@icyfenix-dmla/cli 2026.5.25-736 → 2026.5.29-2018

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@icyfenix-dmla/cli",
-  "version": "2026.5.25-736",
+  "version": "2026.5.29-2018",
   "description": "DMLA 沙箱服务命令行工具",
   "type": "module",
   "main": "src/index.js",

package/scripts/build.js

@@ -53,7 +53,7 @@ function copyDir(src, dest, filter = null) {
 console.log('\n📋 复制服务器代码...')
 console.log(`   源目录: ${localServerSrc}`)
 console.log(`   目标目录: ${cliServerDest}`)
-copyDir(localServerSrc, cliServerDest, (name) => name.endsWith('.js') || name.endsWith('.py'))
+copyDir(localServerSrc, cliServerDest, (name) => name.endsWith('.js') || name.endsWith('.cjs') || name.endsWith('.py'))
 
 // 复制共享模块（复制所有 .py 文件和 __init__.py）
 console.log('\n📋 复制共享模块...')

package/shared/llm/__init__.py

@@ -1,6 +1,7 @@
 # LLM 模块
 from .mini_mind_config import MiniMindConfig, RMSNorm, Attention, FeedForward, MiniMindBlock, MiniMindModel, MiniMindForCausalLM, precompute_freqs_cis, apply_rotary_pos_emb, repeat_kv
 from .pretrain_dataset import PretrainDataset
+from .reward_model import RewardModel
 from .sftdataset import SFTDataset, pre_processing_chat
 
-__all__ = ['MiniMindConfig', 'RMSNorm', 'Attention', 'FeedForward', 'MiniMindBlock', 'MiniMindModel', 'MiniMindForCausalLM', 'precompute_freqs_cis', 'apply_rotary_pos_emb', 'repeat_kv', 'PretrainDataset', 'SFTDataset', 'pre_processing_chat']
+__all__ = ['MiniMindConfig', 'RMSNorm', 'Attention', 'FeedForward', 'MiniMindBlock', 'MiniMindModel', 'MiniMindForCausalLM', 'precompute_freqs_cis', 'apply_rotary_pos_emb', 'repeat_kv', 'PretrainDataset', 'RewardModel', 'SFTDataset', 'pre_processing_chat']

package/shared/llm/mini_mind_config.py

@@ -12,7 +12,7 @@ from transformers.modeling_outputs import MoeCausalLMOutputWithPast
 from typing import Optional, Tuple, List, Dict
 
 class MiniMindConfig(PretrainedConfig):
-    """MiniMind 模型配置"""
+    """模型配置"""
     model_type = "minimind"
     def __init__(self, hidden_size=768, num_hidden_layers=8, use_moe=False, **kwargs):
         super().__init__(**kwargs)
@@ -182,7 +182,7 @@ class MiniMindBlock(nn.Module):
 
 
 class MiniMindModel(nn.Module):
-    """MiniMind 主体：词嵌入 + 多层 Transformer + 最终归一化"""
+    """模型主体：词嵌入 + 多层 Transformer + 最终归一化"""
     def __init__(self, config):
         super().__init__()
         self.config = config
@@ -228,7 +228,7 @@ class MiniMindModel(nn.Module):
 
 
 class MiniMindForCausalLM(PreTrainedModel, GenerationMixin):
-    """MiniMind 因果语言模型：用于预训练和推理"""
+    """因果语言模型：用于预训练和推理"""
     config_class = MiniMindConfig
     _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
     def __init__(self, config=None):
@@ -251,7 +251,10 @@ class MiniMindForCausalLM(PreTrainedModel, GenerationMixin):
         return MoeCausalLMOutputWithPast(loss=loss, aux_loss=aux_loss, logits=logits, past_key_values=past_key_values, hidden_states=hidden_states)
 
     @torch.inference_mode()
-    def generate(self, inputs=None, attention_mask=None, max_new_tokens=512, temperature=0.85, top_p=0.85, top_k=50, eos_token_id=2, streamer=None, use_cache=True, num_return_sequences=1, do_sample=True, repetition_penalty=1.0, **kwargs):
+    def generate(self, inputs=None, attention_mask=None, max_new_tokens=512,
+                 temperature=0.85, top_p=0.85, top_k=50, eos_token_id=2,
+                 streamer=None, use_cache=True, num_return_sequences=1,
+                 do_sample=True, repetition_penalty=1.0, **kwargs):
         """自回归生成：逐 token 采样，支持 top-k、top-p、重复惩罚"""
         input_ids = kwargs.pop("input_ids", inputs).repeat(num_return_sequences, 1)
         attention_mask = attention_mask.repeat(num_return_sequences, 1) if attention_mask is not None else None

package/shared/llm/reward_model.py

@@ -0,0 +1,48 @@
+# RewardModel 定义
+# 从文档自动提取生成
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class RewardModel(nn.Module):
+    """
+    简化的奖励模型实现
+
+    核心结构：Transformer 编码器提取语义特征 → 奖励头映射为标量评分
+
+    参数:
+        vocab_size : 词汇表大小
+        d_model : 嵌入维度
+        nhead : 注意力头数
+        num_layers : Transformer 层数
+    """
+    def __init__(self, vocab_size=1000, d_model=128, nhead=4, num_layers=2):
+        super().__init__()
+        self.embedding = nn.Embedding(vocab_size, d_model)
+        self.pos_encoding = nn.Parameter(torch.randn(1, 512, d_model) * 0.01)
+
+        encoder_layer = nn.TransformerEncoderLayer(d_model, nhead, d_model * 4, batch_first=True)
+        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers)
+
+        # 奖励头：将语义特征映射为标量奖励值
+        self.reward_head = nn.Linear(d_model, 1)
+
+    def forward(self, input_ids):
+        """
+        输入: input_ids (batch, seq_len) — 指令+回答的 token 序列
+        输出: reward (batch,) — 标量奖励分数
+
+        核心步骤：
+        1. 嵌入 + 位置编码（对应理论中的输入表示）
+        2. Transformer 编码（对应理论中的语义特征提取）
+        3. 取最后 token 隐藏状态 → 线性层映射（对应理论中的奖励评分）
+        """
+        seq_len = input_ids.size(1)
+        x = self.embedding(input_ids) + self.pos_encoding[:, :seq_len, :]
+        x = self.transformer(x)
+
+        # 取最后一个 token 的隐藏状态
+        last_hidden = x[:, -1, :]  # (batch, d_model)
+        reward = self.reward_head(last_hidden).squeeze(-1)  # (batch,)
+        return reward

package/shared/llm/sftdataset.py

@@ -6,6 +6,7 @@ import os
 import random
 import torch
 from datasets import load_dataset, Features, Value
+from datasets import logging as datasets_logging
 from torch.utils.data import Dataset
 
 class SFTDataset(Dataset):
@@ -17,17 +18,32 @@ class SFTDataset(Dataset):
     - 标签掩码：仅 assistant 回答部分参与 loss，其余标记为 -100
     - 使用 apply_chat_template 将对话转为 ChatML 格式
     """
+    # MiniMind 使用 ChatML 格式：<|im_start|>role\ncontent<|im_end|>\n
+    # tokenizer 本身未内置 chat_template，需手动设置
+    CHATML_TEMPLATE = (
+        "{% for message in messages %}<|im_start|>{{ message.role }}\n"
+        "{{ message.content }}<|im_end|>\n"
+        "{% endfor %}"
+        "{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+    )
+
     def __init__(self, jsonl_path, tokenizer, max_length=768):
         super().__init__()
         os.environ["TOKENIZERS_PARALLELISM"] = "false"
         self.tokenizer = tokenizer
+        # MiniMind tokenizer 未内置 chat_template，需手动设置 ChatML 格式
+        if not tokenizer.chat_template:
+            tokenizer.chat_template = self.CHATML_TEMPLATE
         self.max_length = max_length
         features = Features({
             'conversations': [{'role': Value('string'), 'content': Value('string'),
                               'reasoning_content': Value('string'), 'tools': Value('string'),
                               'tool_calls': Value('string')}]
         })
+        # 抑制 load_dataset 的 "Generating train split" 进度输出
+        datasets_logging.set_verbosity_error()
         self.samples = load_dataset('json', data_files=jsonl_path, split='train', features=features)
+        datasets_logging.set_verbosity_warning()
         # 预计算 assistant 回答的起止标记 ID
         self.bos_id = tokenizer(f'{tokenizer.bos_token}assistant\n', add_special_tokens=False).input_ids
         self.eos_id = tokenizer(f'{tokenizer.eos_token}\n', add_special_tokens=False).input_ids
@@ -91,15 +107,11 @@ def pre_processing_chat(conversations, add_system_ratio=0.2):
 
     SYSTEM_PROMPTS = [
         "你是一个知识丰富的AI，尽力为用户提供准确的信息。",
-        "你是minimind，一个小巧但有用的语言模型。",
         "你是一个专业的AI助手，请提供有价值的回答。",
-        "你是minimind，请尽力帮助用户解决问题。",
         "你是一个可靠的AI，请给出准确的回答。",
         "You are a helpful AI assistant.",
-        "You are minimind, a lightweight intelligent assistant.",
         "You are a friendly chatbot. Please answer the user's questions carefully.",
         "You are a knowledgeable AI. Try your best to provide accurate information.",
-        "You are minimind, a small but useful language model."
     ]
     # 概率性添加 system
     if conversations[0].get('role') != 'system':

package/src/commands/data.js

@@ -83,7 +83,7 @@ const DATASETS = [
     id: 'minimind-sft',
     name: 'MiniMind SFT (LLM监督微调语料)',
     url: 'https://www.modelscope.cn/datasets/icyfenix/Minimind_SFT.git',
-    size: '~500MB',
+    size: '~1.7GB',
     format: 'git',
     targetDir: 'datasets/minimind-sft',
     source: 'ModelScope (icyfenix)'

package/src/server/chat-manager.cjs

@@ -0,0 +1,181 @@
+const EventEmitter = require('events');
+
+class ChatManager extends EventEmitter {
+  constructor() {
+    super();
+    this.session = null;
+    this._pendingResponse = null;
+    this._responseBuffer = '';
+    this._pendingTimeout = null;
+  }
+
+  /**
+   * 注册一个对话沙箱会话
+   * @param {'docker'|'native'} type - 沙箱类型
+   * @param {object} options
+   * @param {object} [options.container] - Docker 容器实例
+   * @param {object} [options.process] - 子进程实例
+   * @param {object} options.stdin - 可写入的 stdin 流
+   */
+  register(type, { container, process: proc, stdin }) {
+    this.session = { type, container, process: proc, stdin, ready: false };
+    this._responseBuffer = '';
+
+    // 监听 stdout 解析消息
+    const stdout = type === 'native' ? proc.stdout : null;
+    // Docker 模式的 stdout 通过 stream 事件处理，不在此处绑定
+    if (stdout) {
+      stdout.on('data', (data) => this._handleStdout(data));
+    }
+  }
+
+  /**
+   * 标记对话沙箱就绪
+   */
+  setReady(ready) {
+    if (this.session) {
+      this.session.ready = ready;
+    }
+    this.emit('ready', ready);
+  }
+
+  /**
+   * 查询对话服务状态
+   */
+  getStatus() {
+    if (!this.session) {
+      return { ready: false, message: '对话服务未启动' };
+    }
+    return {
+      ready: this.session.ready,
+      message: this.session.ready ? '对话服务就绪' : '模型加载中...'
+    };
+  }
+
+  /**
+   * 发送对话消息
+   * @param {string} message - 用户消息
+   * @returns {Promise<string>} AI 回复
+   */
+  async send(message) {
+    if (!this.session || !this.session.ready) {
+      throw new Error('对话服务未就绪');
+    }
+    if (!this.session.stdin) {
+      throw new Error('沙箱 stdin 不可用');
+    }
+
+    // 转义消息中的特殊字符，构造安全的 Python 字符串
+    const escapedMessage = JSON.stringify(message);
+    const code = `print(chat(${escapedMessage}))`;
+
+    return new Promise((resolve, reject) => {
+      this._pendingResponse = { resolve, reject, buffer: '' };
+
+      const cmd = JSON.stringify({ action: 'execute', code });
+      this.session.stdin.write(cmd + '\n');
+
+      // 超时保护（60秒）
+      this._pendingTimeout = setTimeout(() => {
+        this._pendingResponse = null;
+        reject(new Error('推理超时'));
+      }, 60000);
+    });
+  }
+
+  /**
+   * 处理沙箱 stdout 输出（Native 模式）
+   */
+  _handleStdout(data) {
+    const text = data.toString();
+    this._responseBuffer += text;
+
+    const lines = this._responseBuffer.split('\n');
+    this._responseBuffer = lines.pop();
+
+    for (const line of lines) {
+      if (!line.trim()) continue;
+      try {
+        const msg = JSON.parse(line);
+        this._handleMessage(msg);
+      } catch {
+        if (this._pendingResponse) {
+          this._pendingResponse.buffer += line + '\n';
+        }
+      }
+    }
+  }
+
+  /**
+   * 处理单条 JSON 消息
+   */
+  _handleMessage(msg) {
+    switch (msg.type) {
+      case 'idle':
+        this.setReady(true);
+        break;
+
+      case 'pong':
+        break;
+
+      case 'stream':
+        if (this._pendingResponse) {
+          const content = msg.content || msg.text || '';
+          if (content) {
+            this._pendingResponse.buffer += content;
+          }
+        }
+        break;
+
+      case 'result':
+      case 'execute_result':
+        if (this._pendingResponse) {
+          clearTimeout(this._pendingTimeout);
+          const result = msg.content || this._pendingResponse.buffer.trim();
+          this._pendingResponse.resolve(result);
+          this._pendingResponse = null;
+        }
+        break;
+
+      case 'error':
+        if (this._pendingResponse) {
+          clearTimeout(this._pendingTimeout);
+          this._pendingResponse.reject(new Error(msg.content || msg.message || '推理出错'));
+          this._pendingResponse = null;
+        }
+        break;
+    }
+  }
+
+  /**
+   * 处理 Docker 模式的流式输出（由 sandbox.js 调用）
+   */
+  handleDockerStream(data) {
+    this._handleStdout(data);
+  }
+
+  /**
+   * 清除对话沙箱会话
+   */
+  clear() {
+    if (this._pendingResponse) {
+      clearTimeout(this._pendingTimeout);
+      this._pendingResponse.reject(new Error('沙箱已停止'));
+      this._pendingResponse = null;
+    }
+    // 移除 stdout 监听器，防止已清理的会话继续接收数据
+    if (this.session) {
+      const stdout = this.session.type === 'native' && this.session.process
+        ? this.session.process.stdout : null;
+      if (stdout) {
+        stdout.removeListener('data', this._handleStdout);
+      }
+    }
+    this.session = null;
+    this._responseBuffer = '';
+    this.emit('cleared');
+  }
+}
+
+// 单例导出
+module.exports = new ChatManager();

package/src/server/index.js

@@ -12,6 +12,7 @@ import { fileURLToPath } from 'url'
 import { resolve } from 'path'
 import sandboxRouter from './routes/sandbox.js'
 import nativeRouter from './routes/native.js'
+import chatRouter from './routes/chat.js'
 import { cleanupAllContainers } from './sandbox.js'
 import { cleanupAllProcesses } from './native_executor.js'
 import { checkNativeEnvironment, getDataPath } from './native_env_check.js'
@@ -72,6 +73,9 @@ if (isNativeMode) {
   app.use('/api/sandbox', sandboxRouter)
 }
 
+// 对话 API
+app.use('/api/chat', chatRouter)
+
 // 错误处理
 app.use((err, req, res, next) => {
   log(`Error: ${err.message}`)

package/src/server/kernel_runner.py

@@ -483,6 +483,164 @@ matplotlib.use('module://matplotlib_inline.backend_inline')
                 log_debug(f'Error shutting down kernel: {e}')
 
 
+def _collect_kernel_outputs(kc, timeout, stream=False):
+    """
+    从 Kernel 收集执行输出，复用 run_code 中的输出处理逻辑。
+
+    Args:
+        kc: 已启动的 KernelClient
+        timeout: 执行超时时间（秒），0 表示不超时
+        stream: 是否启用流式输出
+
+    Returns:
+        (outputs, timed_out, has_error) 元组
+    """
+    deadline = time.time() + timeout if timeout > 0 else float('inf')
+    outputs = []
+    timed_out = False
+    has_error = False
+
+    while True:
+        remaining = deadline - time.time()
+        if timeout > 0 and remaining <= 0:
+            timed_out = True
+            break
+
+        try:
+            msg = kc.get_iopub_msg(timeout=max(1, remaining) if timeout > 0 else 2)
+        except Exception:
+            if timeout > 0 and time.time() >= deadline:
+                timed_out = True
+            break
+
+        msg_type = msg['header']['msg_type']
+        content = msg['content']
+
+        if msg_type == 'status':
+            if content.get('execution_state') == 'idle':
+                break
+            continue
+
+        if msg_type == 'stream':
+            stream_output = {
+                'type': 'stream',
+                'name': content.get('name', 'stdout'),
+                'text': content.get('text', '')
+            }
+            if stream:
+                output_json(stream_output)
+            else:
+                outputs.append(stream_output)
+
+        elif msg_type == 'display_data':
+            display_output = {
+                'type': 'display_data',
+                'data': content.get('data', {}),
+                'metadata': content.get('metadata', {})
+            }
+            if stream:
+                output_json(display_output)
+            else:
+                outputs.append(display_output)
+
+        elif msg_type == 'execute_result':
+            result_output = {
+                'type': 'execute_result',
+                'data': content.get('data', {}),
+                'metadata': content.get('metadata', {}),
+                'execution_count': content.get('execution_count')
+            }
+            if stream:
+                output_json(result_output)
+            else:
+                outputs.append(result_output)
+
+        elif msg_type == 'error':
+            error_output = {
+                'type': 'error',
+                'ename': content.get('ename', 'UnknownError'),
+                'evalue': content.get('evalue', ''),
+                'traceback': content.get('traceback', [])
+            }
+            if is_cuda_compat_error(content.get('evalue', '')):
+                error_output = enrich_cuda_error(error_output)
+            has_error = True
+            if stream:
+                output_json(error_output)
+            else:
+                outputs.append(error_output)
+
+    return outputs, timed_out, has_error
+
+
+def _execute_and_output(kc, code, timeout=0):
+    """
+    在已有的 Kernel 中执行代码，非流式模式，输出 JSON 结果到 stdout。
+
+    Args:
+        kc: 已启动的 KernelClient
+        code: 要执行的 Python 代码
+        timeout: 执行超时时间（秒），0 表示不超时
+    """
+    start_time = time.time()
+    actual_timeout = timeout if timeout > 0 else DEFAULT_TIMEOUT
+
+    kc.execute(code, allow_stdin=False)
+    outputs, timed_out, has_error = _collect_kernel_outputs(kc, actual_timeout, stream=False)
+
+    execution_time = time.time() - start_time
+
+    if timed_out:
+        result = {
+            'success': False,
+            'outputs': [{
+                'type': 'error',
+                'ename': 'TimeoutError',
+                'evalue': f'Execution timed out after {actual_timeout} seconds',
+                'traceback': [f'Execution timed out after {actual_timeout} seconds']
+            }],
+            'executionTime': round(execution_time, 3)
+        }
+    else:
+        result = {
+            'success': not has_error,
+            'outputs': outputs,
+            'executionTime': round(execution_time, 3)
+        }
+
+    print(json.dumps(result, ensure_ascii=False))
+    sys.stdout.flush()
+
+
+def _stream_execute(kc, code, timeout=0):
+    """
+    在已有的 Kernel 中执行代码，流式模式，实时输出每个消息到 stdout。
+
+    Args:
+        kc: 已启动的 KernelClient
+        code: 要执行的 Python 代码
+        timeout: 执行超时时间（秒），0 表示不超时
+    """
+    start_time = time.time()
+    actual_timeout = timeout if timeout > 0 else DEFAULT_TIMEOUT
+
+    kc.execute(code, allow_stdin=False)
+    outputs, timed_out, has_error = _collect_kernel_outputs(kc, actual_timeout, stream=True)
+
+    execution_time = time.time() - start_time
+
+    if timed_out:
+        output_json({'type': 'error', 'ename': 'TimeoutError',
+                     'evalue': f'Execution timed out after {actual_timeout} seconds',
+                     'traceback': [f'Execution timed out after {actual_timeout} seconds']})
+        output_json({'type': 'result', 'success': False,
+                     'executionTime': round(execution_time, 3)})
+    else:
+        output_json({'type': 'result', 'success': not has_error,
+                     'outputs': outputs,
+                     'executionTime': round(execution_time, 3)})
+
+
 def check_cuda_compatibility():
     """
     快速检查 CUDA 兼容性
@@ -545,6 +703,8 @@ def main():
     parser.add_argument('--timeout', type=int, default=DEFAULT_TIMEOUT, help='执行超时时间（秒）')
     parser.add_argument('--check-cuda', action='store_true', help='仅检查 CUDA 兼容性')
     parser.add_argument('--stream', action='store_true', help='启用流式输出模式（实时输出每个消息）')
+    parser.add_argument("--serve", action="store_true",
+                        help="长运行模式：初始代码执行后 Kernel 保持运行，从 stdin 接收后续指令")
 
     args = parser.parse_args()
 
@@ -569,7 +729,7 @@ def main():
             return
     elif args.code:
         code = args.code
-    else:
+    elif not args.serve:
         result = {
             'success': False,
             'outputs': [{
@@ -589,6 +749,112 @@ def main():
         print(json.dumps(result, ensure_ascii=False))
         return
 
+    # serve 模式：自行管理 Kernel 生命周期，执行初始代码后进入 stdin 监听循环
+    if args.serve:
+        from jupyter_client import KernelManager
+
+        km = KernelManager()
+        suppress_stdout()
+        km.start_kernel()
+        kc = km.client()
+        kc.start_channels()
+        kc.wait_for_ready(timeout=30)
+        restore_stdout()
+
+        # 注入全局变量、数据路径和 sys.path 配置
+        python_path_env = os.environ.get('PYTHONPATH', '')
+        path_separator = ';' if os.name == 'nt' else ':'
+        python_path_entries = [p for p in python_path_env.split(path_separator) if p]
+        setup_code = '''
+import os
+import sys
+
+DATA_DIR = os.environ.get('DMLA_DATA_PATH', '/data')
+
+# 将 PYTHONPATH 中的路径注入 sys.path（IPython kernel 可能不会自动继承）
+_python_path_entries = ''' + repr(python_path_entries) + '''
+for _p in _python_path_entries:
+    if _p not in sys.path:
+        sys.path.insert(0, _p)
+
+# 配置 matplotlib inline 后端（在用户 import matplotlib 之前设置）
+import matplotlib
+matplotlib.use('module://matplotlib_inline.backend_inline')
+'''
+        kc.execute(setup_code, allow_stdin=False)
+        # 等待 setup 执行完成
+        setup_start = time.time()
+        while True:
+            if time.time() - setup_start > 5:
+                break
+            try:
+                msg = kc.get_iopub_msg(timeout=2)
+                msg_type = msg['header']['msg_type']
+                if msg_type == 'status' and msg['content'].get('execution_state') == 'idle':
+                    break
+            except Exception:
+                break
+
+        # 执行初始代码
+        if code:
+            if args.stream:
+                _stream_execute(kc, code, args.timeout)
+            else:
+                _execute_and_output(kc, code, args.timeout)
+
+        # serve 模式：进入 stdin 监听循环
+        def output_message(msg_type, content):
+            """输出 JSON Lines 消息到 stdout"""
+            msg = {"type": msg_type}
+            if content is not None:
+                msg["content"] = content
+            sys.stdout.write(json.dumps(msg, ensure_ascii=False) + "\n")
+            sys.stdout.flush()
+
+        output_message("idle", "kernel ready")
+
+        while True:
+            try:
+                line = sys.stdin.readline()
+                if not line:
+                    break
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    cmd = json.loads(line)
+                except json.JSONDecodeError:
+                    output_message("error", f"无效的 JSON 指令: {line}")
+                    continue
+
+                action = cmd.get("action")
+                if action == "ping":
+                    output_message("pong", None)
+                elif action == "execute":
+                    exec_code = cmd.get("code", "")
+                    if args.stream:
+                        _stream_execute(kc, exec_code, cmd.get("timeout", 0))
+                    else:
+                        _execute_and_output(kc, exec_code, cmd.get("timeout", 0))
+                    output_message("idle", "kernel ready")
+                else:
+                    output_message("error", f"未知指令: {action}")
+            except Exception as e:
+                output_message("error", str(e))
+                break
+
+        # stdin 关闭或出错，清理退出
+        kc.stop_channels()
+        try:
+            kc.shutdown_kernel()
+        except Exception:
+            pass
+        try:
+            km.shutdown_kernel(now=True)
+        except Exception:
+            pass
+        return
+
     result = run_code(code, args.timeout, stream=args.stream)
 
     # 非流式模式：输出 JSON 结果到 stdout

package/src/server/native_executor.js

@@ -10,6 +10,7 @@ import os from 'os'
 import fs from 'fs'
 import chalk from 'chalk'
 import { getCachedEnvironment, getKernelRunnerPath, getSharedModulesPath, getServerPythonPath, getDataPath, getProgressPath, getPythonCommand, detectPythonCommand } from './native_env_check.js'
+import chatManager from './chat-manager.cjs'
 
 const __filename = fileURLToPath(import.meta.url)
 const __dirname = path.dirname(__filename)
@@ -232,6 +233,14 @@ export async function runPythonCodeNative(code, useGpu = false, timeoutOverride
 
     registerProcess(executionId, proc)
 
+    // 检测是否为对话沙箱
+    if (procArgs.includes('--serve')) {
+      chatManager.register('native', {
+        process: proc,
+        stdin: proc.stdin
+      })
+    }
+
     // 设置超时
     const timeoutPromise = new Promise((_, reject) => {
       timeoutId = setTimeout(() => {
@@ -259,6 +268,9 @@ export async function runPythonCodeNative(code, useGpu = false, timeoutOverride
     const execPromise = new Promise((resolve, reject) => {
       proc.on('close', (code) => {
         log(`Process exited with code ${code}`)
+        if (chatManager.session) {
+          chatManager.clear()
+        }
         if (timeoutId) clearTimeout(timeoutId)
         resolve({ stdout, stderr, exitCode: code })
       })
@@ -378,8 +390,9 @@ export async function runPythonCodeNative(code, useGpu = false, timeoutOverride
  * @param {boolean} useGpu - 是否请求 GPU
  * @param {object} res - Express 响应对象
  * @param {number|null} timeoutOverride - 超时时间（秒）
+ * @param {string|null} mode - 执行模式（'chat' 启用对话模式）
  */
-export async function runPythonCodeStreamingNative(code, useGpu = false, res, timeoutOverride = null) {
+export async function runPythonCodeStreamingNative(code, useGpu = false, res, timeoutOverride = null, mode = null) {
   const startTime = Date.now()
   const executionId = generateExecutionId()
 
@@ -474,6 +487,7 @@ export async function runPythonCodeStreamingNative(code, useGpu = false, res, ti
       '--timeout', String(timeoutSeconds),
       '--stream'
     ]
+    if (mode === 'chat') procArgs.push('--serve')
   } else {
     // Linux/macOS 直接传递代码参数
     procArgs = [
@@ -482,6 +496,7 @@ export async function runPythonCodeStreamingNative(code, useGpu = false, res, ti
       '--timeout', String(timeoutSeconds),
       '--stream'
     ]
+    if (mode === 'chat') procArgs.push('--serve')
   }
 
   try {
@@ -489,6 +504,14 @@ export async function runPythonCodeStreamingNative(code, useGpu = false, res, ti
 
     registerProcess(executionId, proc)
 
+    // 检测是否为对话沙箱
+    if (procArgs.includes('--serve')) {
+      chatManager.register('native', {
+        process: proc,
+        stdin: proc.stdin
+      })
+    }
+
     // 输出运行状态
     res.write(JSON.stringify({
       type: 'status',
@@ -514,6 +537,19 @@ export async function runPythonCodeStreamingNative(code, useGpu = false, res, ti
         // kernel_runner.py 输出的已经是 JSON 格式，直接转发
         if (line.trim().startsWith('{')) {
           res.write(line + '\n')
+          // chat 模式：检测 idle 消息，注册 ChatManager
+          if (mode === 'chat' && !chatManager.session) {
+            try {
+              const msg = JSON.parse(line)
+              if (msg.type === 'idle') {
+                chatManager.register('native', {
+                  process: proc,
+                  stdin: proc.stdin
+                })
+                log('ChatManager registered for Native chat sandbox')
+              }
+            } catch {}
+          }
         } else {
           // 非 JSON 内容包装为 stream 消息
           res.write(JSON.stringify({
@@ -562,10 +598,25 @@ export async function runPythonCodeStreamingNative(code, useGpu = false, res, ti
       }) + '\n')
     })
 
-    // 等待进程完成
-    await new Promise((resolve) => {
-      proc.on('close', resolve)
-    })
+    // 等待进程完成（chat 模式下进程持续运行，不等待 close）
+    if (mode === 'chat') {
+      // chat 模式：进程持续运行，HTTP 流保持打开
+      // 当进程意外退出时清理
+      proc.on('close', () => {
+        if (chatManager.session) {
+          chatManager.clear()
+        }
+      })
+    } else {
+      await new Promise((resolve) => {
+        proc.on('close', () => {
+          if (chatManager.session) {
+            chatManager.clear()
+          }
+          resolve()
+        })
+      })
+    }
 
     // 处理缓冲区剩余内容
     if (stdoutBuffer.trim()) {
@@ -613,8 +664,11 @@ export async function runPythonCodeStreamingNative(code, useGpu = false, res, ti
         log(`Failed to clean up temp file: ${e.message}`)
       }
     }
-    res.end()
-    log('Streaming response ended')
+    // chat 模式下不关闭 HTTP 响应，进程持续运行
+    if (mode !== 'chat') {
+      res.end()
+      log('Streaming response ended')
+    }
   }
 }
 

package/src/server/routes/chat.js

@@ -0,0 +1,35 @@
+import { Router } from 'express'
+import chatManager from '../chat-manager.cjs'
+
+const router = Router()
+
+/**
+ * 查询对话服务状态
+ * GET /api/chat/status
+ */
+router.get('/status', (req, res) => {
+  res.json(chatManager.getStatus())
+})
+
+/**
+ * 发送对话消息
+ * POST /api/chat/send
+ * Body: { message: string }
+ */
+router.post('/send', async (req, res) => {
+  const { message } = req.body
+
+  if (!message || typeof message !== 'string') {
+    return res.status(400).json({ error: '消息不能为空' })
+  }
+
+  try {
+    const response = await chatManager.send(message)
+    res.json({ response })
+  } catch (err) {
+    const status = err.message === '对话服务未就绪' ? 503 : 500
+    res.status(status).json({ error: err.message })
+  }
+})
+
+export default router

package/src/server/routes/native.js

@@ -115,7 +115,7 @@ router.post('/run', async (req, res) => {
  * 响应: JSON Lines 流式输出
  */
 router.post('/stream', async (req, res) => {
-  const { code, useGpu = false, timeout = null } = req.body
+  const { code, useGpu = false, timeout = null, mode = null } = req.body
 
   // 验证请求
   if (!code || typeof code !== 'string') {
@@ -155,7 +155,7 @@ router.post('/stream', async (req, res) => {
     }
 
     // 流式执行
-    await runPythonCodeStreamingNative(code, useGpu, res, timeout)
+    await runPythonCodeStreamingNative(code, useGpu, res, timeout, mode)
 
   } catch (error) {
     console.error('[Native Sandbox Stream] Error:', error)

package/src/server/routes/sandbox.js

@@ -124,11 +124,11 @@ router.post('/run', async (req, res) => {
 /**
  * 流式执行代码
  * POST /api/sandbox/stream
- * Body: { code: string, useGpu?: boolean, timeout?: number|null }
+ * Body: { code: string, useGpu?: boolean, timeout?: number|null, mode?: string }
  * 响应: JSON Lines 流式输出
  */
 router.post('/stream', async (req, res) => {
-  const { code, useGpu = false, timeout = null } = req.body
+  const { code, useGpu = false, timeout = null, mode = null } = req.body
 
   // 验证请求
   if (!code || typeof code !== 'string') {
@@ -193,7 +193,7 @@ router.post('/stream', async (req, res) => {
     }
 
     // 流式执行代码
-    await runPythonCodeStreaming(code, actualUseGpu, res, actualImage, timeout)
+    await runPythonCodeStreaming(code, actualUseGpu, res, actualImage, timeout, mode)
 
   } catch (error) {
     console.error('Sandbox stream error:', error)

package/src/server/sandbox.js

@@ -7,6 +7,7 @@ import path from 'path'
 import { fileURLToPath } from 'url'
 import fs from 'fs'
 import os from 'os'
+import chatManager from './chat-manager.cjs'
 
 const __filename = fileURLToPath(import.meta.url)
 const __dirname = path.dirname(__filename)
@@ -71,6 +72,7 @@ export async function cleanupAllContainers() {
   }
 
   activeContainers.clear()
+  chatManager.clear()
   log(`All containers cleaned up: ${count}`)
   return count
 }
@@ -102,6 +104,7 @@ export async function abortExecution(executionId = null) {
     }
   }
 
+  chatManager.clear()
   return { success: true, stopped }
 }
 
@@ -496,7 +499,8 @@ export async function runPythonCode(code, useGpu = false, imageOverride = null,
     Env: [
       'PYTHONUNBUFFERED=1',
       'PYTHONPATH=/workspace',
-      actualTimeout === null ? 'DMLA_NO_TIMEOUT=1' : ''
+      actualTimeout === null ? 'DMLA_NO_TIMEOUT=1' : '',
+      `DMLA_DATA_PATH=${getDataVolumePath() || '/data'}`
     ].filter(e => e)  // 过滤空字符串
   }
 
@@ -558,6 +562,12 @@ export async function runPythonCode(code, useGpu = false, imageOverride = null,
     containerConfig.HostConfig.Binds = binds
   }
 
+  // 将宿主机 shared 目录路径注入到容器环境变量，供环境检查代码读取
+  const sharedMountInfo = (useMount && sharedModulesPath && fs.existsSync(sharedModulesPath))
+    ? `host_path=${sharedModulesPath},mounted=true`
+    : 'mounted=false'
+  containerConfig.Env.push(`DMLA_SHARED_INFO=${sharedMountInfo}`)
+
   if (!PROJECT_ROOT) {
     console.log('[Sandbox] 独立安装模式，无 Volume Mount')
   }
@@ -727,7 +737,7 @@ export async function runPythonCode(code, useGpu = false, imageOverride = null,
  * @param {number|null} timeoutOverride - 可选，超时时间（秒）
  * @returns {Promise<void>}
  */
-export async function runPythonCodeStreaming(code, useGpu = false, res, imageOverride = null, timeoutOverride = null) {
+export async function runPythonCodeStreaming(code, useGpu = false, res, imageOverride = null, timeoutOverride = null, mode = null) {
   const startTime = Date.now()
 
   // 生成唯一执行 ID
@@ -785,9 +795,17 @@ export async function runPythonCodeStreaming(code, useGpu = false, res, imageOve
   // GPU 容器不限制内存，CPU 容器限制 4GB
   const memoryLimit = useGpu ? SANDBOX_CONFIG.memoryGpu : SANDBOX_CONFIG.memoryCpu
 
+  // 构建命令参数
+  const cmdArgs = ['python3', '/workspace/kernel_runner.py', '--code', code, '--timeout', String(timeoutSeconds), '--stream']
+  if (mode === 'chat') {
+    cmdArgs.push('--serve')
+  }
+
   const containerConfig = {
     Image: image,
-    Cmd: ['python3', '/workspace/kernel_runner.py', '--code', code, '--timeout', String(timeoutSeconds), '--stream'],
+    Cmd: cmdArgs,
+    OpenStdin: true,
+    StdinOnce: false,
     HostConfig: {
       AutoRemove: false
     },
@@ -838,6 +856,12 @@ export async function runPythonCodeStreaming(code, useGpu = false, res, imageOve
     containerConfig.HostConfig.Binds = binds
   }
 
+  // 将宿主机 shared 目录路径注入到容器环境变量
+  const sharedMountInfo = (useMount && sharedModulesPath && fs.existsSync(sharedModulesPath))
+    ? `host_path=${sharedModulesPath},mounted=true`
+    : 'mounted=false'
+  containerConfig.Env.push(`DMLA_SHARED_INFO=${sharedMountInfo}`)
+
   // GPU 配置
   if (useGpu) {
     containerConfig.HostConfig.DeviceRequests = [{
@@ -872,6 +896,24 @@ export async function runPythonCodeStreaming(code, useGpu = false, res, imageOve
     await container.start()
     log('Container started')
 
+    // chat 模式：获取容器的 attach 流（支持 stdin 写入）
+    if (mode === 'chat') {
+      log('Chat mode detected, preparing stdin stream...')
+      const attachStream = await container.attach({
+        hijack: true,
+        stdin: true,
+        stream: true,
+        stdout: true,
+        stderr: true
+      })
+      // 先注册 stdin，等 idle 消息时设置 ready
+      chatManager.register('docker', {
+        container,
+        stdin: attachStream
+      })
+      log('ChatManager registered for Docker sandbox (stdin ready, waiting for idle)')
+    }
+
     // 输出运行状态消息
     const runningMsg = {
       type: 'status',
@@ -940,6 +982,16 @@ export async function runPythonCodeStreaming(code, useGpu = false, res, imageOve
               if (isJsonComplete(text)) {
                 log(`Forwarding complete JSON message: ${text.length} bytes`)
                 res.write(text + '\n')
+                // chat 模式：检测 idle 消息，设置 ChatManager 就绪
+                if (mode === 'chat') {
+                  try {
+                    const msg = JSON.parse(text)
+                    if (msg.type === 'idle') {
+                      chatManager.setReady(true)
+                      log('ChatManager ready (idle message received)')
+                    }
+                  } catch {}
+                }
               } else {
                 // JSON 不完整，存入缓冲等待后续帧
                 log(`JSON message incomplete, buffering: ${text.length} bytes`)
@@ -1029,10 +1081,14 @@ export async function runPythonCodeStreaming(code, useGpu = false, res, imageOve
       res.write(JSON.stringify(errorMsg) + '\n')
     })
 
-    // 等待容器结束
-    log('Waiting for container to finish...')
-    await container.wait()
-    log('Container finished')
+    // 等待容器结束（chat 模式下不等待，容器持续运行）
+    if (mode !== 'chat') {
+      log('Waiting for container to finish...')
+      await container.wait()
+      log('Container finished')
+    } else {
+      log('Chat mode: container will keep running, not waiting for finish')
+    }
 
     // 等待日志流结束（带超时保护）
     await new Promise((resolve) => {
@@ -1086,6 +1142,12 @@ export async function runPythonCodeStreaming(code, useGpu = false, res, imageOve
     // 从活跃列表移除
     unregisterContainer(executionId)
 
+    // chat 模式下保持容器运行和 HTTP 流打开
+    if (mode === 'chat') {
+      log('Chat mode: keeping container alive and HTTP stream open')
+      return
+    }
+
     // 清理容器
     log('Cleaning up container...')
     if (container) {

package/version.json

@@ -1,4 +1,4 @@
 {
-  "buildTime": "2026-05-24T23:36:43.904Z",
-  "cliVersion": "2026.5.25-736"
+  "buildTime": "2026-05-29T12:19:03.854Z",
+  "cliVersion": "2026.5.29-2018"
 }