,7 @@
     "LICENSE"
   ],
   "devDependencies": {
+    "@google/genai": "^1.22.0",
     "@types/marked": "^5.0.2",
     "@types/mime-types": "^3.0.1",
     "@types/minimatch": "^5.1.2",
@@ -96,6 +97,7 @@
   },
   "dependencies": {
     "@lvce-editor/ripgrep": "^1.6.0",
+    "@qwen-code/qwen-code": "^0.0.14",
     "simple-git": "^3.28.0",
     "strip-ansi": "^7.1.0"
   },

packages/core/src/core/client.ts

@@ -703,7 +703,7 @@ export class GeminiClient {
      * We should ignore model for now because some calls use `DEFAULT_GEMINI_FLASH_MODEL`
      * which is not available as `qwen3-coder-flash`
      */
-    const modelToUse = this.config.getModel() || DEFAULT_GEMINI_FLASH_MODEL;
+    const modelToUse = model || this.config.getModel() || DEFAULT_GEMINI_FLASH_MODEL;
     try {
       const userMemory = this.config.getUserMemory();
       const finalSystemInstruction = config.systemInstruction
@@ -789,7 +789,7 @@ export class GeminiClient {
     abortSignal: AbortSignal,
     model?: string,
   ): Promise<GenerateContentResponse> {
-    const modelToUse = model ?? this.config.getModel();
+    const modelToUse = model || this.config.getModel();
     const configToUse: GenerateContentConfig = {
       ...this.generateContentConfig,
       ...generationConfig,

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

@@ -241,10 +241,17 @@ export class ContentGenerationPipeline {
   ): Promise<OpenAI.Chat.ChatCompletionCreateParams> {
     const messages = this.converter.convertGeminiRequestToOpenAI(request);
 
+    // Check if request has a model override
+    let model = this.contentGeneratorConfig.model;
+    if (request.model) {
+      // Use request-specific model if provided
+      model = request.model;
+    }
+
     // Apply provider-specific enhancements
     const baseRequest: OpenAI.Chat.ChatCompletionCreateParams = {
-      model: this.contentGeneratorConfig.model,
       messages,
+      model,
       ...this.buildSamplingParameters(request),
     };
 
@@ -405,9 +412,13 @@ export class ContentGenerationPipeline {
     userPromptId: string,
     isStreaming: boolean,
   ): RequestContext {
+    // For context logging, we use the default model since we don't have access to a request
+    // This is acceptable since context logging doesn't need request-specific model
+    const model = this.contentGeneratorConfig.model;
+    
     return {
       userPromptId,
-      model: this.contentGeneratorConfig.model,
+      model,
       authType: this.contentGeneratorConfig.authType || 'unknown',
       startTime: Date.now(),
       duration: 0,

packages/core/src/services/loopDetectionService.ts

@@ -396,9 +396,18 @@ Please analyze the conversation history to determine the possibility that the co
     };
     let result;
     try {
+      // Get the assistant model from environment variable, falling back to default if not specified
+      let assistantModel = DEFAULT_QWEN_FLASH_MODEL;
+      
+      // Check for OPENAI_ASSISTANT_MODEL environment variable
+      const openaiAssistantModel = process.env['OPENAI_ASSISTANT_MODEL'];
+      if (openaiAssistantModel) {
+        assistantModel = openaiAssistantModel;
+      }
+
       result = await this.config
         .getGeminiClient()
-        .generateJson(contents, schema, signal, DEFAULT_QWEN_FLASH_MODEL);
+        .generateJson(contents, schema, signal, assistantModel);
     } catch (e) {
       // Do nothing, treat it as a non-loop.
       this.config.getDebugMode() ? console.error(e) : console.debug(e);

packages/core/src/tools/web-fetch.ts

@@ -8,6 +8,7 @@ import { convert } from 'html-to-text';
 import { ProxyAgent, setGlobalDispatcher } from 'undici';
 import type { Config } from '../config/config.js';
 import { ApprovalMode } from '../config/config.js';
+import { DEFAULT_QWEN_FLASH_MODEL } from '../config/models.js';
 import { fetchWithTimeout, isPrivateIp } from '../utils/fetch.js';
 import { getResponseText } from '../utils/partUtils.js';
 import { ToolErrorType } from './tool-error.js';
@@ -104,10 +105,22 @@ ${textContent}
         `[WebFetchTool] Processing content with prompt: "${this.params.prompt}"`,
       );
 
+      // For assistant decision-making, we want to use a different model than the one used for regular conversation
+      // This helps avoid KV cache invalidation issues when using the same model for both purposes
+      
+      // Get the assistant model from environment variable, falling back to default if not specified
+      let assistantModel = DEFAULT_QWEN_FLASH_MODEL;
+      
+      // Check for OPENAI_ASSISTANT_MODEL environment variable
+      const openaiAssistantModel = process.env['OPENAI_ASSISTANT_MODEL'];
+      if (openaiAssistantModel) {
+        assistantModel = openaiAssistantModel;
+      }
       const result = await geminiClient.generateContent(
         [{ role: 'user', parts: [{ text: fallbackPrompt }] }],
         {},
         signal,
+        assistantModel
       );
       const resultText = getResponseText(result) || '';
 

packages/core/src/utils/nextSpeakerChecker.test.ts

@@ -242,7 +242,7 @@ describe('checkNextSpeaker', () => {
     expect(result).toBeNull();
   });
 
-  it('should call generateJson with DEFAULT_QWEN_FLASH_MODEL', async () => {
+  it('should call generateJson with DEFAULT_QWEN_FLASH_MODEL when OPENAI_ASSISTANT_MODEL is not set', async () => {
     (chatInstance.getHistory as Mock).mockReturnValue([
       { role: 'model', parts: [{ text: 'Some model output.' }] },
     ] as Content[]);
@@ -259,4 +259,31 @@ describe('checkNextSpeaker', () => {
       .calls[0];
     expect(generateJsonCall[3]).toBe(DEFAULT_QWEN_FLASH_MODEL);
   });
+
+  it('should call generateJson with OPENAI_ASSISTANT_MODEL when set', async () => {
+    // Mock the environment variable
+    const originalEnv = process.env['OPENAI_ASSISTANT_MODEL'];
+    process.env['OPENAI_ASSISTANT_MODEL'] = 'qwen2.5-0.5b';
+    
+    try {
+      (chatInstance.getHistory as Mock).mockReturnValue([
+        { role: 'model', parts: [{ text: 'Some model output.' }] },
+      ] as Content[]);
+      const mockApiResponse: NextSpeakerResponse = {
+        reasoning: 'Model made a statement, awaiting user input.',
+        next_speaker: 'user',
+      };
+      (mockGeminiClient.generateJson as Mock).mockResolvedValue(mockApiResponse);
+
+      await checkNextSpeaker(chatInstance, mockGeminiClient, abortSignal);
+
+      expect(mockGeminiClient.generateJson).toHaveBeenCalled();
+      const generateJsonCall = (mockGeminiClient.generateJson as Mock).mock
+        .calls[0];
+      expect(generateJsonCall[3]).toBe('qwen2.5-0.5b');
+    } finally {
+      // Restore original environment variable
+      process.env['OPENAI_ASSISTANT_MODEL'] = originalEnv;
+    }
+  });
 });

packages/core/src/utils/nextSpeakerChecker.ts

@@ -108,11 +108,23 @@ export async function checkNextSpeaker(
   ];
 
   try {
+    // For assistant decision-making, we want to use a different model than the one used for regular conversation
+    // This helps avoid KV cache invalidation issues when using the same model for both purposes
+    
+    // Get the assistant model from environment variable, falling back to default if not specified
+    let assistantModel = DEFAULT_QWEN_FLASH_MODEL;
+    
+    // Check for OPENAI_ASSISTANT_MODEL environment variable
+    const openaiAssistantModel = process.env['OPENAI_ASSISTANT_MODEL'];
+    if (openaiAssistantModel) {
+      assistantModel = openaiAssistantModel;
+    }
+
     const parsedResponse = (await geminiClient.generateJson(
       contents,
       RESPONSE_SCHEMA,
       abortSignal,
-      DEFAULT_QWEN_FLASH_MODEL,
+      assistantModel,
     )) as unknown as NextSpeakerResponse;
 
     if (

packages/core/src/utils/summarizer.ts

@@ -11,7 +11,7 @@ import type {
   GenerateContentResponse,
 } from '@google/genai';
 import type { GeminiClient } from '../core/client.js';
-import { DEFAULT_GEMINI_FLASH_LITE_MODEL } from '../config/models.js';
+import { DEFAULT_QWEN_FLASH_MODEL } from '../config/models.js';
 import { getResponseText, partToString } from './partUtils.js';
 
 /**
@@ -81,12 +81,24 @@ export async function summarizeToolOutput(
   const toolOutputSummarizerConfig: GenerateContentConfig = {
     maxOutputTokens,
   };
+  // For assistant decision-making, we want to use a different model than the one used for regular conversation
+  // This helps avoid KV cache invalidation issues when using the same model for both purposes
+  
+  // Get the assistant model from environment variable, falling back to default if not specified
+  let assistantModel = DEFAULT_QWEN_FLASH_MODEL;
+  
+  // Check for OPENAI_ASSISTANT_MODEL environment variable
+  const openaiAssistantModel = process.env['OPENAI_ASSISTANT_MODEL'];
+  if (openaiAssistantModel) {
+    assistantModel = openaiAssistantModel;
+  }
+
   try {
     const parsedResponse = (await geminiClient.generateContent(
       contents,
       toolOutputSummarizerConfig,
       abortSignal,
-      DEFAULT_GEMINI_FLASH_LITE_MODEL,
+      assistantModel,
     )) as unknown as GenerateContentResponse;
     return getResponseText(parsedResponse) || textToSummarize;
   } catch (error) {