Switch desktop execution to Sonnet 4.5

Author: jbutler1980

packages/bytebot-agent/src/proxy/proxy.service.spec.ts

@@ -252,6 +252,77 @@ describe('ProxyService endpoint failover', () => {
     expect(localCreate).toHaveBeenCalledTimes(0);
   });
 
+  it('treats claude-sonnet-4-5 as desktop-vision for endpoint ordering', async () => {
+    const eventEmitter = { emit: jest.fn() };
+    const llmResilienceService = makeResilience(eventEmitter);
+
+    const configService = {
+      get: jest.fn((key: string) => {
+        const map: Record<string, string> = {
+          BYTEBOT_LLM_PROXY_URL: 'http://local-proxy:4000',
+          BYTEBOT_LLM_PROXY_ENDPOINTS:
+            'http://local-proxy:4000,http://global-proxy:4000',
+          // Desktop-vision ordering is global-first
+          BYTEBOT_LLM_PROXY_DESKTOP_VISION_ENDPOINTS:
+            'http://global-proxy:4000,http://local-proxy:4000',
+          BYTEBOT_LLM_PROXY_API_KEY: 'dummy',
+          BYTEBOT_LLM_PROXY_ENDPOINT_PREFLIGHT_ENABLED: 'false',
+        };
+        return map[key] ?? '';
+      }),
+    } as any;
+
+    const localCreate = jest.fn(async () => {
+      const error = new Error('connect ECONNREFUSED 10.0.0.1:4000');
+      (error as any).code = 'ECONNREFUSED';
+      throw error;
+    });
+    const globalCreate = jest.fn(async () => {
+      return {
+        model: 'claude-sonnet-4-5',
+        choices: [{ message: { content: 'ok' } }],
+        usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+      };
+    });
+
+    class TestProxyService extends ProxyService {
+      protected override createOpenAIClient(baseURL: string): any {
+        if (baseURL.includes('local-proxy')) {
+          return { chat: { completions: { create: localCreate } } };
+        }
+        return { chat: { completions: { create: globalCreate } } };
+      }
+    }
+
+    const service = new TestProxyService(
+      configService,
+      llmResilienceService,
+      eventEmitter as any,
+    );
+
+    const messages = [
+      {
+        id: 'm1',
+        createdAt: new Date(),
+        updatedAt: new Date(),
+        taskId: 't1',
+        summaryId: null,
+        role: Role.USER,
+        content: [{ type: MessageContentType.Text, text: 'hello' }],
+      },
+    ] as any;
+
+    await service.generateMessage(
+      'system',
+      messages,
+      'claude-sonnet-4-5',
+      { useTools: false },
+    );
+
+    expect(globalCreate).toHaveBeenCalledTimes(1);
+    expect(localCreate).toHaveBeenCalledTimes(0);
+  });
+
   it('disables LiteLLM caching for desktop-vision model requests', async () => {
     const eventEmitter = { emit: jest.fn() };
     const llmResilienceService = makeResilience(eventEmitter);
@@ -310,6 +381,64 @@ describe('ProxyService endpoint failover', () => {
     expect(create.mock.calls[0][0].cache).toEqual({ 'no-cache': true });
   });
 
+  it('disables LiteLLM caching for claude-sonnet-4-5 desktop-vision requests', async () => {
+    const eventEmitter = { emit: jest.fn() };
+    const llmResilienceService = makeResilience(eventEmitter);
+
+    const configService = {
+      get: jest.fn((key: string) => {
+        const map: Record<string, string> = {
+          BYTEBOT_LLM_PROXY_URL: 'http://proxy:4000',
+          BYTEBOT_LLM_PROXY_ENDPOINTS: 'http://proxy:4000',
+          BYTEBOT_LLM_PROXY_DESKTOP_VISION_ENDPOINTS: 'http://proxy:4000',
+          BYTEBOT_LLM_PROXY_API_KEY: 'dummy',
+          BYTEBOT_LLM_PROXY_ENDPOINT_PREFLIGHT_ENABLED: 'false',
+        };
+        return map[key] ?? '';
+      }),
+    } as any;
+
+    const create = jest.fn(async (request: any) => {
+      return {
+        model: 'claude-sonnet-4-5',
+        choices: [{ message: { content: 'ok' } }],
+        usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+        __request: request,
+      };
+    });
+
+    class TestProxyService extends ProxyService {
+      protected override createOpenAIClient(): any {
+        return { chat: { completions: { create } } };
+      }
+    }
+
+    const service = new TestProxyService(
+      configService,
+      llmResilienceService,
+      eventEmitter as any,
+    );
+
+    const messages = [
+      {
+        id: 'm1',
+        createdAt: new Date(),
+        updatedAt: new Date(),
+        taskId: 't1',
+        summaryId: null,
+        role: Role.USER,
+        content: [{ type: MessageContentType.Text, text: 'hello' }],
+      },
+    ] as any;
+
+    await service.generateMessage('system', messages, 'claude-sonnet-4-5', {
+      useTools: false,
+    });
+
+    expect(create).toHaveBeenCalledTimes(1);
+    expect(create.mock.calls[0][0].cache).toEqual({ 'no-cache': true });
+  });
+
   it('does not replay Thinking blocks into Chat Completions history', async () => {
     const eventEmitter = { emit: jest.fn() };
     const llmResilienceService = makeResilience(eventEmitter);

packages/bytebot-agent/src/proxy/proxy.service.ts

@@ -325,8 +325,12 @@ export class ProxyService implements BytebotAgentService {
     const trimmed = (model || '').trim();
     return (
       trimmed === 'desktop-vision' ||
+      trimmed === 'claude-sonnet-4-5' ||
+      trimmed === 'claude-opus-4-5' ||
       trimmed === 'qwen3-vl-32b' ||
-      trimmed.endsWith('/qwen3-vl-32b')
+      trimmed.endsWith('/qwen3-vl-32b') ||
+      trimmed.endsWith('/claude-sonnet-4-5-20250929') ||
+      trimmed.endsWith('/claude-opus-4-5-20251101')
     );
   }
 

packages/bytebot-temporal-worker/src/activities/planning.activities.ts

@@ -13,7 +13,7 @@
  *
  * Model Flow (Phase 13):
  * - Planning (planGoal, refinePlan): Uses PLANNING_MODEL (default: gpt-oss-120b)
- * - Execution (via activities): Uses EXECUTION_MODEL (default: qwen3-vl-32b when available)
+ * - Execution (via activities): Uses EXECUTION_MODEL (default: claude-sonnet-4-5)
  * - Fallback: Uses FALLBACK_MODEL (default: claude-sonnet-4-5 - verified in LiteLLM)
  *
  * Agent Paradigm (Phase 13.3):
@@ -57,10 +57,10 @@ const LLM_API_KEY = process.env.LLM_API_KEY ?? process.env.OPENAI_API_KEY ?? '';
 const PLANNING_TIMEOUT_MS = parseInt(process.env.PLANNING_TIMEOUT_MS ?? '120000', 10);
 
 // Phase 13: In-house model configuration with proper fallback
-// Use gpt-oss-120b for planning (oversight), qwen3-vl-32b for execution (when available)
+// Use gpt-oss-120b for planning (oversight), claude-sonnet-4-5 for execution
 // Fallback uses claude-sonnet-4-5 which is configured in LiteLLM
 const PLANNING_MODEL = process.env.PLANNING_MODEL ?? 'gpt-oss-120b';
-const EXECUTION_MODEL = process.env.EXECUTION_MODEL ?? 'qwen3-vl-32b';
+const EXECUTION_MODEL = process.env.EXECUTION_MODEL ?? 'claude-sonnet-4-5';
 // Phase 13: Fixed fallback model - gpt-4 was not in LiteLLM config
 // Using claude-sonnet-4-5 as fallback (reliable, available in LiteLLM)
 const FALLBACK_MODEL = process.env.FALLBACK_MODEL ?? 'claude-sonnet-4-5';

packages/bytebot-workflow-orchestrator/src/services/task-dispatch.service.ts

@@ -49,9 +49,9 @@ import { hasDesktopExecutionTool } from '../contracts/planner-tools';
 // - gpt-oss-120b is running (1/1 replicas) with 131K context window
 // - 120B parameter model provides superior reasoning for browser automation
 //
-// Routes tasks to appropriate in-house models based on execution surface:
+// Routes tasks to appropriate in-house/external models based on execution surface:
 // - Browser tasks (requiresDesktop: false) → gpt-oss-120b (high-capability reasoning)
-// - Desktop tasks (requiresDesktop: true) → desktop-vision (vision-capable model group with fallbacks)
+// - Desktop tasks (requiresDesktop: true) → claude-sonnet-4-5 (vision-capable model)
 // - Fallback → claude-sonnet-4-5 (external API for complex reasoning)
 //
 // Model configuration for bytebot-agent's task schema
@@ -63,18 +63,18 @@ const BROWSER_TASK_MODEL = {
 };
 
 const DESKTOP_TASK_MODEL = {
-  // IMPORTANT: Use the model group alias (not a single backend) to ensure fallbacks are always enabled.
-  // See: docs/bytebot/contracts/VISION_MODEL_GROUPS.md
-  name: 'desktop-vision',
-  title: 'desktop-vision',
+  // Phase 15: Switch desktop execution to Sonnet 4.5 for reliability.
+  // Note: Proxy layer treats claude-sonnet-4-5 as a desktop-vision model for endpoint ordering + no-cache.
+  name: 'claude-sonnet-4-5',
+  title: 'claude-sonnet-4-5',
   provider: 'proxy',
-  contextWindow: 32000,
+  contextWindow: 128000,
 };
 
 // Fallback model for tasks that don't specify execution surface
 // or when in-house models fail
 const FALLBACK_TASK_MODEL = {
-  name: 'anthropic/claude-sonnet-4-5-20250929',
+  name: 'claude-sonnet-4-5',
   title: 'claude-sonnet-4-5',
   provider: 'proxy',
   contextWindow: 128000,
@@ -335,7 +335,7 @@ export class TaskDispatchService implements OnModuleInit {
    */
   private selectModelForTask(requiresDesktop?: boolean): typeof BROWSER_TASK_MODEL {
     if (requiresDesktop === true) {
-      this.logger.debug('Selected DESKTOP_TASK_MODEL (desktop-vision) for desktop task');
+      this.logger.debug('Selected DESKTOP_TASK_MODEL (claude-sonnet-4-5) for desktop task');
       return DESKTOP_TASK_MODEL;
     }