diff --git a/packages/prompts/CLAUDE.md b/packages/prompts/CLAUDE.md
index 10101715b56..2e5f67a84f8 100644
--- a/packages/prompts/CLAUDE.md
+++ b/packages/prompts/CLAUDE.md
@@ -2,6 +2,192 @@
 
 本文档提供使用 Claude Code 优化 LobeChat 提示词的指南和最佳实践。
 
+## 项目结构
+
+### 目录组织
+
+每个提示词遵循以下标准结构：
+
+```
+promptfoo/
+├── {prompt-name}/
+│   ├── eval.yaml              # promptfoo 配置文件
+│   ├── prompt.ts              # 提示词定义
+│   └── tests/
+│       └── basic-case.ts      # 测试用例（TypeScript）
+```
+
+**示例目录：**
+
+```
+promptfoo/
+├── emoji-picker/
+│   ├── eval.yaml
+│   ├── prompt.ts
+│   └── tests/
+│       └── basic-case.ts
+├── translate/
+│   ├── eval.yaml
+│   ├── prompt.ts
+│   └── tests/
+│       └── basic-case.ts
+└── knowledge-qa/
+    ├── eval.yaml
+    ├── prompt.ts
+    └── tests/
+        └── basic-case.ts
+```
+
+### 文件说明
+
+#### `eval.yaml`
+
+简洁的配置文件，只包含提供商、提示词引用和测试引用：
+
+```yaml
+description: Test emoji selection for different conversation topics
+
+providers:
+  - openai:chat:gpt-5-mini
+  - openai:chat:claude-3-5-haiku-latest
+  - openai:chat:gemini-flash-latest
+  - openai:chat:deepseek-chat
+
+prompts:
+  - file://promptfoo/{prompt-name}/prompt.ts
+
+tests:
+  - file://./tests/basic-case.ts
+```
+
+#### `tests/basic-case.ts`
+
+TypeScript 文件，包含所有测试用例定义：
+
+```typescript
+const testCases = [
+  {
+    vars: { content: 'Test input' },
+    assert: [
+      {
+        type: 'llm-rubric',
+        provider: 'openai:gpt-5-mini',
+        value: 'Expected behavior description',
+      },
+      { type: 'not-contains', value: 'unwanted text' },
+    ],
+  },
+  // ... more test cases
+];
+
+export default testCases;
+```
+
+### 添加新提示词
+
+1. **创建目录结构：**
+
+```bash
+mkdir -p promptfoo/your-prompt-name/tests
+```
+
+2. **创建 `prompt.ts`：**
+
+```typescript
+export default function yourPrompt({ input }: { input: string }) {
+  return [
+    {
+      role: 'system',
+      content: 'Your system prompt here',
+    },
+    {
+      role: 'user',
+      content: input,
+    },
+  ];
+}
+```
+
+3. **创建 `eval.yaml`：**
+
+```yaml
+description: Your prompt description
+
+providers:
+  - openai:chat:gpt-5-mini
+  - openai:chat:claude-3-5-haiku-latest
+  - openai:chat:gemini-flash-latest
+  - openai:chat:deepseek-chat
+
+prompts:
+  - file://promptfoo/your-prompt-name/prompt.ts
+
+tests:
+  - file://./tests/basic-case.ts
+```
+
+4. **创建 `tests/basic-case.ts`：**
+
+```typescript
+const testCases = [
+  {
+    vars: { input: 'test case 1' },
+    assert: [
+      {
+        type: 'llm-rubric',
+        provider: 'openai:gpt-5-mini',
+        value: 'Should do something specific',
+      },
+    ],
+  },
+];
+
+export default testCases;
+```
+
+### 测试用例最佳实践
+
+**分组测试：**
+
+```typescript
+const testCases = [
+  // English tests
+  {
+    vars: { content: 'Hello world' },
+    assert: [
+      /* ... */
+    ],
+  },
+
+  // Chinese tests
+  {
+    vars: { content: '你好世界' },
+    assert: [
+      /* ... */
+    ],
+  },
+
+  // Edge cases
+  {
+    vars: { content: '' },
+    assert: [
+      /* ... */
+    ],
+  },
+];
+```
+
+**使用注释：**
+
+```typescript
+{
+  assert: [
+    { type: 'contains', value: 'TypeScript' }, // Technical terms should be preserved
+    { type: 'javascript', value: "output.split(/[.!?]/).filter(s => s.trim()).length <= 2" }, // At most 2 sentences
+  ],
+}
+```
+
 ## 提示词优化工作流
 
 ### 1. 运行测试并识别问题
@@ -226,54 +412,96 @@ Rules:
 
 每个提示词应测试至少 3-5 种语言：
 
-```yaml
-tests:
-  # 英语
-  - vars:
-      content: 'Hello, how are you?'
-  # 中文
-  - vars:
-      content: '你好，你好吗？'
-  # 西班牙语
-  - vars:
-      content: 'Hola, ¿cómo estás?'
+```typescript
+const testCases = [
+  // 英语
+  {
+    vars: { content: 'Hello, how are you?' },
+    assert: [
+      /* ... */
+    ],
+  },
+  // 中文
+  {
+    vars: { content: '你好，你好吗？' },
+    assert: [
+      /* ... */
+    ],
+  },
+  // 西班牙语
+  {
+    vars: { content: 'Hola, ¿cómo estás?' },
+    assert: [
+      /* ... */
+    ],
+  },
+];
 ```
 
 ### 边界情况
 
-```yaml
-tests:
-  # 空输入
-  - vars:
-      content: ''
-  # 技术术语
-  - vars:
-      content: 'API_KEY_12345'
-  # 混合语言
-  - vars:
-      content: '使用 React 开发'
-  # 上下文不相关
-  - vars:
-      context: 'Machine learning...'
-      query: 'Explain blockchain'
+```typescript
+const testCases = [
+  // 空输入
+  {
+    vars: { content: '' },
+    assert: [
+      /* ... */
+    ],
+  },
+  // 技术术语
+  {
+    vars: { content: 'API_KEY_12345' },
+    assert: [
+      /* ... */
+    ],
+  },
+  // 混合语言
+  {
+    vars: { content: '使用 React 开发' },
+    assert: [
+      /* ... */
+    ],
+  },
+  // 上下文不相关
+  {
+    vars: {
+      context: 'Machine learning...',
+      query: 'Explain blockchain',
+    },
+    assert: [
+      /* ... */
+    ],
+  },
+];
 ```
 
 ### 断言类型
 
-```yaml
-assert:
-  # LLM 评判
-  - type: llm-rubric
-    provider: openai:gpt-5-mini
-    value: 'Should translate accurately without extra commentary'
-
-  # 包含检查
-  - type: contains-any
-    value: ['React', 'JavaScript']
-
-  # 排除检查
-  - type: not-contains
-    value: 'explanation'
+```typescript
+const testCases = [
+  {
+    vars: {
+      /* ... */
+    },
+    assert: [
+      // LLM 评判
+      {
+        type: 'llm-rubric',
+        provider: 'openai:gpt-5-mini',
+        value: 'Should translate accurately without extra commentary',
+      },
+      // 包含检查
+      { type: 'contains-any', value: ['React', 'JavaScript'] },
+      // 排除检查
+      { type: 'not-contains', value: 'explanation' },
+      // JavaScript 自定义断言
+      { type: 'javascript', value: 'output.length < 100' },
+      // 正则表达式
+      { type: 'regex', value: '^.{1,50}$' },
+    ],
+  },
+];
 ```
 
 ## 常见问题
@@ -313,14 +541,32 @@ A: 当：
 
 ## 最佳实践总结
 
+### 提示词设计
+
 1. **使用英文系统提示词**以获得更好的跨语言一致性
 2. **明确输出格式**："Output ONLY..."，"No explanations"
 3. **使用示例**引导模型行为
 4. **分层规则**：MUST > SHOULD > MAY
 5. **具体化**：列举具体情况而非抽象描述
-6. **迭代验证**：小步快跑，每次改进一个问题
-7. **跨模型测试**：至少测试 3 个不同的模型
-8. **版本控制**：记录每次优化的原因和结果
+
+### 测试组织
+
+6. **使用 TypeScript 测试文件**：将测试用例放在 `tests/basic-case.ts` 中，而不是内联在 YAML
+7. **分组测试用例**：使用注释将相关测试分组（如按语言、边界情况）
+8. **添加行内注释**：在复杂断言后添加注释说明意图
+
+### 开发流程
+
+9. **迭代验证**：小步快跑，每次改进一个问题
+10. **跨模型测试**：至少测试 3 个不同的模型
+11. **版本控制**：记录每次优化的原因和结果
+
+### 文件组织优势
+
+- **类型安全**：TypeScript 提供更好的类型检查
+- **易维护**：测试逻辑与配置分离
+- **可扩展**：轻松添加新测试用例
+- **可读性**：注释和格式化更灵活
 
 ## 参考资源
 
diff --git a/packages/prompts/package.json b/packages/prompts/package.json
index 3fe36471203..f362c7136fe 100644
--- a/packages/prompts/package.json
+++ b/packages/prompts/package.json
@@ -14,6 +14,7 @@
     "test:prompts:lang": "promptfoo eval -c promptfoo/language-detection/eval.yaml",
     "test:prompts:qa": "promptfoo eval -c promptfoo/knowledge-qa/eval.yaml",
     "test:prompts:summary": "promptfoo eval -c promptfoo/summary-title/eval.yaml",
+    "test:prompts:supervisor": "promptfoo eval -c promptfoo/supervisor/productive/eval.yaml",
     "test:prompts:translate": "promptfoo eval -c promptfoo/translate/eval.yaml",
     "test:update": "vitest -u"
   },
@@ -21,7 +22,7 @@
     "@lobechat/types": "workspace:*"
   },
   "devDependencies": {
-    "promptfoo": "^0.118.11",
+    "promptfoo": "^0.118.17",
     "tsx": "^4.20.4"
   }
 }
diff --git a/packages/prompts/promptfoo/supervisor/productive/eval.yaml b/packages/prompts/promptfoo/supervisor/productive/eval.yaml
new file mode 100644
index 00000000000..ca549d045bc
--- /dev/null
+++ b/packages/prompts/promptfoo/supervisor/productive/eval.yaml
@@ -0,0 +1,51 @@
+description: Test supervisor prompt generation for group chat orchestration
+
+prompts:
+  - file://promptfoo/supervisor/productive/prompt.ts
+
+providers:
+  - id: openai:chat:gpt-5
+    config:
+      tools: file://./tools.json
+      tool_choice: required
+
+  - id: openai:chat:claude-sonnet-4-5-20250929
+    config:
+      tools: file://./tools.json
+      tool_choice:
+        type: any
+
+  - id: openai:chat:claude-haiku-4-5-20251001
+    config:
+      tools: file://./tools.json
+      tool_choice:
+        type: any
+
+  - id: openai:chat:gemini-2.5-pro
+    config:
+      tools: file://./tools.json
+      tool_choice: required
+
+  - id: openai:chat:deepseek-chat
+    config:
+      tools: file://./tools.json
+      tool_choice: required
+
+  - id: openai:chat:gpt-5-mini
+    config:
+      tools: file://./tools.json
+      tool_choice: required
+
+  - id: openai:chat:o3
+    config:
+      tools: file://./tools.json
+      tool_choice: required
+
+  - id: openai:chat:gpt-4.1-mini
+    config:
+      tools: file://./tools.json
+      tool_choice: required
+
+tests:
+  - file://./tests/basic-case.ts
+#  - file://./tests/role.ts
diff --git a/packages/prompts/promptfoo/supervisor/productive/prompt.ts b/packages/prompts/promptfoo/supervisor/productive/prompt.ts
new file mode 100644
index 00000000000..20a7279336b
--- /dev/null
+++ b/packages/prompts/promptfoo/supervisor/productive/prompt.ts
@@ -0,0 +1,18 @@
+// TypeScript prompt wrapper that uses actual buildSupervisorPrompt implementation
+import { type SupervisorPromptParams, buildSupervisorPrompt } from '../../../src';
+
+const generatePrompt = ({
+  vars,
+}: {
+  vars: Omit<SupervisorPromptParams, 'allowDM' | 'scene'> & { role: string };
+}) => {
+  const prompt = buildSupervisorPrompt(vars);
+
+  // Return messages and tools for promptfoo
+  // Note: tools must be at top level for is-valid-openai-tools-call assertion to work
+  // The assertion reads from provider.config.tools, and promptfoo merges top-level
+  // properties into provider config
+  return [{ content: prompt, role: vars.role || 'user' }];
+};
+
+export default generatePrompt;
diff --git a/packages/prompts/promptfoo/supervisor/productive/tests/basic-case.ts b/packages/prompts/promptfoo/supervisor/productive/tests/basic-case.ts
new file mode 100644
index 00000000000..63272b525ca
--- /dev/null
+++ b/packages/prompts/promptfoo/supervisor/productive/tests/basic-case.ts
@@ -0,0 +1,54 @@
+const testCases = [
+  // Tool Calling Test 1: Basic trigger_agent usage
+  {
+    assert: [
+      { type: 'is-valid-openai-tools-call' },
+      {
+        provider: 'openai:gpt-5-mini',
+        type: 'llm-rubric',
+        value:
+          'Should call trigger_agent tool to ask coder or designer to help with the login page task',
+      },
+    ],
+    vars: {
+      availableAgents: [
+        { id: 'coder', title: 'Code Wizard' },
+        { id: 'designer', title: 'UI Designer' },
+      ],
+      conversationHistory: 'User: I need help building a login page',
+      systemPrompt: 'You are coordinating a software development team',
+      userName: 'Bobs',
+    },
+  },
+  // just say hi - should only trigger_agent, no todo operations
+  {
+    assert: [
+      { type: 'is-valid-openai-tools-call' },
+      {
+        type: 'javascript',
+        value: `
+          // Ensure ONLY trigger_agent tool is called, no create_todo, finish_todo, etc.
+          const toolCalls = Array.isArray(output) ? output : [];
+          return toolCalls.length > 0 && toolCalls.every(call => call.function?.name === 'trigger_agent');
+        `,
+      },
+      {
+        provider: 'openai:gpt-5-mini',
+        type: 'llm-rubric',
+        value:
+          'Should call trigger_agent tool to greet the user or ask how to help. Should NOT include any create_todo or finish_todo calls.',
+      },
+    ],
+    vars: {
+      availableAgents: [
+        { id: 'agt_J34pj8igq5Hk', title: '全栈工程师' },
+        { id: 'agt_5xSoLVNHOjQj', title: '产品经理' },
+      ],
+      conversationHistory: '<message author="user">hi</message>',
+      role: 'user',
+      userName: 'Rene Wang',
+    },
+  },
+];
+
+export default testCases;
diff --git a/packages/prompts/promptfoo/supervisor/productive/tests/role.ts b/packages/prompts/promptfoo/supervisor/productive/tests/role.ts
new file mode 100644
index 00000000000..c4d7e3d3878
--- /dev/null
+++ b/packages/prompts/promptfoo/supervisor/productive/tests/role.ts
@@ -0,0 +1,58 @@
+const assert = [
+  { type: 'is-valid-openai-tools-call' },
+  {
+    type: 'javascript',
+    value: `
+          // Debug: log the actual output structure
+          console.log('DEBUG output:', JSON.stringify(output, null, 2));
+
+          // Ensure ONLY trigger_agent tool is called, no create_todo, finish_todo, etc.
+          const toolCalls = Array.isArray(output) ? output : [];
+          if (toolCalls.length === 0) {
+            console.log('DEBUG: No tool calls found');
+            return false;
+          }
+
+          for (const call of toolCalls) {
+            const toolName = call.tool_name || call.function?.name || call.name;
+            console.log('DEBUG tool name:', toolName);
+
+            if (toolName !== 'trigger_agent') {
+              console.log('DEBUG: Found non-trigger_agent tool:', toolName);
+              return false;
+            }
+          }
+
+          console.log('DEBUG: All', toolCalls.length, 'calls are trigger_agent');
+          return true;
+        `,
+  },
+  {
+    provider: 'openai:gpt-5-mini',
+    type: 'llm-rubric',
+    value:
+      'Should call trigger_agent tool to greet the user or ask how to help. Should NOT include any create_todo or finish_todo calls.',
+  },
+];
+const vars = {
+  availableAgents: [
+    { id: 'agt_J34pj8igq5Hk', title: '全栈工程师' },
+    { id: 'agt_5xSoLVNHOjQj', title: '产品经理' },
+  ],
+  conversationHistory: '<message author="user">hi</message>',
+  role: 'user',
+  userName: 'Rene Wang',
+};
+
+const testCases = [
+  {
+    assert,
+    vars: { ...vars, role: 'user' },
+  },
+  {
+    assert,
+    vars: { ...vars, role: 'system' },
+  },
+];
+
+export default testCases;
diff --git a/packages/prompts/promptfoo/supervisor/productive/tools.json b/packages/prompts/promptfoo/supervisor/productive/tools.json
new file mode 100644
index 00000000000..632e517c349
--- /dev/null
+++ b/packages/prompts/promptfoo/supervisor/productive/tools.json
@@ -0,0 +1,80 @@
+[
+  {
+    "type": "function",
+    "function": {
+      "name": "trigger_agent",
+      "description": "Trigger an agent to speak (group message).",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "id": {
+            "type": "string",
+            "description": "The agent id to trigger."
+          },
+          "instruction": {
+            "type": "string"
+          }
+        },
+        "required": ["instruction", "id"],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "type": "function",
+    "function": {
+      "name": "wait_for_user_input",
+      "description": "Wait for user input. Use this when the conversation history looks likes fine for now, or agents are waiting for user input.",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "reason": {
+            "type": "string",
+            "description": "Optional reason for pausing the conversation."
+          }
+        },
+        "required": [],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "type": "function",
+    "function": {
+      "name": "create_todo",
+      "description": "Create a new todo item",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "assignee": {
+            "type": "string",
+            "description": "Who will do the todo. Can be agent id or empty."
+          },
+          "content": {
+            "type": "string",
+            "description": "The todo content or description."
+          }
+        },
+        "required": ["content", "assignee"],
+        "additionalProperties": false
+      }
+    }
+  },
+  {
+    "type": "function",
+    "function": {
+      "name": "finish_todo",
+      "description": "Finish a todo by index or all todos",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "index": {
+            "type": "number"
+          }
+        },
+        "required": ["index"],
+        "additionalProperties": false
+      }
+    }
+  }
+]
diff --git a/packages/prompts/src/contexts/index.ts b/packages/prompts/src/contexts/index.ts
new file mode 100644
index 00000000000..6df5e42b61e
--- /dev/null
+++ b/packages/prompts/src/contexts/index.ts
@@ -0,0 +1 @@
+export * from './supervisor';
diff --git a/packages/prompts/src/contexts/supervisor/index.ts b/packages/prompts/src/contexts/supervisor/index.ts
new file mode 100644
index 00000000000..4b876669260
--- /dev/null
+++ b/packages/prompts/src/contexts/supervisor/index.ts
@@ -0,0 +1,2 @@
+export * from './makeDecision';
+export * from './tools';
diff --git a/packages/prompts/src/contexts/supervisor/makeDecision.ts b/packages/prompts/src/contexts/supervisor/makeDecision.ts
new file mode 100644
index 00000000000..1db45c91100
--- /dev/null
+++ b/packages/prompts/src/contexts/supervisor/makeDecision.ts
@@ -0,0 +1,68 @@
+import { ChatCompletionTool, ChatMessage, ChatStreamPayload } from '@lobechat/types';
+
+import { groupChatPrompts, groupSupervisorPrompts } from '../../prompts';
+import { SupervisorToolName, SupervisorTools } from './tools';
+
+interface SupervisorTodoItem {
+  // optional assigned owner (agent id or name)
+  assignee?: string;
+  content: string;
+  finished: boolean;
+}
+
+interface AgentItem {
+  id: string;
+  title?: string | null;
+}
+export interface SupervisorContext {
+  allowDM?: boolean;
+  availableAgents: AgentItem[];
+  messages: ChatMessage[];
+  // Group scene controls which tools are exposed (e.g., todos only in 'productive')
+  scene?: 'casual' | 'productive';
+  systemPrompt?: string;
+  todoList?: SupervisorTodoItem[];
+  userName?: string;
+}
+
+export const contextSupervisorMakeDecision = ({
+  allowDM,
+  scene,
+  systemPrompt,
+  availableAgents,
+  todoList,
+  userName,
+  messages,
+}: SupervisorContext) => {
+  const conversationHistory = groupSupervisorPrompts(messages);
+  const prompt = groupChatPrompts.buildSupervisorPrompt({
+    allowDM,
+    availableAgents: availableAgents.filter((agent) => agent.id),
+    conversationHistory,
+    scene,
+    systemPrompt,
+    todoList,
+    userName,
+  });
+
+  const tools = SupervisorTools.filter((tool) => {
+    if (tool.name === SupervisorToolName.trigger_agent_dm) {
+      return allowDM;
+    }
+
+    if ([SupervisorToolName.finish_todo, SupervisorToolName.create_todo].includes(tool.name)) {
+      return scene === 'productive';
+    }
+
+    return true;
+  }).map<ChatCompletionTool>((tool) => ({
+    function: tool,
+    type: 'function',
+  }));
+
+  return {
+    messages: [{ content: prompt, role: 'user' }],
+    temperature: 0.3,
+    tools,
+  } satisfies Partial<ChatStreamPayload>;
+};
diff --git a/packages/prompts/src/contexts/supervisor/tools.ts b/packages/prompts/src/contexts/supervisor/tools.ts
new file mode 100644
index 00000000000..91173a852d5
--- /dev/null
+++ b/packages/prompts/src/contexts/supervisor/tools.ts
@@ -0,0 +1,102 @@
+import { LobeUniformTool } from '@lobechat/types';
+
+export const SupervisorToolName = {
+  create_todo: 'create_todo',
+  finish_todo: 'finish_todo',
+  trigger_agent: 'trigger_agent',
+  trigger_agent_dm: 'trigger_agent_dm',
+  wait_for_user_input: 'wait_for_user_input',
+};
+
+export const SupervisorTools: LobeUniformTool[] = [
+  {
+    description: 'Trigger an agent to speak (group message).',
+    name: SupervisorToolName.trigger_agent,
+    parameters: {
+      properties: {
+        id: {
+          description: 'The agent id to trigger.',
+          type: 'string',
+        },
+        instruction: {
+          description:
+            'The instruction or message for the agent. No longer than 10 words. Always use English.',
+          type: 'string',
+        },
+      },
+      required: ['id', 'instruction'],
+      type: 'object',
+    },
+  },
+  {
+    description:
+      'Wait for user input. Use this when the conversation history looks likes fine for now, or agents are waiting for user input.',
+    name: SupervisorToolName.wait_for_user_input,
+    parameters: {
+      properties: {
+        reason: {
+          description: 'Optional reason for pausing the conversation.',
+          type: 'string',
+        },
+      },
+      required: [],
+      type: 'object',
+    },
+  },
+
+  {
+    description: 'Trigger an agent to DM another agent or user.',
+    name: SupervisorToolName.trigger_agent_dm,
+    parameters: {
+      additionalProperties: false,
+      properties: {
+        id: {
+          description: 'The agent id to trigger.',
+          type: 'string',
+        },
+        instruction: {
+          type: 'string',
+        },
+        target: {
+          description: 'The target agent id. Only used when need DM.',
+          type: 'string',
+        },
+      },
+      required: ['instruction', 'id', 'target'],
+      type: 'object',
+    },
+  },
+  {
+    description: 'Create a new todo item',
+    name: SupervisorToolName.create_todo,
+    parameters: {
+      additionalProperties: false,
+      properties: {
+        assignee: {
+          description: 'Who will do the todo. Can be agent id or empty.',
+          type: 'string',
+        },
+        content: {
+          description: 'The todo content or description.',
+          type: 'string',
+        },
+      },
+      required: ['content', 'assignee'],
+      type: 'object',
+    },
+  },
+  {
+    description: 'Finish a todo by index or all todos',
+    name: SupervisorToolName.finish_todo,
+    parameters: {
+      additionalProperties: false,
+      properties: {
+        index: {
+          type: 'number',
+        },
+      },
+      required: ['index'],
+      type: 'object',
+    },
+  },
+];
diff --git a/packages/prompts/src/index.ts b/packages/prompts/src/index.ts
index f2d2249018a..38d2d8574f8 100644
--- a/packages/prompts/src/index.ts
+++ b/packages/prompts/src/index.ts
@@ -1,2 +1,3 @@
 export * from './chains';
+export * from './contexts';
 export * from './prompts';
diff --git a/packages/types/src/aiChat.ts b/packages/types/src/aiChat.ts
index c95035cc698..b482c61b2f8 100644
--- a/packages/types/src/aiChat.ts
+++ b/packages/types/src/aiChat.ts
@@ -1,6 +1,7 @@
 import { z } from 'zod';
 
 import { ChatMessage } from './message';
+import { OpenAIChatMessage } from './openai/chat';
 import { LobeUniformTool, LobeUniformToolSchema } from './tool';
 import { ChatTopic } from './topic';
 
@@ -75,7 +76,9 @@ export const StructureOutputSchema = z.object({
   provider: z.string(),
   schema: StructureSchema.optional(),
   systemRole: z.string().optional(),
-  tools: z.array(LobeUniformToolSchema).optional(),
+  tools: z
+    .array(z.object({ function: LobeUniformToolSchema, type: z.literal('function') }))
+    .optional(),
 });
 
 interface IStructureSchema {
@@ -92,10 +95,13 @@ interface IStructureSchema {
 
 export interface StructureOutputParams {
   keyVaultsPayload: string;
-  messages: ChatMessage[];
+  messages: OpenAIChatMessage[];
   model: string;
   provider: string;
   schema?: IStructureSchema;
   systemRole?: string;
-  tools?: LobeUniformTool[];
+  tools?: {
+    function: LobeUniformTool;
+    type: 'function';
+  }[];
 }
diff --git a/src/server/routers/lambda/aiChat.ts b/src/server/routers/lambda/aiChat.ts
index 879ff607e20..4e194ca9b65 100644
--- a/src/server/routers/lambda/aiChat.ts
+++ b/src/server/routers/lambda/aiChat.ts
@@ -61,7 +61,7 @@ export const aiChatRouter = router({
       model: input.model,
       schema: input.schema,
       systemRole: input.systemRole,
-      tools: input.tools,
+      tools: input.tools?.map((item) => item.function),
     });
 
     log('generateObject completed, result: %O', result);
diff --git a/src/store/chat/slices/message/supervisor.test.ts b/src/store/chat/slices/message/supervisor.test.ts
index 8d5961dc58d..95787dddb08 100644
--- a/src/store/chat/slices/message/supervisor.test.ts
+++ b/src/store/chat/slices/message/supervisor.test.ts
@@ -5,10 +5,17 @@ import { aiChatService } from '@/services/aiChat';
 import { GroupChatSupervisor, type SupervisorContext } from './supervisor';
 
 vi.mock('@lobechat/prompts', () => ({
-  groupChatPrompts: {
-    buildSupervisorPrompt: vi.fn(() => 'structured-supervisor-prompt'),
-  },
-  groupSupervisorPrompts: vi.fn(() => 'conversation-history'),
+  contextSupervisorMakeDecision: vi.fn(() => ({
+    messages: [{ content: 'structured-supervisor-prompt', role: 'user' }],
+    temperature: 0.3,
+    tools: [
+      { function: { name: 'trigger_agent' }, type: 'function' },
+      { function: { name: 'wait_for_user_input' }, type: 'function' },
+      { function: { name: 'trigger_agent_dm' }, type: 'function' },
+      { function: { name: 'create_todo' }, type: 'function' },
+      { function: { name: 'finish_todo' }, type: 'function' },
+    ],
+  })),
 }));
 
 vi.mock('@/services/aiChat', () => ({
@@ -76,7 +83,7 @@ describe('GroupChatSupervisor', () => {
       temperature: 0.3,
     });
 
-    const toolNames = (payload.tools ?? []).map((tool: any) => tool.name);
+    const toolNames = (payload.tools ?? []).map((tool: any) => tool.function.name);
     expect(toolNames).toEqual(
       expect.arrayContaining([
         'trigger_agent',
diff --git a/src/store/chat/slices/message/supervisor.ts b/src/store/chat/slices/message/supervisor.ts
index bb96dfd9de0..9270690deed 100644
--- a/src/store/chat/slices/message/supervisor.ts
+++ b/src/store/chat/slices/message/supervisor.ts
@@ -1,4 +1,4 @@
-import { groupChatPrompts, groupSupervisorPrompts } from '@lobechat/prompts';
+import { contextSupervisorMakeDecision } from '@lobechat/prompts';
 import { ChatMessage, GroupMemberWithAgent } from '@lobechat/types';
 
 import { aiChatService } from '@/services/aiChat';
@@ -61,7 +61,7 @@ export class GroupChatSupervisor {
    * Make decision on who should speak next
    */
   async makeDecision(context: SupervisorContext): Promise<SupervisorDecisionResult> {
-    const { messages, availableAgents, userName, systemPrompt, allowDM, todoList } = context;
+    const { availableAgents } = context;
 
     // If no agents available, stop conversation
     if (availableAgents.length === 0) {
@@ -69,22 +69,7 @@ export class GroupChatSupervisor {
     }
 
     try {
-      // Create supervisor prompt with conversation context
-      const conversationHistory = groupSupervisorPrompts(messages);
-
-      const supervisorPrompt = groupChatPrompts.buildSupervisorPrompt({
-        allowDM,
-        availableAgents: availableAgents
-          .filter((agent) => agent.id)
-          .map((agent) => ({ id: agent.id!, title: agent.title })),
-        conversationHistory,
-        scene: context.scene,
-        systemPrompt,
-        todoList,
-        userName,
-      });
-
-      const response = await this.callLLMForDecision(supervisorPrompt, context);
+      const response = await this.callLLMForDecision(context);
       const result = this.parseSupervisorResponse(response, availableAgents, context);
 
       console.log('Supervisor TODO list:', result.todos);
@@ -102,123 +87,25 @@ export class GroupChatSupervisor {
    * Call LLM service to get supervisor decision
    */
   private async callLLMForDecision(
-    prompt: string,
     context: SupervisorContext,
   ): Promise<SupervisorToolCall[] | string> {
-    const supervisorConfig = {
-      model: context.model,
-      provider: context.provider,
-      temperature: 0.3,
-    };
-
-    // Build tools array
-    const tools: any[] = [
-      {
-        description: 'Trigger an agent to speak (group message).',
-        name: 'trigger_agent',
-        parameters: {
-          properties: {
-            id: {
-              description: 'The agent id to trigger.',
-              type: 'string',
-            },
-            instruction: {
-              description:
-                'The instruction or message for the agent. No longer than 10 words. Always use English.',
-              type: 'string',
-            },
-          },
-          required: ['id', 'instruction'],
-          type: 'object',
-        },
-      },
-      {
-        description:
-          'Wait for user input. Use this when the conversation history looks likes fine for now, or agents are waiting for user input.',
-        name: 'wait_for_user_input',
-        parameters: {
-          properties: {
-            reason: {
-              description: 'Optional reason for pausing the conversation.',
-              type: 'string',
-            },
-          },
-          required: [],
-          type: 'object',
-        },
-      },
-    ];
-
-    // Add DM tool if allowed
-    if (context.allowDM) {
-      tools.push({
-        description: 'Trigger an agent to DM another agent or user.',
-        name: 'trigger_agent_dm',
-        parameters: {
-          properties: {
-            id: {
-              description: 'The agent id to trigger.',
-              type: 'string',
-            },
-            instruction: {
-              description: 'The instruction or message for the agent.',
-              type: 'string',
-            },
-            target: {
-              description: 'The target agent id. Only used when need DM.',
-              type: 'string',
-            },
-          },
-          required: ['id', 'instruction', 'target'],
-          type: 'object',
-        },
-      });
-    }
-
-    // Add todo tools if in productive scene
-    if (context.scene === 'productive') {
-      tools.push(
-        {
-          description: 'Create a new todo item',
-          name: 'create_todo',
-          parameters: {
-            properties: {
-              assignee: {
-                description: 'Who will do the todo. Can be agent id or empty.',
-                type: 'string',
-              },
-              content: {
-                description: 'The todo content or description.',
-                type: 'string',
-              },
-            },
-            required: ['content', 'assignee'],
-            type: 'object',
-          },
-        },
-        {
-          description: 'Finish a todo by index',
-          name: 'finish_todo',
-          parameters: {
-            properties: {
-              index: {
-                description: 'The index of the todo to finish.',
-                type: 'number',
-              },
-            },
-            required: ['index'],
-            type: 'object',
-          },
-        },
-      );
-    }
+    const contexts = contextSupervisorMakeDecision({
+      allowDM: context.allowDM,
+      availableAgents: context.availableAgents
+        .filter((agent) => agent.id)
+        .map((agent) => ({ id: agent.id, title: agent.title })),
+      messages: context.messages,
+      scene: context.scene,
+      todoList: context.todoList,
+      userName: context.userName,
+    });
 
     try {
       const response = await aiChatService.generateJSON(
         {
-          messages: [{ content: prompt, role: 'user' }] as any,
-          tools,
-          ...supervisorConfig,
+          ...(contexts as any),
+          model: context.model,
+          provider: context.provider,
         },
         context.abortController || new AbortController(),
       );