宜博 / lemonboss

const fs = require('fs');
const path = require('path');

const SYSTEM_AGENTS_DIR = path.resolve(__dirname, '../../system/agents');
const COMPANIES_DIR = path.resolve(__dirname, '../../data/companies');
const TASK_META_DIR = path.resolve(__dirname, '../../system/tasklog');

/**
 * 递归扫描并读取系统中所有 Agent 的会话，
 * 从 .tasks/ meta 和 conversation 模式中推断父子关系，构建任务树。
 */
class TaskService {

  getAllConversations() {
    const list = [];
    
    if (fs.existsSync(SYSTEM_AGENTS_DIR)) {
      const agents = fs.readdirSync(SYSTEM_AGENTS_DIR).filter(d => !d.startsWith('.'));
      for (const agentId of agents) {
        const convDir = path.join(SYSTEM_AGENTS_DIR, agentId, 'conversations');
        this._scanConversationsDir(convDir, 'system', agentId, list);
      }
    }

    if (fs.existsSync(COMPANIES_DIR)) {
      const dbNames = fs.readdirSync(COMPANIES_DIR).filter(d => !d.startsWith('.'));
      for (const companyId of dbNames) {
        if (companyId === '_archived_agents') continue;
        const agentsDir = path.join(COMPANIES_DIR, companyId, 'agents');
        if (fs.existsSync(agentsDir)) {
          const agents = fs.readdirSync(agentsDir).filter(d => !d.startsWith('.'));
          for (const agentId of agents) {
            const convDir = path.join(agentsDir, agentId, 'conversations');
            this._scanConversationsDir(convDir, companyId, agentId, list);
          }
        }
      }
    }

    return list;
  }

  _loadTaskMetas() {
    const metas = new Map();
    if (!fs.existsSync(TASK_META_DIR)) return metas;
    const files = fs.readdirSync(TASK_META_DIR).filter(f => f.endsWith('.json'));
    for (const f of files) {
      try {
        const tm = JSON.parse(fs.readFileSync(path.join(TASK_META_DIR, f), 'utf-8'));
        if (tm.taskId) metas.set(tm.taskId, tm);
      } catch(e) {}
    }
    return metas;
  }

  /**
   * 从 Conversation 聚合 Task 列表，并通过多种手段推断父子关系：
   * 1) .tasks/ meta 中的显式 parentTaskId
   * 2) conversationId 模式推断（task-T-xxx 从 tpl-train-* 派出，tpl-train 从 conv-* 派出）
   * 3) title 中可解析的 task reference（[TASK from xxx]）
   */
  getAggregatedTasks() {
    const convs = this.getAllConversations();
    const taskMetas = this._loadTaskMetas();
    const taskMap = new Map();

    // 第一遍：建立所有 task group
    for (const c of convs) {
      let taskId = this._extractTaskId(c.conversationId);

      if (!taskMap.has(taskId)) {
        const meta = taskMetas.get(taskId);
        taskMap.set(taskId, {
          taskId,
          title: c.meta.title || taskId,
          createdAt: c.meta.createdAt || new Date().toISOString(),
          updatedAt: c.meta.updatedAt || c.meta.createdAt,
          status: meta?.status || c.meta.status || 'unknown',
          from: meta?.from || null,
          to: meta?.to || null,
          parentTaskId: meta?.parentTaskId || null,
          agents: new Set(),
          conversations: [],
          children: [],
        });
      }

      const t = taskMap.get(taskId);
      t.agents.add(`${c.companyId}:${c.agentId}`);
      t.conversations.push(c);
      
      if (c.meta.createdAt && new Date(c.meta.createdAt) < new Date(t.createdAt)) t.createdAt = c.meta.createdAt;
      if (c.meta.updatedAt && new Date(c.meta.updatedAt) > new Date(t.updatedAt)) t.updatedAt = c.meta.updatedAt;
    }

    // 第二遍：推断父子关系
    this._inferParentChild(taskMap, taskMetas);

    // 第三遍：构建 children 列表
    for (const t of taskMap.values()) {
      if (t.parentTaskId && taskMap.has(t.parentTaskId)) {
        const parent = taskMap.get(t.parentTaskId);
        if (!parent.children.includes(t.taskId)) {
          parent.children.push(t.taskId);
        }
      }
    }

    const tasks = Array.from(taskMap.values()).map(t => ({
      ...t,
      agents: Array.from(t.agents),
      isRoot: !t.parentTaskId || !taskMap.has(t.parentTaskId),
      conversationCount: t.conversations.length,
    }));
    
    tasks.sort((a, b) => new Date(b.createdAt).getTime() - new Date(a.createdAt).getTime());
    return tasks;
  }

  /**
   * 从 conversationId 提取逻辑 taskId
   */
  _extractTaskId(convId) {
    // "task-T-xxx" / "callback-T-xxx" → "T-xxx"
    if (convId.includes('-T-')) {
      const match = convId.match(/(T-[a-z0-9]+)/);
      if (match) return match[1];
    }
    // 其他 (conv-xxx, tpl-train-xxx) 保持原样
    return convId;
  }

  /**
   * 多策略推断父子关系：
   * 
   * 策略1: .tasks/ meta 中的 parentTaskId（直接有的情况）
   * 策略2: 通过 "from" agent 的 活跃会话 推断
   *   - T-xxx 的 meta.from = "innovator" 
   *   - innovator 同时有一个 tpl-train-* 会话（时间上包含 T-xxx）
   *   → tpl-train-* 就是 T-xxx 的父任务
   * 策略3: 通过 master conv 触发 tpl-train 推断
   *   - conv-xxx (master) 的时间 < tpl-train-xxx 的时间 且 title 包含相同模板名
   *   → conv-xxx 是 tpl-train-xxx 的父任务
   */
  _inferParentChild(taskMap, taskMetas) {
    const allTasks = Array.from(taskMap.values());
    
    // 策略1: 已有 parentTaskId (来自 .tasks/ meta)，已在构建时设置
    
    // 策略2: T-xxx 任务，推断父任务为 from agent 的活跃 tpl-train/task 会话
    for (const t of allTasks) {
      if (t.parentTaskId) continue;
      if (!t.taskId.startsWith('T-')) continue;
      
      const meta = taskMetas.get(t.taskId);
      if (!meta) continue;
      
      const fromAgent = meta.from;
      const taskCreatedAt = new Date(meta.createdAt);
      
      // 找 fromAgent 的其他会话，时间上在这个 task 之前开始且之后结束/仍活跃
      let bestParent = null;
      let bestTimeDiff = Infinity;
      
      for (const candidate of allTasks) {
        if (candidate.taskId === t.taskId) continue;
        if (candidate.taskId.startsWith('T-') && !taskMetas.has(candidate.taskId)) continue;
        
        const hasAgent = candidate.conversations.some(c => c.agentId === fromAgent);
        if (!hasAgent) continue;
        
        const candidateCreated = new Date(candidate.createdAt);
        const candidateUpdated = new Date(candidate.updatedAt);
        
        if (candidateCreated < taskCreatedAt && candidateUpdated >= taskCreatedAt) {
          const diff = taskCreatedAt - candidateCreated;
          if (diff < bestTimeDiff) {
            bestTimeDiff = diff;
            bestParent = candidate;
          }
        }
      }
      
      if (bestParent) {
        t.parentTaskId = bestParent.taskId;
        // console.log(`[TaskService] 🔗 Inferred parent: ${t.taskId} → ${bestParent.taskId}`);
      }
    }

    // 策略4: T-xxx from master，但 master 的 conv-* 已经结束的情况
    // 这种情况发生在：master 原始对话结束后，通过 callback 链路继续派发新任务
    // 例如 conv-xxx → tpl-train → T-a (callback → master → 新 dispatch T-b)
    // 此时 T-b 应该关联到 conv-xxx（同一个用户请求）
    for (const t of allTasks) {
      if (t.parentTaskId) continue;
      if (!t.taskId.startsWith('T-')) continue;
      
      const meta = taskMetas.get(t.taskId);
      if (!meta || meta.from !== 'master') continue;
      
      const taskCreatedAt = new Date(meta.createdAt);
      
      // 找最近的 master conv-* 会话，窗口放宽到 30 分钟
      let bestConv = null;
      let bestDiff = 30 * 60 * 1000;
      
      for (const candidate of allTasks) {
        if (!candidate.taskId.startsWith('conv-')) continue;
        if (!candidate.conversations.some(c => c.agentId === 'master')) continue;
        
        const cCreated = new Date(candidate.createdAt);
        const diff = taskCreatedAt - cCreated;
        
        if (diff > 0 && diff < bestDiff) {
          bestDiff = diff;
          bestConv = candidate;
        }
      }
      
      if (bestConv) {
        t.parentTaskId = bestConv.taskId;
        // console.log(`[TaskService] 🔗 Inferred parent (master→T): ${t.taskId} → ${bestConv.taskId}`);
      }
    }

    // 策略3: tpl-train-* / tpl-gen-* / task-* 与 conv-* (master) 的关联
    // master 的 conv-xxx 在时间上紧接着触发了 tpl-train-xxx / tpl-gen-xxx
    for (const t of allTasks) {
      if (t.parentTaskId) continue;
      
      const isTemplateTask = t.taskId.startsWith('tpl-train-') || t.taskId.startsWith('train-') || t.taskId.startsWith('tpl-gen-');
      if (!isTemplateTask) continue;
      
      const tCreated = new Date(t.createdAt);
      
      // 找 master 的 conv-xxx，时间上在此刻之前且间隔小于 15 分钟
      let bestConv = null;
      let bestDiff = 15 * 60 * 1000; // 15分钟窗口
      
      for (const candidate of allTasks) {
        if (!candidate.taskId.startsWith('conv-')) continue;
        // 必须有 master 参与
        if (!candidate.conversations.some(c => c.agentId === 'master')) continue;
        
        const cCreated = new Date(candidate.createdAt);
        const diff = tCreated - cCreated;
        
        if (diff > 0 && diff < bestDiff) {
          // 额外检查：title 是否有模板名关联
          const templateName = t.taskId.replace(/^(tpl-train-|train-|tpl-gen-)/, '').replace(/-[a-z0-9]+$/, '');
          if (candidate.title && candidate.title.toLowerCase().includes(templateName.replace(/_/g, ' ').split(' ')[0])) {
            bestDiff = diff;
            bestConv = candidate;
          } else if (!bestConv) {
            // 如果没有更好的匹配，用时间最近的
            bestDiff = diff;
            bestConv = candidate;
          }
        }
      }
      
      if (bestConv) {
        t.parentTaskId = bestConv.taskId;
        // console.log(`[TaskService] 🔗 Inferred parent (conv→tpl): ${t.taskId} → ${bestConv.taskId}`);
      }
    }
  }

  getTaskTimeline(filterTaskId, includeChildren = true) {
    const tasks = this.getAggregatedTasks();
    const taskGroup = tasks.find(t => t.taskId === filterTaskId);
    if (!taskGroup) return [];

    const taskIds = new Set([filterTaskId]);
    if (includeChildren) {
      this._collectDescendants(filterTaskId, tasks, taskIds);
    }

    const relevantConvs = [];
    for (const t of tasks) {
      if (taskIds.has(t.taskId)) {
        relevantConvs.push(...t.conversations);
      }
    }

    let timeline = [];
    
    for (const conv of relevantConvs) {
      const logPath = path.join(path.dirname(conv.histPath), 'messages.log');
      
      if (fs.existsSync(logPath)) {
        try {
          // 限制最大读取 512KB，超出则只读尾部
          const MAX_LOG_READ = 512 * 1024;
          const stat = fs.statSync(logPath);
          let logContent;
          if (stat.size > MAX_LOG_READ) {
            const fd = fs.openSync(logPath, 'r');
            const buf = Buffer.alloc(MAX_LOG_READ);
            fs.readSync(fd, buf, 0, MAX_LOG_READ, stat.size - MAX_LOG_READ);
            fs.closeSync(fd);
            logContent = buf.toString('utf-8');
          } else {
            logContent = fs.readFileSync(logPath, 'utf-8');
          }
          const blocks = this._parseMessagesLog(logContent, conv);
          timeline.push(...blocks);
        } catch (e) {
          console.warn('Error parsing messages.log', logPath, e.message);
        }
      } else if (fs.existsSync(conv.histPath)) {
        try {
          const history = JSON.parse(fs.readFileSync(conv.histPath, 'utf-8'));
          history.forEach(msg => {
            if (!msg.content && !msg.tool_calls) return;
            timeline.push({
              key: `${conv.conversationId}-${msg.messageId || Math.random()}`,
              companyId: conv.companyId,
              agentId: conv.agentId,
              conversationId: conv.conversationId,
              role: msg.role,
              content: msg.content,
              timestamp: msg.timestamp || conv.meta.updatedAt,
            });
          });
        } catch (e) {
          console.warn('Error parsing history.json', conv.histPath);
        }
      }
    }

    timeline.sort((a, b) => new Date(a.timestamp).getTime() - new Date(b.timestamp).getTime());
    return timeline;
  }

  _collectDescendants(taskId, allTasks, collected) {
    const task = allTasks.find(t => t.taskId === taskId);
    if (!task) return;
    for (const childId of task.children) {
      if (!collected.has(childId)) {
        collected.add(childId);
        this._collectDescendants(childId, allTasks, collected);
      }
    }
  }

  _parseMessagesLog(logContent, conv) {
    const blocks = [];
    const blockRegex = /\[(\d{4}-\d{2}-\d{2}T[\d:.]+Z?)\]\s*\[(USER|AGENT|SYSTEM|TOOL)\]\n/g;
    
    const splits = [];
    let match;
    while ((match = blockRegex.exec(logContent)) !== null) {
      splits.push({
        timestamp: match[1],
        role: match[2].toLowerCase(),
        startIndex: match.index + match[0].length,
      });
    }

    for (let i = 0; i < splits.length; i++) {
      const s = splits[i];
      const endIndex = (i + 1 < splits.length) ? splits[i + 1].startIndex - splits[i + 1].timestamp.length - 10 : logContent.length;
      const content = logContent.substring(s.startIndex, endIndex).trim();
      if (!content) continue;

      blocks.push({
        key: `${conv.conversationId}-log-${i}`,
        companyId: conv.companyId,
        agentId: conv.agentId,
        conversationId: conv.conversationId,
        role: s.role === 'agent' ? 'assistant' : s.role,
        content,
        timestamp: s.timestamp.endsWith('Z') ? s.timestamp : s.timestamp + 'Z',
      });
    }
    return blocks;
  }

  _scanConversationsDir(convDir, companyId, agentId, list) {
    if (!fs.existsSync(convDir)) return;
    const items = fs.readdirSync(convDir, { withFileTypes: true });
    for (const item of items) {
      if (item.isDirectory()) {
        const metaPath = path.join(convDir, item.name, 'meta.json');
        const histPath = path.join(convDir, item.name, 'history.json');
        if (fs.existsSync(metaPath)) {
          let meta = {};
          try { meta = JSON.parse(fs.readFileSync(metaPath, 'utf-8')); } catch(e){}
          list.push({ companyId, agentId, conversationId: item.name, meta, histPath });
        }
      }
    }
  }
}

module.exports = { TaskService };