From f66cc93781b139f76126ade4c901c108a8cf3da5 Mon Sep 17 00:00:00 2001
From: NathanAlex7 <18898599977@163.com>
Date: Tue, 1 Sep 2026 17:13:31 +0800
Subject: [PATCH] =?UTF-8?q?fix(xingbang):=20reward=20=E8=AE=A1=E7=AE=97?=
 =?UTF-8?q?=E5=90=8E=E5=86=99=E5=9B=9E=E8=BD=A8=E8=BF=B9=E6=96=87=E4=BB=B6?=
 =?UTF-8?q?,=20=E8=BF=9B=E5=8C=96=E9=97=AD=E7=8E=AF=E6=95=B0=E6=8D=AE?=
 =?UTF-8?q?=E5=AE=8C=E6=95=B4?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

此前 cumulative_reward 只存在于返回对象, 未写回磁盘轨迹文件
(finalize 先落盘, 之后改的 reward 只在内存)。新增 rewrite() 在
reward 计算后重新覆盖轨迹, 确保自进化数据(含奖励)完整可追溯。
---
 backend/xingbang/evolution/trajectory.py | 13 +++++++++++++
 backend/xingbang/workflow.py             |  2 ++
 2 files changed, 15 insertions(+)

diff --git a/backend/xingbang/evolution/trajectory.py b/backend/xingbang/evolution/trajectory.py
index 4d9d4d0..6cd05d4 100644
--- a/backend/xingbang/evolution/trajectory.py
+++ b/backend/xingbang/evolution/trajectory.py
@@ -90,6 +90,19 @@ class TrajectoryRecorder:
         )
         return path
 
+    def set_reward(self, reward: float) -> None:
+        """计算奖励后写回当前轨迹并重新落盘(进化闭环数据完整)。"""
+        if self._current is not None:
+            self._current["cumulative_reward"] = reward
+            # 若已有 trajectory_id, 覆盖落盘(full path 已知)
+            tid = self._current.get("trajectory_id")
+            if tid:
+                self._save(self._current)
+
+    def rewrite(self, trajectory: dict[str, Any]) -> Path:
+        """写回一条已存在的轨迹(非当前进行中的), 返回路径。"""
+        return self._save(trajectory)
+
     # ── 从调用侧采集 ────────────────────────────────────
     def on_evidence(self, kind: str, data: dict) -> None:
         """订阅 agents 的 evidence 回调,累积成本步记录。"""
diff --git a/backend/xingbang/workflow.py b/backend/xingbang/workflow.py
index 4ca78e2..85df112 100644
--- a/backend/xingbang/workflow.py
+++ b/backend/xingbang/workflow.py
@@ -122,6 +122,7 @@ class XingbangWorkflow:
         )
         reward = compute_reward_v1(trajectory)
         trajectory["cumulative_reward"] = reward
+        self._recorder.rewrite(trajectory)  # 写回含 reward 轨迹, 进化数据完整
         insights_path = self._evolution.save_insights()
 
         return {
@@ -180,6 +181,7 @@ class XingbangWorkflow:
         trajectory = self._recorder.finalize(success=True)
         reward = compute_reward_v1(trajectory)
         trajectory["cumulative_reward"] = reward
+        self._recorder.rewrite(trajectory)  # 写回含 reward 轨迹, 进化数据完整
         insights_path = self._evolution.save_insights()
         yield {"event": "done", "workflow_id": workflow_id,
                "trajectory_id": traj_id, "reward": reward,