From ef20ceb6efa1a520056f7b07082a912557719ea6 Mon Sep 17 00:00:00 2001
From: yibo <bboonnee@gmail.com>
Date: Wed, 2 Sep 2026 10:54:06 +0800
Subject: [PATCH] fix(graph-engine): reopenRunForRetry/孤儿清扫重挂 auto 驱动循环 (T-mtjh8sdt)

---
 server/services/graph-engine.js | 30 ++++++++++++++++++++++++++++++
 1 file changed, 30 insertions(+), 0 deletions(-)

diff --git a/server/services/graph-engine.js b/server/services/graph-engine.js
index f7cdfaf..69a06a0 100644
--- a/server/services/graph-engine.js
+++ b/server/services/graph-engine.js
@@ -3204,6 +3204,7 @@ function stopWallclockWatchdog() {
  */
 function _reapStaleRunningNodes(rows, summary) {
   const swept = [];
+  const runIdsToRedrive = new Set();
   for (const row of rows) {
     try {
       const run = getRun(row.run_id);
@@ -3220,10 +3221,26 @@ function _reapStaleRunningNodes(rows, summary) {
       const out = { activated: [], backwardEdgeId: null, loop: null, contractDiff: null, crossLayer: null, reset: [], infraRerun: false };
       routeFailure(row.run_id, def, nodeRun, out, null);
       swept.push({ runId: row.run_id, nodeRunId: row.id, infraRerun: out.infraRerun });
+      runIdsToRedrive.add(row.run_id);
     } catch (err) {
       console.error(`[graph-engine] _reapStaleRunningNodes node=${row.id} 异常: ${err.message}`);
     }
   }
+  // T-mtjh8sdt 同一个洞的第二个入口：routeFailure 把节点重置回 ready/running 只改了状态，
+  // 不会重新挂上 runAuto 的自驱动循环（reopenRunForRetry 那次已经补过人工重试路径，这里
+  // 是开机孤儿清扫 sweepOrphanedRunningNodes 与 45 号文档持续性存活扫描
+  // sweepStaleRunningAgentCalls 共用的结算内层循环，同一缺陷的另一条触发路径）。
+  // 异步补、不 await：sweepOrphanedRunningNodes 在服务器启动路径上同步调用，扫描可能
+  // 命中好几个 run，不能让重新驱动的 agent 调用顶着启动耗时走；按 run 去重，避免同一个
+  // run 里多个孤儿节点各自触发一次 runAuto。
+  for (const runId of runIdsToRedrive) {
+    const after = getRun(runId);
+    if (after && after.run_status === 'running' && after.execution_mode === 'auto') {
+      runAuto(runId).catch((err) => {
+        console.error(`[graph-engine] 孤儿清扫重挂 auto 驱动异常 run=${runId}: ${err.message}`);
+      });
+    }
+  }
   return { scanned: rows.length, swept };
 }
 
@@ -3802,6 +3819,13 @@ function descendantRuns(run) {
  *
  * @returns {void} 没有「失败」这个出口：要么推进了，要么被别人抢先推走了（都可以继续动节点），
  *   要么抛 RunSlotBusyError（整条重试作废，节点一个字都没动）。
+ *
+ * **auto 模式重挂驱动循环**（T-mtjh8sdt）：这里把 run 从 blocked/ready 拉回 running 走的是
+ * 裸 `admitOrRequeue`，不是 `startRun`——所以以前只改了状态，从没重新挂上 `runAuto` 的自驱动
+ * 循环。手动 `nextStep` 推一步后没人接着推，agentRunId 仍是 null，backward 重试每次都会在
+ * 这里再卡一次。`startRun` 的收尾那段（判 execution_mode==='auto' 才挂 driver）搬过来一份：
+ * 只在真的把 run 推成了 running 且 execution_mode 是 auto 时才挂，manual 模式或没推动（requeued/
+ * lost）都不挂，语义与 startRun 对齐。
  */
 function reopenRunForRetry(runId, reopenRun) {
   if (!reopenRun) return;
@@ -3813,6 +3837,12 @@ function reopenRunForRetry(runId, reopenRun) {
   if (cur === 'blocked' || cur === 'ready') {
     admitOrRequeue(runId, cur, 'running', { summary: `人工重试重开链路：${cur} → running` });
   }
+  const after = getRun(runId);
+  if (after && after.run_status === 'running' && after.execution_mode === 'auto') {
+    runAuto(runId).catch((err) => {
+      console.error(`[graph-engine] 人工重试重挂 auto 驱动异常 run=${runId}: ${err.message}`);
+    });
+  }
 }
 
 /**
--
libgit2 0.21.4