perf: reuse existing session for task callbacks to save ~30K tokens/callback

When a task completes and triggers a callback to the parent agent, previously it always created a new convId (callback-{taskId}), spawning a fresh CLI process with full system prompt re-injection. Now it checks the sessions pool for an existing connected session of the target agent and reuses its convId. This avoids creating a new CLI process and saves the system prompt overhead (~30K tokens for master). Falls back to creating new session if no active one exists.

perf: reuse existing session for task callbacks to save ~30K tokens/callback
When a task completes and triggers a callback to the parent agent, previously it always created a new convId (callback-{taskId}), spawning a fresh CLI process with full system prompt re-injection. Now it checks the sessions pool for an existing connected session of the target agent and reuses its convId. This avoids creating a new CLI process and saves the system prompt overhead (~30K tokens for master). Falls back to creating new session if no active one exists.
yibo
1 parent 598bef1e
Showing 1 changed file with 25 additions and 1 deletions
server/routes/agent-chat.js
@@ -665,10 +665,25 @@ async function getOrCreateProtocol(agentId, convId) {
  
   const protocol = new ClaudeCliProtocol({ name: identity.name }, cliOptions);
  
+  const releaseLock = () => {
+    const lock = agentChatLocks.get(agentId);
+    if (lock && lock.convId === convId) {
+      agentChatLocks.delete(agentId);
+      console.log(`[AgentChat] 🔓 ${identity.name} 锁已释放 (conv: ${convId})`);
+    }
+  };
+
   // 绑定错误监听器，防止 EventEmitter 抛出未捕获异常导致整个 Node 进程奔溃 (例如 spawn claude ENOENT)
   protocol.on('error', (err) => {
     console.error(`[AgentChat] 🚨 ${identity.name} 协议级异常:`, err.message);
     sessions.delete(key); // 下次提问将重建进程
+    releaseLock();
+  });
+
+  protocol.on('disconnected', ({ code, signal }) => {
+    sessions.delete(key);
+    releaseLock();
+    console.log(`[AgentChat] 💤 ${identity.name} 会话断开 (code=${code}, signal=${signal})`);
   });
  
   await protocol.connect(null, agentDir);
@@ -1733,7 +1748,16 @@ router.post(&#39;/task/report&#39;, (req, res) =&gt; {
         try {
           // 读取 task 元数据获取上下文
           const taskMeta = loadTaskMeta(taskId);
-          const callbackConvId = `callback-${taskId}`;
+
+          // 💰 Token 优化：尝试复用目标 Agent 的已有 session，避免创建新 CLI 进程
+          let callbackConvId = `callback-${taskId}`;
+          for (const [key, entry] of sessions.entries()) {
+            if (key.startsWith(`${to}:`) && entry.protocol.isConnected) {
+              callbackConvId = entry.convId;
+              console.log(`[TaskCallback] ♻️ 复用 ${to} 已有 session: ${callbackConvId}`);
+              break;
+            }
+          }
  
           const statusText = status === 'done' ? '✅ 已完成' : '❌ 失败';
           const callbackMessage = `[TASK_CALLBACK]\n任务编号: ${taskId}\n执行者: ${from}\n状态: ${statusText}\n摘要: ${summary || '(无摘要)'}\n产出文件: ${outputs.length > 0 ? outputs.join(', ') : '(无)'}\n${taskMeta ? `原始任务: ${taskMeta.task}` : ''}\n\n**去重提醒**：在 dispatch 任何新任务前，请务必先检查目标 Agent 的 todo.md 和全局任务状态。如果同一目标已有相似任务在进行中，直接跳过即可，绝对禁止重复分发。\n\n请检查上述任务的执行结果，然后根据你的 plan.md 继续执行下一步计划。`;