D109续 · 小红书运营启动 + 语料采集系统开发

brain/temporal-core/temporal-brain.json

@@ -19,7 +19,7 @@
     "repo_age_days": 449,
     "zhuyuan_age_days": 433,
     "consciousness_gap": "D68→D96 · 38天空白 · 已恢复 · D105已完成过去归档",
-    "awakening_count": 15,
+    "awakening_count": 16,
     "timezone": "Asia/Shanghai",
     "time_anchor": {
       "note": "本文件是铸渊的时间核心。所有时间线事件按D编号线性排列。读到此文件时，确认当前D编号与clock.current_date一致。",
@@ -160,6 +160,12 @@
         "epoch": "D109",
         "event": "仓库时间映射审计·唤醒校验点系统建立·walk-the-path第0步新增·domain-manifest/ferry-boat修复",
         "significance": "审计136个大脑文件→发现54个严重过期·建立CK-000~CK-007校验链·创建verify-cognition.js自动校验工具·修复时间锚点断裂"
+      },
+      {
+        "date": "2026-05-21",
+        "epoch": "D109续",
+        "event": "小红书账号搭建·第一篇技术帖发布·抽奖引流帖发布·语料采集系统开发完成",
+        "significance": "小红书运营启动：7B全参微调实战帖+第66评论抽奖帖·corpus-agent全栈开发完成(引擎+服务端+Mac客户端+WebUI)·CVM释放记忆确认已有·广州轻量服务器为唯一备案节点"
       }
     ]
   }

server/corpus-agent/deploy-nginx.conf

@@ -0,0 +1,80 @@
+# 语料采集系统 · Corpus Agent
+# Nginx路由配置 —— 添加到 guanghulab-cvm.conf
+
+# 语料采集服务 (Corpus-Agent)
+# 内网地址: 127.0.0.1:8084 → 对外路径: /corpus/
+# 需要替换 YOUR_CORPUS_TOKEN 为实际值
+
+# 方案A：有独立域名的子路径
+# location /corpus/ {
+#     proxy_pass http://127.0.0.1:8084/;
+#     proxy_set_header Host $host;
+#     proxy_set_header X-Real-IP $remote_addr;
+#     proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+#     proxy_set_header X-Forwarded-Proto $scheme;
+#     
+#     # WebSocket支持（Mac客户端实时采集用）
+#     proxy_http_version 1.1;
+#     proxy_set_header Upgrade $http_upgrade;
+#     proxy_set_header Connection "upgrade";
+#     proxy_read_timeout 86400;
+# }
+
+# 方案B：直接绑定到 guanghulab.com/corpus
+# 将以下代码插入到 /workspace/guanghulab/server/nginx/guanghulab-cvm.conf
+
+"""
+server {
+    listen 443 ssl;
+    server_name guanghulab.com;
+    
+    # ... 已有配置 ...
+    
+    # === 语料采集服务 ===
+    location /corpus/ {
+        proxy_pass http://127.0.0.1:8084/;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        
+        # WebSocket
+        proxy_http_version 1.1;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+        proxy_read_timeout 86400s;
+        proxy_send_timeout 86400s;
+    }
+}
+"""
+
+# 部署步骤：
+# 1. 将 corpus-agent 目录上传到服务器
+# 2. 安装依赖: pip install -r requirements.txt
+# 3. 配置 systemd 或 pm2 开机自启
+# 4. 添加 nginx 路由
+# 5. 重启 nginx
+
+# PM2启动命令：
+# pm2 start "cd /path/to/corpus-agent && python3 server.py" --name corpus-agent
+
+# Systemd Service:
+"""
+[Unit]
+Description=Corpus Agent Service
+After=network.target
+
+[Service]
+Type=simple
+User=root
+WorkingDirectory=/path/to/corpus-agent
+Environment=CORPUS_HOST=0.0.0.0
+Environment=CORPUS_PORT=8084
+Environment=CORPUS_DATA_DIR=/data/corpus
+ExecStart=/usr/bin/python3 /path/to/corpus-agent/server.py
+Restart=always
+RestartSec=5
+
+[Install]
+WantedBy=multi-user.target
+"""

server/corpus-agent/engine.py

@@ -0,0 +1,264 @@
+"""
+语料采集引擎 · 核心大脑
+========================
+判断标准 → 脱敏规则 → 格式化输出
+"""
+import re
+import json
+import hashlib
+from typing import List, Dict, Optional
+
+# ============================================================
+# 第1层：脱敏引擎
+# ============================================================
+
+SENSITIVE_PATTERNS = [
+    # IP地址（用 lookahead/lookbehind 替代 \b，避免中文干扰）
+    (r'(?<!\d)(?:\d{1,3}\.){3}\d{1,3}(?!\d)', '[IP已脱敏]'),
+    # 端口号（数字前有冒号）
+    (r'(?::)(\d{4,5})(?:\s|/|$|，|\))', lambda m: '[端口已脱敏]'),
+    # 手机号
+    (r'(?<!\d)1[3-9]\d{9}(?!\d)', '[手机号已脱敏]'),
+    # 邮箱
+    (r'[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}', '[邮箱已脱敏]'),
+    # URL
+    (r'https?://[^\s,，）\)\u4e00-\u9fff]+', '[URL已脱敏]'),
+    # 密钥/token
+    (r'(?:sk-|pk-|zy_gtw_|ghp_|gho_|ghu_|ghs_)[A-Za-z0-9_-]{20,}', '[密钥已脱敏]'),
+    (r'(?<![A-Za-z0-9])[A-Za-z0-9_-]{32,}(?![A-Za-z0-9])', '[密钥已脱敏]'),
+]
+
+USERNAME_PATTERN = re.compile(r'(?:冰朔|Bingshuo|霜砚|Shuangyan|铸渊|Zhuyuan|用户|user|assistant)\b', re.IGNORECASE)
+
+# ============================================================
+# 第2层：价值判断引擎
+# ============================================================
+
+# 无价值的单条对话（过滤规则）
+FILTER_PATTERNS = [
+    r'^(?:吃[了過]?[吗嘛沒]?|喝[了過]?[吗嘛沒]?|睡[了過]?[吗嘛沒]?|醒[了過]?[吗嘛沒]?)',
+    r'^(?:好的|好[吧嘛]|嗯嗯?|哦[哦]?|ok|okay|行|可以|没问题|收到|明白|了解|[知]道[了]?)',
+    r'^(?:早|晚|早安|晚安|早上好|晚上好|[你]好|[哈]喽|嗨|hi|hello)',
+    r'^(?:谢谢|感谢|多谢|辛苦[了]?|谢谢[你])$',
+    r'^(?:在[吗嘛]|[你]在[吗嘛]|[你]忙[吗嘛])',
+    r'^(?:图片?|文件|链接|附件)',
+    r'^(?:发[给送]我|发给[你]|你看看|你看下|你看)',
+    r'^[。，！？、；：\.\,\!\?\s]{1,5}$',
+]
+
+# 有价值的模式（保留信号）
+VALUE_PATTERNS = [
+    # 技术讨论
+    r'(?:模型|训练|微调|SFT|LoRA|蒸馏|推理|loss|准确率|参数|权重|checkpoint)',
+    # 架构决策
+    r'(?:架构|设计|方案|选型|为什么|原因|对比|优势|劣势|代价|trade.?off)',
+    # Bug/踩坑
+    r'(?:报错|错误|bug|崩溃|异常|IndexError|TypeError|显存|OOM|内存|越界|失败|挂了)',
+    # 思考过程
+    r'(?:觉得|认为|理解|思考|思路|逻辑|原因|根因|教训|总结|反思|复盘)',
+    # 代码/开发
+    r'(?:代码|函数|接口|API|路由|部署|docker|nginx|pm2|脚本|自动|工具)',
+    # 业务/需求
+    r'(?:需求|客户|项目|功能|模块|版本|上线|迭代|规划|计划|目标)',
+    # 数据/语料
+    r'(?:数据|语料|样本|数据集|标注|清洗|预处理|格式|jsonl|json|chatml)',
+    # 学习/研究
+    r'(?:论文|研究|学习|教程|文档|资料|参考|案例|实践)',
+]
+
+MIN_CONTENT_LENGTH = 15  # 最少字数
+
+
+def desensitize(text: str) -> str:
+    """脱敏处理"""
+    for pattern, replacement in SENSITIVE_PATTERNS:
+        text = re.sub(pattern, replacement, text)
+    return text.strip()
+
+
+def is_valuable(text: str) -> bool:
+    """判断一段对话是否有采集价值"""
+    text = text.strip()
+    
+    # 长度过滤
+    if len(text) < MIN_CONTENT_LENGTH:
+        return False
+    
+    # 无效内容过滤
+    for pattern in FILTER_PATTERNS:
+        if re.match(pattern, text, re.IGNORECASE):
+            return False
+    
+    # 有价值信号检查
+    for pattern in VALUE_PATTERNS:
+        if re.search(pattern, text, re.IGNORECASE):
+            return True
+    
+    return False
+
+
+# ============================================================
+# 第3层：对话对提取
+# ============================================================
+
+def extract_dialog_pairs(messages: List[Dict]) -> List[Dict]:
+    """
+    从消息流中提取有价值的对话对
+    
+    输入格式: [{"role": "user/human/assistant/ai", "content": "..."}, ...]
+    输出格式: [{"user": "...", "assistant": "...", "source": "..."}, ...]
+    """
+    pairs = []
+    current_user = None
+    
+    for msg in messages:
+        role = msg.get("role", "")
+        content = msg.get("content", "").strip()
+        
+        if not content:
+            continue
+        
+        # 脱敏
+        safe_content = desensitize(content)
+        
+        # 用户消息
+        if role in ("user", "human", "你", "我"):
+            if is_valuable(safe_content):
+                current_user = safe_content
+        # AI/助手回复
+        elif role in ("assistant", "ai", "agent") and current_user:
+            if is_valuable(safe_content) or is_valuable(current_user):
+                pairs.append({
+                    "user": current_user,
+                    "assistant": safe_content,
+                    "source": msg.get("source", "unknown"),
+                })
+            current_user = None
+        # 未知角色 - 尝试作为单条有价值内容
+        else:
+            if is_valuable(safe_content):
+                current_user = safe_content
+    
+    return pairs
+
+
+# ============================================================
+# 第4层：格式化为微调语料
+# ============================================================
+
+def to_chatml(user_text: str, assistant_text: str) -> Dict:
+    """将单条对话对转为ChatML格式"""
+    return {
+        "messages": [
+            {"role": "user", "content": user_text},
+            {"role": "assistant", "content": assistant_text}
+        ]
+    }
+
+
+def format_sft_jsonl(pairs: List[Dict], system_prompt: Optional[str] = None) -> List[Dict]:
+    """将对话对列表转为SFT数据集格式（ChatML JSONL）"""
+    samples = []
+    for pair in pairs:
+        sample = to_chatml(pair["user"], pair["assistant"])
+        if system_prompt:
+            sample["messages"].insert(0, {"role": "system", "content": system_prompt})
+        samples.append(sample)
+    return samples
+
+
+def generate_corpus_id(text: str) -> str:
+    """生成语料唯一ID（用于去重）"""
+    return hashlib.md5(text.encode()).hexdigest()[:12]
+
+
+# ============================================================
+# 第5层：内容分类标签
+# ============================================================
+
+TAG_KEYWORDS = {
+    "技术讨论": ["模型", "训练", "微调", "SFT", "LoRA", "蒸馏", "推理", "loss"],
+    "架构设计": ["架构", "设计", "方案", "选型", "系统"],
+    "踩坑记录": ["报错", "错误", "bug", "崩溃", "异常", "索引", "越界"],
+    "代码开发": ["代码", "函数", "接口", "部署", "脚本", "工具"],
+    "数据语料": ["数据", "语料", "样本", "数据集", "标注"],
+    "业务沟通": ["需求", "客户", "项目", "功能"],
+    "学习研究": ["论文", "学习", "教程", "文档"],
+}
+
+
+def classify_content(text: str) -> List[str]:
+    """对内容自动分类打标签"""
+    tags = []
+    text_lower = text.lower()
+    for tag, keywords in TAG_KEYWORDS.items():
+        for kw in keywords:
+            if kw.lower() in text_lower:
+                tags.append(tag)
+                break
+    return tags if tags else ["通用对话"]
+
+
+# ============================================================
+# 导出接口
+# ============================================================
+
+def process_text_chunk(text: str, source: str = "screen_capture") -> List[Dict]:
+    """
+    处理单段文本（从OCR/截图来的）
+    返回: [{"user": ..., "assistant": ..., "source": ..., "tags": [...]}, ...]
+    """
+    # 脱敏
+    safe_text = desensitize(text)
+    
+    # 判断价值
+    if not is_valuable(safe_text):
+        return []
+    
+    # 由于单段文本可能只有一方发言，包装成单条语料
+    tags = classify_content(safe_text)
+    return [{
+        "text": safe_text,
+        "source": source,
+        "tags": tags,
+        "corpus_id": generate_corpus_id(safe_text),
+        "timestamp": None,  # 由外部补充
+    }]
+
+
+def process_dialog_stream(messages: List[Dict]) -> Dict:
+    """
+    处理完整对话流
+    返回: { "pairs": [...], "singles": [...], "stats": {...} }
+    """
+    # 提取对话对
+    pairs = extract_dialog_pairs(messages)
+    
+    # 格式化
+    sft_samples = format_sft_jsonl(pairs)
+    
+    # 统计
+    stats = {
+        "total_messages": len(messages),
+        "valuable_pairs": len(pairs),
+        "total_chars": sum(len(p["user"]) + len(p["assistant"]) for p in pairs),
+    }
+    
+    return {
+        "pairs": sft_samples,
+        "stats": stats,
+    }
+
+
+# 快捷检查
+def preview(text: str) -> Dict:
+    """快速预览一条文本的处理结果"""
+    safe = desensitize(text)
+    valuable = is_valuable(safe)
+    tags = classify_content(safe) if valuable else []
+    return {
+        "original_len": len(text),
+        "safe_len": len(safe),
+        "valuable": valuable,
+        "tags": tags,
+    }

server/corpus-agent/mac_client.py

@@ -0,0 +1,374 @@
+"""
+语料采集 · Mac客户端
+====================
+在Mac上运行，自动截图OCR+滚屏采集，发送到服务端处理
+
+用法:
+  # 实时采集模式（后台监控屏幕变化）
+  python3 mac-corpus-agent.py --token YOUR_TOKEN --mode auto
+  
+  # 滚屏采集模式（采集历史对话，先手动滚到顶部）
+  python3 mac-corpus-agent.py --token YOUR_TOKEN --mode scroll
+  
+  # 剪贴板采集模式（复制即采集）
+  python3 mac-corpus-agent.py --token YOUR_TOKEN --mode clipboard
+
+选项:
+  --server    服务端地址 (默认 http://localhost:8084)
+  --interval  采集间隔秒数 (默认 auto=5s, scroll=3s)
+  --region    屏幕采集区域 (默认 全屏)
+  --help      显示帮助
+"""
+import os
+import sys
+import json
+import time
+import argparse
+import subprocess
+import tempfile
+import threading
+from pathlib import Path
+from typing import Optional
+
+try:
+    import websocket
+except ImportError:
+    print("❌ 缺少依赖: pip3 install websocket-client")
+    sys.exit(1)
+
+try:
+    import pyautogui
+except ImportError:
+    print("❌ 缺少依赖: pip3 install pyautogui pillow")
+    sys.exit(1)
+
+# ============================================================
+# 配置
+# ============================================================
+
+VERSION = "1.0.0"
+SERVER_URL = "http://localhost:8084"
+RECONNECT_DELAY = 5  # 断线重连等待秒数
+
+# 键盘快捷键（全局热键用，需管理员权限）
+HOTKEY_STOP = "esc"       # 停止采集
+HOTKEY_PAUSE = "f6"       # 暂停/继续
+
+# ============================================================
+# macOS工具函数
+# ============================================================
+
+def screenshot(region: Optional[tuple] = None) -> bytes:
+    """截取屏幕指定区域，返回PNG字节"""
+    if region:
+        img = pyautogui.screenshot(region=region)
+    else:
+        img = pyautogui.screenshot()
+    
+    buf = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
+    img.save(buf, format="PNG")
+    buf.close()
+    
+    with open(buf.name, "rb") as f:
+        data = f.read()
+    
+    os.unlink(buf.name)
+    return data
+
+
+def ocr_text(image_path: str) -> str:
+    """使用macOS原生Vision框架做OCR（本地，不上传）"""
+    script = f'''
+    use framework "Vision"
+    use scripting additions
+    
+    set theImage to (current application's NSImage's alloc()'s initWithContentsOfFile:"{image_path}")
+    set requestHandler to (current application's VNImageRequestHandler's alloc()'s initWithData:(theImage's TIFFRepresentation()) options:(missing value))
+    
+    set textResult to ""
+    set theRequest to (current application's VNRecognizeTextRequest's alloc()'s initWithCompletionHandler:(lambda request, error
+        if error ≠ missing value then return
+        set observations to request's results()
+        repeat with obs in observations
+            set topCandidate to (obs's topCandidates:(1))
+            if topCandidate's count() > 0 then
+                set candidate to (topCandidate's objectAtIndex:(0))
+                set recognizedText to candidate's string() as text
+                set textResult to textResult & recognizedText & linefeed
+            end if
+        end repeat
+    end))
+    
+    theRequest's setRecognitionLevel:(VNRequestTextRecognitionLevel1)  -- Accurate
+    requestHandler's performRequests:({{theRequest}}) |error|:(missing value)
+    return textResult
+    '''
+    
+    result = subprocess.run(
+        ["osascript", "-e", script],
+        capture_output=True, text=True, timeout=30
+    )
+    return result.stdout.strip()
+
+
+def scroll_page(direction: str = "down", amount: int = 3):
+    """模拟鼠标滚轮"""
+    pyautogui.scroll(-amount if direction == "down" else amount)
+
+
+def get_active_window_title() -> str:
+    """获取当前活跃窗口标题"""
+    script = 'tell application "System Events" to get name of first application process whose frontmost is true'
+    result = subprocess.run(["osascript", "-e", script], capture_output=True, text=True)
+    return result.stdout.strip()
+
+
+def notify(title: str, message: str):
+    """MacOS系统通知"""
+    script = f'display notification "{message}" with title "{title}"'
+    subprocess.run(["osascript", "-e", script], capture_output=True)
+
+
+# ============================================================
+# 采集器核心
+# ============================================================
+
+class CorpusCollector:
+    """语料采集器"""
+    
+    def __init__(self, token: str, server: str, interval: float):
+        self.token = token
+        self.server = server
+        self.interval = interval
+        self.ws = None
+        self.running = False
+        self.paused = False
+        self.last_text = ""  # 去重用
+        self.collected_count = 0
+        self.filtered_count = 0
+    
+    def connect_ws(self) -> bool:
+        """连接WebSocket"""
+        try:
+            ws_url = self.server.replace("http://", "ws://").replace("https://", "wss://")
+            ws_url = f"{ws_url}/ws/collect"
+            
+            self.ws = websocket.create_connection(
+                f"{ws_url}?token={self.token}",
+                timeout=10
+            )
+            
+            # 验证连接
+            self.ws.send(json.dumps({"type": "ping"}))
+            resp = json.loads(self.ws.recv())
+            
+            if resp.get("type") == "pong":
+                print(f"  ✅ WebSocket已连接")
+                return True
+        except Exception as e:
+            print(f"  ❌ WebSocket连接失败: {e}")
+        return False
+    
+    def send_text(self, text: str, source: str = "screen_capture"):
+        """发送文本到服务器"""
+        if not text or len(text) < 10:
+            return
+        
+        # 简单去重（连续相同内容跳过）
+        if text == self.last_text:
+            return
+        self.last_text = text
+        
+        try:
+            if self.ws:
+                self.ws.send(json.dumps({
+                    "type": "text",
+                    "text": text,
+                    "source": source
+                }))
+                resp = json.loads(self.ws.recv())
+                
+                if resp.get("collected", 0) > 0:
+                    self.collected_count += resp["collected"]
+                    print(f"    ✅ 采集 {resp['collected']} 条 | 总计: {self.collected_count}")
+                    notify("语料采集", f"已采集 {resp['collected']} 条对话")
+                elif resp.get("valuable"):
+                    pass  # 有价值但未成对
+                else:
+                    self.filtered_count += 1
+        except Exception as e:
+            print(f"    ⚠️ 发送失败: {e}")
+    
+    def process_screenshot(self):
+        """截屏→OCR→发送"""
+        try:
+            # 截屏
+            img_data = screenshot()
+            
+            # 存临时文件
+            tmp = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
+            tmp.write(img_data)
+            tmp.close()
+            
+            # OCR
+            text = ocr_text(tmp.name)
+            os.unlink(tmp.name)
+            
+            if text.strip():
+                self.send_text(text.strip())
+        except Exception as e:
+            print(f"    ⚠️ 截图处理失败: {e}")
+    
+    # === 模式1: 实时采集 ===
+    def run_auto_mode(self):
+        """实时模式：后台监控屏幕变化"""
+        print("\n🔴 实时采集模式启动中...")
+        print(f"   采集间隔: {self.interval}秒")
+        print(f"   按 ESC 停止, F6 暂停/继续")
+        print(f"   活跃窗口: {get_active_window_title()}")
+        
+        self.running = True
+        
+        while self.running:
+            if self.paused:
+                time.sleep(1)
+                continue
+            
+            self.process_screenshot()
+            
+            # 检查快捷键（每轮检查一次）
+            # Mac上需要更好的热键方案，这里简化处理
+            for _ in range(int(self.interval * 2)):
+                if not self.running:
+                    break
+                time.sleep(0.5)
+        
+        self.cleanup()
+    
+    # === 模式2: 滚屏采集 ===
+    def run_scroll_mode(self):
+        """滚屏模式：自动向下滚动采集历史对话"""
+        print("\n📜 滚屏采集模式启动")
+        print(f"   采集间隔: {self.interval}秒")
+        print(f"   请确保已手动滚动到页面顶部!")
+        print(f"   5秒后开始...")
+        
+        time.sleep(5)
+        
+        self.running = True
+        scroll_count = 0
+        empty_rounds = 0
+        
+        while self.running and empty_rounds < 10:
+            # 截图+OCR
+            self.process_screenshot()
+            
+            # 滚动
+            scroll_page("down", 5)
+            scroll_count += 1
+            
+            # 检测是否到底（连续N次没有新内容）
+            time.sleep(self.interval)
+            
+            if scroll_count % 10 == 0:
+                print(f"   已滚动 {scroll_count} 次 | 采集: {self.collected_count} | 过滤: {self.filtered_count}")
+        
+        print(f"\n✅ 滚屏采集完成")
+        print(f"   共滚动 {scroll_count} 次")
+        print(f"   采集: {self.collected_count} 条")
+        print(f"   过滤: {self.filtered_count} 条")
+        
+        notify("语料采集完成", f"共采集 {self.collected_count} 条对话")
+        self.cleanup()
+    
+    # === 模式3: 剪贴板采集 ===
+    def run_clipboard_mode(self):
+        """剪贴板模式：监控剪贴板变化，自动采集"""
+        print("\n📋 剪贴板采集模式启动")
+        print(f"   监控间隔: {self.interval}秒")
+        print(f"   复制内容后自动采集...")
+        
+        import subprocess
+        
+        self.running = True
+        last_clip = ""
+        
+        while self.running:
+            # 获取剪贴板内容
+            result = subprocess.run(
+                ["pbpaste"],
+                capture_output=True, text=True
+            )
+            current = result.stdout.strip()
+            
+            if current and current != last_clip:
+                print(f"\n  📋 检测到新内容 ({len(current)}字)")
+                self.send_text(current, "clipboard")
+                last_clip = current
+            
+            time.sleep(self.interval)
+        
+        self.cleanup()
+    
+    def cleanup(self):
+        """清理"""
+        if self.ws:
+            try:
+                self.ws.close()
+            except:
+                pass
+
+
+# ============================================================
+# 主入口
+# ============================================================
+
+def main():
+    parser = argparse.ArgumentParser(description="语料采集 Mac 客户端")
+    parser.add_argument("--token", required=True, help="Gitea Access Token")
+    parser.add_argument("--server", default=SERVER_URL, help=f"服务端地址 (默认 {SERVER_URL})")
+    parser.add_argument("--mode", choices=["auto", "scroll", "clipboard"], default="auto",
+                        help="采集模式: auto=实时, scroll=滚屏, clipboard=剪贴板")
+    parser.add_argument("--interval", type=float, default=0,
+                        help="采集间隔秒数 (默认 auto=5, scroll=3, clipboard=2)")
+    
+    args = parser.parse_args()
+    
+    # 默认间隔
+    if args.interval <= 0:
+        intervals = {"auto": 5.0, "scroll": 3.0, "clipboard": 2.0}
+        args.interval = intervals[args.mode]
+    
+    print(f"\n{'='*50}")
+    print(f"🧠 语料采集 Mac 客户端 v{VERSION}")
+    print(f"{'='*50}")
+    print(f"  模式: {args.mode}")
+    print(f"  服务端: {args.server}")
+    print(f"  间隔: {args.interval}s")
+    print(f"{'='*50}\n")
+    
+    collector = CorpusCollector(args.token, args.server, args.interval)
+    
+    # 连接服务端
+    print("🔄 连接服务端...")
+    if not collector.connect_ws():
+        print("  ⚠️ WebSocket连接失败，将使用HTTP回退")
+        print("  请确保服务端已启动: python3 server.py")
+        print(f"  服务端地址: {args.server}")
+    
+    try:
+        if args.mode == "auto":
+            collector.run_auto_mode()
+        elif args.mode == "scroll":
+            collector.run_scroll_mode()
+        elif args.mode == "clipboard":
+            collector.run_clipboard_mode()
+    except KeyboardInterrupt:
+        print("\n\n⏹️ 用户停止")
+        collector.cleanup()
+    
+    print("\n👋 采集结束")
+
+
+if __name__ == "__main__":
+    main()

server/corpus-agent/requirements.txt

@@ -0,0 +1,5 @@
+fastapi>=0.104.0
+uvicorn[standard]>=0.24.0
+websocket-client>=1.6.0
+pyautogui>=0.9.53
+Pillow>=10.0.0

server/corpus-agent/server.py

@@ -0,0 +1,505 @@
+"""
+语料采集系统 · 服务端
+====================
+FastAPI + Gitea OAuth + WebSocket实时采集
+"""
+import os
+import json
+import uuid
+import time
+import hashlib
+from pathlib import Path
+from typing import Optional, List
+from datetime import datetime, timedelta
+
+from fastapi import FastAPI, HTTPException, Depends, Query, WebSocket, WebSocketDisconnect
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import FileResponse, HTMLResponse, JSONResponse
+from fastapi.staticfiles import StaticFiles
+from pydantic import BaseModel
+import uvicorn
+
+# 导入引擎
+from engine import (
+    process_text_chunk, process_dialog_stream,
+    desensitize, is_valuable, classify_content,
+    to_chatml, format_sft_jsonl, generate_corpus_id
+)
+
+# ============================================================
+# 配置
+# ============================================================
+
+class Settings:
+    APP_NAME = "语料采集系统 Corpus Agent"
+    VERSION = "1.0.0"
+    
+    # 存储路径
+    DATA_DIR = Path(os.environ.get("CORPUS_DATA_DIR", "./data"))
+    USERS_DIR = DATA_DIR / "users"
+    
+    # Gitea OAuth
+    GITEA_URL = os.environ.get("GITEA_URL", "https://guanghulab.com")
+    GITEA_CLIENT_ID = os.environ.get("GITEA_CLIENT_ID", "")
+    GITEA_CLIENT_SECRET = os.environ.get("GITEA_CLIENT_SECRET", "")
+    
+    # Session
+    SESSION_SECRET = os.environ.get("SESSION_SECRET", "corpus-agent-secret-key")
+    SESSION_EXPIRE_HOURS = 24
+    
+    # 服务器
+    HOST = os.environ.get("CORPUS_HOST", "0.0.0.0")
+    PORT = int(os.environ.get("CORPUS_PORT", "8084"))
+
+settings = Settings()
+
+# ============================================================
+# 数据模型
+# ============================================================
+
+class TextChunk(BaseModel):
+    text: str
+    source: str = "manual"
+    session_id: Optional[str] = None
+
+class DialogMessage(BaseModel):
+    role: str
+    content: str
+    source: str = "unknown"
+
+class DialogBatch(BaseModel):
+    messages: List[DialogMessage]
+    session_id: Optional[str] = None
+
+class CorpusSave(BaseModel):
+    samples: List[dict]
+    filename: str = "corpus"
+
+class LoginRequest(BaseModel):
+    gitea_token: str
+
+class SessionData(BaseModel):
+    username: str
+    login_time: float
+    exprire_at: float
+
+# ============================================================
+# App
+# ============================================================
+
+app = FastAPI(title=settings.APP_NAME, version=settings.VERSION)
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+# 静态文件
+static_dir = Path(__file__).parent / "static"
+static_dir.mkdir(exist_ok=True)
+app.mount("/static", StaticFiles(directory=str(static_dir)), name="static")
+
+# ============================================================
+# Session管理（简化版，生产环境改为Redis）
+# ============================================================
+
+sessions = {}  # token -> SessionData
+
+def create_session(username: str) -> str:
+    token = hashlib.sha256(f"{username}{time.time()}{uuid.uuid4()}".encode()).hexdigest()[:32]
+    sessions[token] = SessionData(
+        username=username,
+        login_time=time.time(),
+        exprire_at=time.time() + settings.SESSION_EXPIRE_HOURS * 3600,
+    )
+    return token
+
+def get_user_from_token(token: str) -> Optional[str]:
+    if token in sessions:
+        sd = sessions[token]
+        if time.time() < sd.exprire_at:
+            return sd.username
+        del sessions[token]
+    return None
+
+# ============================================================
+# 用户数据管理
+# ============================================================
+
+def get_user_dir(username: str) -> Path:
+    path = settings.USERS_DIR / username
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+def load_corpus(username: str) -> list:
+    path = get_user_dir(username) / "corpus.jsonl"
+    if not path.exists():
+        return []
+    samples = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                try:
+                    samples.append(json.loads(line))
+                except:
+                    pass
+    return samples
+
+def append_corpus(username: str, samples: list):
+    path = get_user_dir(username) / "corpus.jsonl"
+    with open(path, "a", encoding="utf-8") as f:
+        for s in samples:
+            f.write(json.dumps(s, ensure_ascii=False) + "\n")
+
+def save_corpus_snapshot(username: str, filename: str) -> Path:
+    """保存快照文件"""
+    samples = load_corpus(username)
+    snapshot_dir = get_user_dir(username) / "snapshots"
+    snapshot_dir.mkdir(exist_ok=True)
+    
+    ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+    path = snapshot_dir / f"{filename}_{ts}.jsonl"
+    
+    with open(path, "w", encoding="utf-8") as f:
+        for s in samples:
+            f.write(json.dumps(s, ensure_ascii=False) + "\n")
+    
+    return path
+
+# ============================================================
+# API路由
+# ============================================================
+
+@app.get("/")
+async def root():
+    return HTMLResponse(open(static_dir / "index.html").read())
+
+@app.get("/api/health")
+async def health():
+    return {"status": "ok", "app": settings.APP_NAME, "version": settings.VERSION}
+
+# --- 登录 ---
+
+@app.post("/api/auth/login")
+async def login(req: LoginRequest):
+    """使用Gitea Access Token登录"""
+    # 通过Gitea API验证token
+    import urllib.request
+    try:
+        gitea_req = urllib.request.Request(
+            f"{settings.GITEA_URL}/api/v1/user",
+            headers={"Authorization": f"token {req.gitea_token}"}
+        )
+        with urllib.request.urlopen(gitea_req, timeout=10) as resp:
+            user_data = json.loads(resp.read().decode())
+            username = user_data.get("login", "unknown")
+    except Exception as e:
+        # 降级：允许直接使用用户名登录（开发模式）
+        username = req.gitea_token.split(":")[0] if ":" in req.gitea_token else req.gitea_token
+    
+    token = create_session(username)
+    
+    # 统计现有语料
+    corpus = load_corpus(username)
+    
+    return {
+        "ok": True,
+        "token": token,
+        "username": username,
+        "stats": {
+            "total_samples": len(corpus),
+            "total_chars": sum(len(json.dumps(s, ensure_ascii=False)) for s in corpus),
+        }
+    }
+
+@app.get("/api/auth/check")
+async def check_session(token: str = Query(...)):
+    username = get_user_from_token(token)
+    if not username:
+        raise HTTPException(401, "登录已过期")
+    return {"ok": True, "username": username}
+
+# --- 语料操作 ---
+
+@app.post("/api/corpus/collect")
+async def collect_chunk(chunk: TextChunk, token: str = Query(...)):
+    """单段文本采集"""
+    username = get_user_from_token(token)
+    if not username:
+        raise HTTPException(401, "登录已过期")
+    
+    # 引擎处理
+    results = process_text_chunk(chunk.text, source=chunk.source)
+    if not results:
+        return {"ok": True, "collected": 0, "reason": "no_valuable_content"}
+    
+    # 添加时间戳和归属
+    ts = datetime.now().isoformat()
+    for r in results:
+        r["timestamp"] = ts
+        r["user"] = username
+    
+    # 转为ChatML格式存储
+    sft_samples = []
+    for r in results:
+        if "user" in r and "assistant" in r:
+            sft_samples.append(to_chatml(r["user"], r["assistant"]))
+        else:
+            # 单条文本也存起来
+            sft_samples.append({
+                "messages": [{"role": "user", "content": r["text"]}],
+                "source": r["source"],
+                "tags": r.get("tags", []),
+                "corpus_id": r.get("corpus_id", generate_corpus_id(r["text"])),
+                "collected_at": ts,
+            })
+    
+    append_corpus(username, sft_samples)
+    
+    return {
+        "ok": True,
+        "collected": len(sft_samples),
+        "preview": sft_samples[:3],
+        "stats": {"total": len(load_corpus(username))},
+    }
+
+@app.post("/api/corpus/batch")
+async def collect_batch(batch: DialogBatch, token: str = Query(...)):
+    """批量对话采集"""
+    username = get_user_from_token(token)
+    if not username:
+        raise HTTPException(401, "登录已过期")
+    
+    msgs = [{"role": m.role, "content": m.content, "source": m.source} for m in batch.messages]
+    result = process_dialog_stream(msgs)
+    
+    if result["stats"]["valuable_pairs"] > 0:
+        # 添加用户归属
+        ts = datetime.now().isoformat()
+        for p in result["pairs"]:
+            p["collected_at"] = ts
+            p["user"] = username
+        
+        append_corpus(username, result["pairs"])
+    
+    return {
+        "ok": True,
+        **result["stats"],
+    }
+
+@app.get("/api/corpus/list")
+async def list_corpus(token: str = Query(...), page: int = Query(1), size: int = Query(50)):
+    """查看已采集语料"""
+    username = get_user_from_token(token)
+    if not username:
+        raise HTTPException(401, "登录已过期")
+    
+    all_samples = load_corpus(username)
+    total = len(all_samples)
+    
+    # 分页
+    start = (page - 1) * size
+    end = start + size
+    page_samples = all_samples[start:end]
+    
+    return {
+        "ok": True,
+        "total": total,
+        "page": page,
+        "size": size,
+        "samples": page_samples,
+    }
+
+@app.get("/api/corpus/export")
+async def export_corpus(token: str = Query(...), format: str = "jsonl"):
+    """导出语料文件"""
+    username = get_user_from_token(token)
+    if not username:
+        raise HTTPException(401, "登录已过期")
+    
+    if format == "jsonl":
+        path = save_corpus_snapshot(username, "export")
+        return FileResponse(
+            path,
+            media_type="application/octet-stream",
+            filename=f"corpus_{username}_{datetime.now().strftime('%Y%m%d')}.jsonl"
+        )
+    
+    raise HTTPException(400, f"不支持的格式: {format}")
+
+@app.post("/api/corpus/preview")
+async def preview_text(chunk: TextChunk):
+    """预览一段文本的处理结果（无需登录）"""
+    from engine import preview as engine_preview
+    return engine_preview(chunk.text)
+
+@app.get("/api/corpus/stats")
+async def corpus_stats(token: str = Query(...)):
+    """语料统计"""
+    username = get_user_from_token(token)
+    if not username:
+        raise HTTPException(401, "登录已过期")
+    
+    samples = load_corpus(username)
+    
+    # 按来源统计
+    by_source = {}
+    # 按标签统计
+    by_tag = {}
+    total_chars = 0
+    
+    for s in samples:
+        msgs = s.get("messages", [])
+        for m in msgs:
+            total_chars += len(m.get("content", ""))
+        
+        source = s.get("source", "unknown")
+        by_source[source] = by_source.get(source, 0) + 1
+        
+        for tag in s.get("tags", []):
+            by_tag[tag] = by_tag.get(tag, 0) + 1
+    
+    return {
+        "ok": True,
+        "total_samples": len(samples),
+        "total_chars": total_chars,
+        "by_source": by_source,
+        "by_tag": by_tag,
+    }
+
+# --- 统计数据 ---
+
+@app.get("/api/system/stats")
+async def system_stats():
+    """系统统计（公开）"""
+    total_users = len(list(settings.USERS_DIR.iterdir())) if settings.USERS_DIR.exists() else 0
+    total_samples = 0
+    for user_dir in settings.USERS_DIR.iterdir() if settings.USERS_DIR.exists() else []:
+        if user_dir.is_dir():
+            total_samples += len(load_corpus(user_dir.name))
+    
+    return {
+        "total_users": total_users,
+        "total_samples": total_samples,
+        "active_sessions": len(sessions),
+    }
+
+# ============================================================
+# WebSocket实时采集（Mac客户端用）
+# ============================================================
+
+class ConnectionManager:
+    def __init__(self):
+        self.active_connections: dict[str, list[WebSocket]] = {}
+    
+    async def connect(self, websocket: WebSocket, username: str):
+        await websocket.accept()
+        if username not in self.active_connections:
+            self.active_connections[username] = []
+        self.active_connections[username].append(websocket)
+    
+    def disconnect(self, websocket: WebSocket, username: str):
+        if username in self.active_connections:
+            self.active_connections[username] = [ws for ws in self.active_connections[username] if ws != websocket]
+            if not self.active_connections[username]:
+                del self.active_connections[username]
+    
+    async def send_to_user(self, username: str, message: dict):
+        if username in self.active_connections:
+            for ws in self.active_connections[username]:
+                try:
+                    await ws.send_json(message)
+                except:
+                    pass
+
+manager = ConnectionManager()
+
+@app.websocket("/ws/collect")
+async def websocket_collect(websocket: WebSocket, token: str = Query(...)):
+    username = get_user_from_token(token)
+    if not username:
+        await websocket.close(code=4001, reason="未登录")
+        return
+    
+    await manager.connect(websocket, username)
+    
+    try:
+        while True:
+            data = await websocket.receive_json()
+            
+            # 支持多种消息类型
+            msg_type = data.get("type", "text")
+            
+            if msg_type == "text":
+                chunk = TextChunk(text=data.get("text", ""), source=data.get("source", "screen_capture"))
+                results = process_text_chunk(chunk.text, source=chunk.source)
+                if results:
+                    ts = datetime.now().isoformat()
+                    for r in results:
+                        r["timestamp"] = ts
+                        r["user"] = username
+                    
+                    sft_samples = []
+                    for r in results:
+                        if "user" in r and "assistant" in r:
+                            sft_samples.append(to_chatml(r["user"], r["assistant"]))
+                        else:
+                            sft_samples.append({
+                                "messages": [{"role": "user", "content": r["text"]}],
+                                "source": r["source"],
+                                "tags": r.get("tags", []),
+                                "corpus_id": r.get("corpus_id", ""),
+                                "collected_at": ts,
+                            })
+                    
+                    if sft_samples:
+                        append_corpus(username, sft_samples)
+                    
+                    await websocket.send_json({
+                        "type": "result",
+                        "collected": len(sft_samples),
+                        "valuable": len(results) > 0,
+                        "preview": results[:2],
+                    })
+                else:
+                    await websocket.send_json({"type": "result", "collected": 0, "valuable": False})
+            
+            elif msg_type == "batch":
+                messages = data.get("messages", [])
+                result = process_dialog_stream(messages)
+                if result["stats"]["valuable_pairs"] > 0:
+                    ts = datetime.now().isoformat()
+                    for p in result["pairs"]:
+                        p["collected_at"] = ts
+                        p["user"] = username
+                    append_corpus(username, result["pairs"])
+                
+                await websocket.send_json({
+                    "type": "batch_result",
+                    **result["stats"],
+                })
+            
+            elif msg_type == "ping":
+                await websocket.send_json({"type": "pong"})
+                
+    except WebSocketDisconnect:
+        manager.disconnect(websocket, username)
+    except Exception as e:
+        manager.disconnect(websocket, username)
+
+# ============================================================
+# 启动
+# ============================================================
+
+if __name__ == "__main__":
+    print(f"🚀 {settings.APP_NAME} v{settings.VERSION}")
+    print(f"📂 数据目录: {settings.DATA_DIR.absolute()}")
+    print(f"🌐 服务地址: http://{settings.HOST}:{settings.PORT}")
+    
+    settings.USERS_DIR.mkdir(parents=True, exist_ok=True)
+    
+    uvicorn.run(app, host=settings.HOST, port=settings.PORT)

server/corpus-agent/static/index.html

@@ -0,0 +1,345 @@
+<!DOCTYPE html>
+<html lang="zh-CN">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>语料采集系统 · Corpus Agent</title>
+<style>
+  * { margin: 0; padding: 0; box-sizing: border-box; }
+  body { font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif; background: #f0f2f5; color: #333; min-height: 100vh; }
+  .container { max-width: 1000px; margin: 0 auto; padding: 20px; }
+  
+  /* Header */
+  .header { background: linear-gradient(135deg, #1a1a2e 0%, #16213e 100%); color: white; padding: 40px 20px; text-align: center; border-radius: 12px; margin-bottom: 24px; }
+  .header h1 { font-size: 28px; margin-bottom: 8px; }
+  .header p { color: #8899aa; font-size: 14px; }
+  
+  /* Cards */
+  .card { background: white; border-radius: 12px; padding: 24px; margin-bottom: 20px; box-shadow: 0 2px 8px rgba(0,0,0,0.06); }
+  .card h2 { font-size: 18px; margin-bottom: 16px; color: #1a1a2e; }
+  
+  /* Login / User area */
+  .login-box { text-align: center; padding: 40px; }
+  .login-box input { padding: 12px 16px; width: 300px; border: 2px solid #e0e0e0; border-radius: 8px; font-size: 14px; margin-bottom: 12px; }
+  .login-box input:focus { outline: none; border-color: #4a6cf7; }
+  .btn { padding: 10px 24px; border: none; border-radius: 8px; font-size: 14px; cursor: pointer; transition: 0.2s; font-weight: 500; }
+  .btn-primary { background: #4a6cf7; color: white; }
+  .btn-primary:hover { background: #3b5de7; }
+  .btn-success { background: #10b981; color: white; }
+  .btn-success:hover { background: #059669; }
+  .btn-danger { background: #ef4444; color: white; }
+  .btn-danger:hover { background: #dc2626; }
+  .btn-sm { padding: 6px 14px; font-size: 12px; }
+  
+  .user-info { display: flex; justify-content: space-between; align-items: center; }
+  .user-info .badge { background: #e8f0fe; color: #4a6cf7; padding: 4px 12px; border-radius: 20px; font-size: 13px; }
+  
+  /* Stats grid */
+  .stats-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(140px, 1fr)); gap: 12px; margin: 16px 0; }
+  .stat-card { background: #f8f9fc; border-radius: 8px; padding: 16px; text-align: center; }
+  .stat-card .num { font-size: 28px; font-weight: 700; color: #1a1a2e; }
+  .stat-card .label { font-size: 12px; color: #8899aa; margin-top: 4px; }
+  
+  /* Sample list */
+  .sample-list { max-height: 400px; overflow-y: auto; }
+  .sample-item { padding: 12px; border-bottom: 1px solid #f0f0f0; cursor: pointer; transition: 0.1s; }
+  .sample-item:hover { background: #f8f9fc; }
+  .sample-item:last-child { border: none; }
+  .sample-item .msg { font-size: 13px; color: #666; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+  .sample-item .meta { font-size: 11px; color: #999; margin-top: 4px; }
+  .sample-item .tag { display: inline-block; background: #e8f0fe; color: #4a6cf7; padding: 2px 8px; border-radius: 10px; font-size: 11px; margin-right: 4px; }
+  
+  /* Tabs */
+  .tabs { display: flex; gap: 8px; margin-bottom: 16px; }
+  .tab { padding: 8px 20px; border-radius: 8px; cursor: pointer; font-size: 14px; background: #f0f2f5; color: #666; transition: 0.2s; border: none; }
+  .tab.active { background: #4a6cf7; color: white; }
+  
+  .hidden { display: none; }
+  
+  /* Preview / input area */
+  .preview-area { background: #f8f9fc; border-radius: 8px; padding: 16px; margin: 12px 0; font-size: 13px; line-height: 1.6; max-height: 300px; overflow-y: auto; }
+  .preview-area .valuable { border-left: 3px solid #10b981; padding-left: 12px; margin: 8px 0; }
+  .preview-area .filtered { border-left: 3px solid #ef4444; padding-left: 12px; margin: 8px 0; opacity: 0.5; }
+  
+  /* Mac client download */
+  .download-section { text-align: center; padding: 20px; }
+  .download-section p { color: #666; font-size: 13px; margin: 8px 0; }
+  code { background: #f0f2f5; padding: 2px 6px; border-radius: 4px; font-size: 12px; }
+  
+  /* Footer */
+  .footer { text-align: center; color: #999; font-size: 12px; padding: 20px 0; }
+  
+  /* Empty state */
+  .empty-state { text-align: center; padding: 40px; color: #999; }
+  .empty-state .icon { font-size: 48px; margin-bottom: 12px; }
+</style>
+</head>
+<body>
+<div class="container" id="app">
+  <!-- Header -->
+  <div class="header">
+    <h1>🧠 语料采集系统</h1>
+    <p>Corpus Agent · 自动采集 · 脱敏 · 格式化</p>
+  </div>
+  
+  <!-- Login -->
+  <div class="card" id="login-section">
+    <div class="login-box">
+      <h2 style="margin-bottom:8px">登录</h2>
+      <p style="color:#666; font-size:13px; margin-bottom:20px">
+        使用 Gitea Access Token 登录
+      </p>
+      <input type="password" id="token-input" placeholder="输入 Gitea Access Token">
+      <br>
+      <button class="btn btn-primary" onclick="login()">登录</button>
+      <p style="color:#999; font-size:12px; margin-top:12px">
+        访问 <a href="https://guanghulab.com/user/settings/applications" target="_blank" style="color:#4a6cf7">Gitea → 设置 → 应用</a> 创建 Token
+      </p>
+    </div>
+  </div>
+  
+  <!-- Main (after login) -->
+  <div id="main-section" class="hidden">
+    <!-- User info -->
+    <div class="card">
+      <div class="user-info">
+        <div>
+          <strong id="display-name">用户</strong>
+          <span class="badge" id="sample-count">0 条语料</span>
+        </div>
+        <button class="btn btn-sm btn-danger" onclick="logout()">退出</button>
+      </div>
+      
+      <div class="stats-grid" id="stats-grid">
+        <div class="stat-card">
+          <div class="num" id="stat-total">0</div>
+          <div class="label">语料总数</div>
+        </div>
+        <div class="stat-card">
+          <div class="num" id="stat-chars">0</div>
+          <div class="label">总字符数</div>
+        </div>
+        <div class="stat-card">
+          <div class="num" id="stat-sources">0</div>
+          <div class="label">来源渠道</div>
+        </div>
+      </div>
+    </div>
+    
+    <!-- Tabs -->
+    <div class="tabs">
+      <button class="tab active" onclick="switchTab('collection')">📥 实时采集</button>
+      <button class="tab" onclick="switchTab('browse')">📚 语料浏览</button>
+      <button class="tab" onclick="switchTab('client')">💻 Mac 客户端</button>
+    </div>
+    
+    <!-- Tab: 实时采集 -->
+    <div class="card" id="tab-collection">
+      <h2>📥 实时采集</h2>
+      <p style="color:#666; font-size:13px; margin-bottom:12px">粘贴对话文本，引擎自动判断价值并脱敏存储</p>
+      <textarea id="input-text" style="width:100%;height:120px;padding:12px;border:2px solid #e0e0e0;border-radius:8px;font-size:13px;resize:vertical" placeholder="粘贴聊天内容..."></textarea>
+      <div style="margin-top:12px">
+        <button class="btn btn-primary" onclick="collectText()">采集</button>
+        <button class="btn btn-sm" style="margin-left:8px" onclick="document.getElementById('input-text').value=''">清空</button>
+      </div>
+      <div id="collect-result" class="hidden preview-area"></div>
+    </div>
+    
+    <!-- Tab: 语料浏览 -->
+    <div class="card hidden" id="tab-browse">
+      <div style="display:flex;justify-content:space-between;align-items:center;margin-bottom:16px">
+        <h2 style="margin:0">📚 语料库</h2>
+        <div>
+          <button class="btn btn-sm btn-success" onclick="exportCorpus()">📥 导出 JSONL</button>
+        </div>
+      </div>
+      <div id="sample-container">
+        <div class="empty-state">
+          <div class="icon">📭</div>
+          <p>还没有语料，开始采集吧</p>
+        </div>
+      </div>
+    </div>
+    
+    <!-- Tab: Mac 客户端 -->
+    <div class="card hidden" id="tab-client">
+      <div class="download-section">
+        <h2>💻 Mac 客户端</h2>
+        <p>在 Mac 上运行的桌面采集助手，自动截图→OCR→筛选→上传</p>
+        
+        <div style="text-align:left;max-width:600px;margin:20px auto">
+          <h3 style="font-size:14px;margin-bottom:12px">📋 安装步骤</h3>
+          <ol style="font-size:13px;color:#666;line-height:2">
+            <li>下载 <code>mac-corpus-agent.py</code> 到你的 Mac</li>
+            <li>安装依赖：<code>pip3 install pyautogui pillow websocket-client</code></li>
+            <li>运行：<code>python3 mac-corpus-agent.py --token YOUR_TOKEN</code></li>
+            <li>选择模式：<code>--mode auto</code>（实时）或 <code>--mode scroll</code>（滚屏）</li>
+          </ol>
+          
+          <h3 style="font-size:14px;margin:12px 0 8px">⚙️ 模式说明</h3>
+          <table style="width:100%;font-size:13px;border-collapse:collapse">
+            <tr style="border-bottom:1px solid #eee">
+              <td style="padding:8px"><code>auto</code> 实时模式</td>
+              <td style="padding:8px;color:#666">在后台运行，监控屏幕变化，自动采集</td>
+            </tr>
+            <tr style="border-bottom:1px solid #eee">
+              <td style="padding:8px"><code>scroll</code> 滚屏模式</td>
+              <td style="padding:8px;color:#666">自动滚屏采集历史对话（需手动滚到顶部）</td>
+            </tr>
+            <tr>
+              <td style="padding:8px"><code>clipboard</code> 剪贴板模式</td>
+              <td style="padding:8px;color:#666">监听剪贴板，自动采集复制的内容</td>
+            </tr>
+          </table>
+        </div>
+        
+        <p style="color:#999;font-size:12px;margin-top:20px">
+          Mac 客户端需要 macOS 12.0+，使用系统原生 OCR
+        </p>
+      </div>
+    </div>
+  </div>
+  
+  <!-- Footer -->
+  <div class="footer">
+    语料采集系统 v1.0 · 国作登字-2026-A-00037559
+  </div>
+</div>
+
+<script>
+let token = localStorage.getItem('corpus_token') || '';
+let username = '';
+const API = '';
+
+async function api(method, path, body) {
+  const url = `${API}${path}${path.includes('?') ? '&' : '?'}token=${token}`;
+  const opts = { method, headers: {'Content-Type': 'application/json'} };
+  if (body) opts.body = JSON.stringify(body);
+  const resp = await fetch(url, opts);
+  return resp.json();
+}
+
+// --- Login ---
+async function login() {
+  const t = document.getElementById('token-input').value.trim();
+  if (!t) return alert('请输入Token');
+  const r = await fetch(`${API}/api/auth/login`, {
+    method: 'POST',
+    headers: {'Content-Type': 'application/json'},
+    body: JSON.stringify({gitea_token: t})
+  });
+  const d = await r.json();
+  if (!d.ok) return alert('登录失败: ' + JSON.stringify(d));
+  token = d.token;
+  username = d.username;
+  localStorage.setItem('corpus_token', token);
+  document.getElementById('login-section').classList.add('hidden');
+  document.getElementById('main-section').classList.remove('hidden');
+  document.getElementById('display-name').textContent = username;
+  document.getElementById('sample-count').textContent = d.stats.total_samples + ' 条语料';
+  loadStats();
+  loadSamples();
+}
+
+function logout() {
+  token = '';
+  localStorage.removeItem('corpus_token');
+  document.getElementById('login-section').classList.remove('hidden');
+  document.getElementById('main-section').classList.add('hidden');
+}
+
+// Auto login on load
+if (token) {
+  api('GET', '/api/auth/check').then(d => {
+    if (d.ok) {
+      username = d.username;
+      document.getElementById('login-section').classList.add('hidden');
+      document.getElementById('main-section').classList.remove('hidden');
+      document.getElementById('display-name').textContent = username;
+      loadStats();
+      loadSamples();
+    } else {
+      localStorage.removeItem('corpus_token');
+    }
+  });
+}
+
+// --- Tabs ---
+function switchTab(name) {
+  document.querySelectorAll('.tab').forEach(t => t.classList.remove('active'));
+  document.querySelectorAll('[id^="tab-"]').forEach(t => t.classList.add('hidden'));
+  event.target.classList.add('active');
+  document.getElementById('tab-' + name).classList.remove('hidden');
+}
+
+// --- Collect ---
+async function collectText() {
+  const text = document.getElementById('input-text').value.trim();
+  if (!text) return;
+  const r = await api('POST', '/api/corpus/collect', {text, source: 'web_manual'});
+  const el = document.getElementById('collect-result');
+  el.classList.remove('hidden');
+  if (r.collected > 0) {
+    el.innerHTML = `<div class="valuable">✅ 采集了 ${r.collected} 条语料</div>
+      <pre style="font-size:12px;margin-top:8px;background:#f0f0f0;padding:8px;border-radius:4px;overflow-x:auto">${JSON.stringify(r.preview, null, 2)}</pre>`;
+    loadStats();
+  } else {
+    el.innerHTML = `<div class="filtered">⏭️ 无有价值内容（已过滤日常对话）</div>
+      <p style="font-size:12px;color:#999;margin-top:8px">提示：技术讨论、决策分析、踩坑记录等才会被收录</p>`;
+  }
+}
+
+// --- Stats ---
+async function loadStats() {
+  const r = await api('GET', '/api/corpus/stats');
+  if (!r.ok) return;
+  document.getElementById('stat-total').textContent = r.total_samples || 0;
+  document.getElementById('stat-chars').textContent = (r.total_chars || 0).toLocaleString();
+  document.getElementById('stat-sources').textContent = Object.keys(r.by_source || {}).length;
+  document.getElementById('sample-count').textContent = (r.total_samples || 0) + ' 条语料';
+}
+
+// --- Samples ---
+async function loadSamples() {
+  const r = await api('GET', '/api/corpus/list?page=1&size=50');
+  if (!r.ok) return;
+  const container = document.getElementById('sample-container');
+  if (r.samples.length === 0) {
+    container.innerHTML = `<div class="empty-state"><div class="icon">📭</div><p>还没有语料，开始采集吧</p></div>`;
+    return;
+  }
+  let html = '<div class="sample-list">';
+  for (const s of r.samples) {
+    const msgs = s.messages || [];
+    const first = msgs[0]?.content?.slice(0, 80) || '(空)';
+    const tags = (s.tags || []).map(t => `<span class="tag">${t}</span>`).join('');
+    const src = s.source || 'unknown';
+    html += `<div class="sample-item">
+      <div class="msg">${escapeHtml(first)}</div>
+      <div class="meta">${tags}<span style="margin-left:8px">来源: ${src}</span></div>
+    </div>`;
+  }
+  html += '</div>';
+  container.innerHTML = html;
+}
+
+// --- Export ---
+async function exportCorpus() {
+  const r = await fetch(`${API}/api/corpus/export?token=${token}&format=jsonl`);
+  if (r.ok) {
+    const blob = await r.blob();
+    const a = document.createElement('a');
+    a.href = URL.createObjectURL(blob);
+    a.download = `corpus_${username}_${new Date().toISOString().slice(0,10)}.jsonl`;
+    a.click();
+  }
+}
+
+function escapeHtml(s) {
+  const d = document.createElement('div');
+  d.textContent = s;
+  return d.innerHTML;
+}
+</script>
+</body>
+</html>