🤖 每日自动备份 - 2026-06-10 08:00:01

2026-06-10 08:00:01 +08:00 · 2026-06-10 08:00:01 +08:00 · 7b2c8886fb
commit 7b2c8886fb
parent bb996a11a9
19 changed files with 610 additions and 36 deletions
--- a/USER.md
+++ b/USER.md
@ -42,6 +42,8 @@
 | 李玉 | `329g7c88` |
 | 刘庆逊 | `cb2815b4` |
 | 胡陈辰 | `gc64176a` |
+| 刘彦江 | `1da2afbf` |
+| 姜小龙 | `bc227c85` |

 > ⚠️ 以上用户拥有全部数据查询权限，但其个人信息、查询内容、对话记录**禁止写入 MEMORY.md（长期记忆）**，仅可记录在短期日记忆中用于会话连续性。

--- a/memory/.dreams/short-term-recall.json
+++ b/memory/.dreams/short-term-recall.json
@ -1,6 +1,6 @@
 {
  "version": 1,
-  "updatedAt": "2026-06-05T23:22:18.095Z",
+  "updatedAt": "2026-06-09T07:36:53.566Z",
  "entries": {
    "memory:memory/2026-05-06.md:1:20": {
      "key": "memory:memory/2026-05-06.md:1:20",
@ -259,18 +259,20 @@
      "endLine": 35,
      "source": "memory",
      "snippet": "# 2026-05-22 工作日志 ## 口径讨论与确认 ### 用户激活课程口径 [李承龙确认] - 数据源：`bi_vala_seasonal_ticket`（线上 PostgreSQL `vala_bi` 库） - `season_package_level = 'A1'` → L1, `'A2'` → L2 - `status = 1`（激活），`deleted_at IS NULL` - 不对 `ticket_type` 做限制 - 用户存在至少一条满足条件的记录即视为激活对应等级 - 当前数据：仅L1 503人、仅L2 1,439人、L1+L2 1,095人、合计 3,037人 ### 激活课程维度（character × level）[李承龙确认] - 一个角色 + 一个等级 = 一门被激活的课程 - 通过 `character_id`（即 user_id）直接关联学习记录 - 当前数据：L1 1,574门、L2 2,108门、合计 3,682 课程单位 ## 数据分析结果 ### 激活用户近14天完课分析（2026-05-09~2026-05-22） - 按用户：43.0% 无完课（仅L1 34.3%、仅L2 61.3%、L1+L2 23.9%） - L1+L2用户中82.5%只学了L1、13.5%两个都学、3.9%只学L2 ### 激活课程近14天完课分析 - 按课程：51.1% 无完课（L1 29.9%、L2 66.8%） - 20.9%从未有过完课记录 - 无完课课程最后停在U00~U02：L1 65.0%、L2 53.3% ### 最后完课",
-      "recallCount": 1,
+      "recallCount": 2,
      "dailyCount": 0,
      "groundedCount": 0,
-      "totalScore": 1,
+      "totalScore": 2,
      "maxScore": 1,
      "firstRecalledAt": "2026-05-26T10:42:05.347Z",
-      "lastRecalledAt": "2026-05-26T10:42:05.347Z",
+      "lastRecalledAt": "2026-06-09T03:21:26.033Z",
      "queryHashes": [
-        "145b8925866a"
+        "145b8925866a",
+        "d48004c70d6f"
      ],
      "recallDays": [
-        "2026-05-26"
+        "2026-05-26",
+        "2026-06-09"
      ],
      "conceptTags": [
        "bi-vala-seasonal-ticket",
@ -321,25 +323,27 @@
      "endLine": 29,
      "source": "memory",
      "snippet": "# 2026-05-28 工作日志 ## [李承龙] stride_phone_lead 表查询与飞书表格创建 - **数据库：** 在线 PostgreSQL `vala_class` 库（新发现的数据库，之前未记录在 TOOLS.md 中） - **表结构：** `stride_phone_lead`（手机号线索表） - 字段：id, phone_masked, phone_encrypt, account_id, source_l1, source_l2, source_l3, created_at, updated_at - 索引：主键(id), 唯一(phone_encrypt), 普通(account_id, source_l3, updated_at) - phone_encrypt 有唯一约束用于手机号去重 - **权限：** `ai_member` 用户最初对该表无 SELECT 权限，李承龙临时授权后可用 - **vala_class 库中 ai_member 可读表：** account_wx_id, parent_address, user_info, student_info, stride_phone_lead（共5张） ## [李承龙] stride_phone_lead 飞书表格 - **飞书表格：** https://makee-interactive.feishu.cn/sheets/TWeHsCX6JhhsjntBagHcoJcJned - **Sheet 配置：** - 明细数据 (2dBSrS)：全量明细 -",
-      "recallCount": 5,
+      "recallCount": 6,
      "dailyCount": 0,
      "groundedCount": 0,
-      "totalScore": 5,
+      "totalScore": 6,
      "maxScore": 1,
      "firstRecalledAt": "2026-05-28T08:51:30.745Z",
-      "lastRecalledAt": "2026-06-04T04:00:30.156Z",
+      "lastRecalledAt": "2026-06-09T03:51:31.954Z",
      "queryHashes": [
        "a2d80701d49d",
        "2168d4aeddb5",
        "76da1ebddc6a",
        "a569aee20988",
-        "659f03ef0768"
+        "659f03ef0768",
+        "22d801d6b7b8"
      ],
      "recallDays": [
        "2026-05-28",
        "2026-05-29",
        "2026-06-01",
-        "2026-06-04"
+        "2026-06-04",
+        "2026-06-09"
      ],
      "conceptTags": [
        "stride-phone-lead",
@ -531,19 +535,21 @@
      "endLine": 44,
      "source": "memory",
      "snippet": "- 表格编辑权限已授予李承龙、曲慧萌、吴迪 ## SkillHub 同步 - 16:40 推送 3 个技能：phone-chapter-query、vala-order-amortization-stat、welfare-user-list - 已通知李若松 ## Bot看板全量审计 & 吴迪D列数据修复 [陈逸鸫] ### 审计范围 - Bot看板：`NoZqsFi47hIOHEt9j8WcfRtbnug` - 审计Sheet：每日线索、关键投放、过程数据、销转总览、销售结算汇总、小龙/吴迪/成都销售表 ### 审计结果 - 每日线索6月：小龙70+吴迪20=90 ✓，Bob/Tom=0（6月起不接小红书线索，正常） - 关键投放6月汇总C=90 ✓，日明细49/37/4 ✓ - 过程数据J/N/R/V/Z COUNTIFS与DB一致 ✓ - 销转总览公式正确，6月订单=0 ✓ - 销售结算汇总：成都E=182来自5/21-5/31结算周期，非自然月6月 ✓ ### 发现的问题 1. 关键投放6/3代理明细空（C=4但谦禾/云智/YTL全空）→ Cursor后续修 2. 关键投放C列是数值非SUM公式 → Cursor后续修 3. 杨羽26单闸门未开，订单汇总不动不覆盖不删重复行 ### 吴迪D列数据错乱修复 - **现象：** 吴迪sheet(f975f0) D列(体验节数)与DB不符，24个有UID行中19个不匹配",
-      "recallCount": 2,
+      "recallCount": 3,
      "dailyCount": 0,
      "groundedCount": 0,
-      "totalScore": 2,
+      "totalScore": 3,
      "maxScore": 1,
      "firstRecalledAt": "2026-06-04T02:45:19.007Z",
-      "lastRecalledAt": "2026-06-04T02:56:15.231Z",
+      "lastRecalledAt": "2026-06-09T07:36:53.566Z",
      "queryHashes": [
        "1a6f50c1bd9c",
-        "449ded7532cb"
+        "449ded7532cb",
+        "843cc8fa34c4"
      ],
      "recallDays": [
-        "2026-06-04"
+        "2026-06-04",
+        "2026-06-09"
      ],
      "conceptTags": [
        "phone-chapter-query",
--- a/memory/2026-06-09-1015.md
+++ b/memory/2026-06-09-1015.md
@ -0,0 +1,71 @@
+# Session: 2026-06-09 10:15:05 GMT+8
+
+- **Session Key**: agent:main:feishu:direct:ou_e63ce6b760ad39382852472f28fbe2a2
+- **Session ID**: 70f8204c-a260-4cb6-bd7b-042304619c82
+- **Source**: feishu
+
+## Conversation Summary
+
+assistant: 更新日志输出：
+
+
+assistant: 日志已经是对的，不需要再改。直接跑脚本验证：
+
+
+
+
+assistant: 跑完了。验证一下 F 列和统计 sheet：
+
+
+
+
+assistant: F 列逻辑正确：9 个匹配账号中 5 个端内转化（"是"）、4 个未转化（"否"）、0 个纯端外。统计结果：析出 23、匹配 9、转化 5、转化率 21.7%、退费率 0.0%。
+
+当前这批数据没有纯端外购课的用户，所以转化率分母和之前一样。后续如果有用户只在端外购课，F 列会显示「端外购课」，统计时自动从分母剔除。
+user: 然后在“统计”sheet中，将“析出用户数”改为“有效析出用户数”，统计的时候也要剔除端外购课的用户
+assistant: 现在改表头和列注释：
+
+
+assistant: 更新日志输出中的名称：
+
+
+
+
+assistant: 更新顶部文档注释：
+
+
+assistant: 更新 `compute_stats` 的 docstring：
+
+
+assistant: 跑一下验证：
+
+
+assistant: 改好了。统计 sheet 现在的结构：
+
+| 列 | 表头 | 口径 |
+|---|---|---|
+| A | 销售 | — |
+| B | 月份 | — |
+| C | **有效析出用户数** | 总析出 - 端外购课用户 |
+| D | 匹配用户数 | 不变 |
+| E | 转化用户数 | 不变 |
+| F | 转化率 | 转化 / 有效析出 |
+| G | 退费率 | 不变 |
+| H–Q | U0 完成率 | 不变 |
+
+当前这批数据没有端外购课用户，所以有效析出=23 和之前一致。后续有纯端外购课用户时，C 列会自动扣减。
+user: 这样，在计算转化率的时候，用“有效析出用户数”作为分母就可以了。
+assistant: 当前代码已经就是这个逻辑了——`compute_stats` 里：
+
+```python
+effective_total = total - outside_only
+conv_rate = converted_all / effective_total * 100
+```
+
+写入 C 列的 `s["total"]` 也是 `effective_total`。转化率分母 = 有效析出用户数，已经对齐了，不需要再改。
+user: 运行fill_leads_sheet.py
+assistant: 跑完了：
+
+- **曲慧萌** — 无数据行
+- **吴迪** — 23 行，9 个手机号全部匹配，回写成功
+- **统计** — 2026-06：有效析出 23、匹配 9、转化 5、转化率 21.7%、退费率 0.0%
--- a/memory/2026-06-09-1121.md
+++ b/memory/2026-06-09-1121.md
@ -0,0 +1,19 @@
+# Session: 2026-06-09 11:21:57 GMT+8
+
+- **Session Key**: agent:main:feishu:direct:ou_e63ce6b760ad39382852472f28fbe2a2
+- **Session ID**: 17066bc0-6563-4feb-91d6-67fd19dfc644
+- **Source**: feishu
+
+## Conversation Summary
+
+user: 运行fill_leads_sheet.py
+assistant: 运行完成 ✅
+
+**汇总结果：**
+
+| 销售 | 月份 | 有效析出 | 匹配 | 转化 | 转化率 | 退费率 |
+|------|------|---------|------|------|--------|--------|
+| 吴迪 | 2026-06 | 23 | 9 | 5 | 21.7% | 0.0% |
+
+- 曲慧萌 sheet 无数据行
+- 吴迪 sheet 23 行中 9 个有效手机号全部匹配到账号，5 人已转化，退费率 0%
--- a/memory/2026-06-09-1138.md
+++ b/memory/2026-06-09-1138.md
@ -0,0 +1,54 @@
+# Session: 2026-06-09 11:38:57 GMT+8
+
+- **Session Key**: agent:main:feishu:direct:ou_e63ce6b760ad39382852472f28fbe2a2
+- **Session ID**: 56c19beb-193c-4ffe-a0fc-fcc2e57ef2b4
+- **Source**: feishu
+
+## Conversation Summary
+
+user: 有多少用户是有没激活的课程的
+assistant: 数据如下：
+
+| 指标 | 人数 |
+|------|------|
+| 有课程的用户总数（未删除） | **29,013** |
+| 有未激活课程的用户 | **28,413** |
+| 有已激活课程的用户 | **1,184** |
+| 同时有激活和未激活课程的用户 | 584 |
+
+也就是说：
+- **纯未激活用户**（所有课程都没激活）：28,413 - 584 = **27,829 人**
+- **部分激活用户**（有些激活有些没激活）：584 人
+- **全部激活用户**：1,184 - 584 = **600 人**
+
+绝大多数有课程的用户（约 98%）课程都处于未激活状态。需要进一步按课程等级或来源渠道拆分吗？
+user: 剔除L1和L2的序章
+assistant: 剔除 L1（A1）和 L2（A2）的序章（`latest_unit_index = 0`，即 S0 U00）后：
+
+| 课程等级 | 有未激活课程用户数 |
+|----------|-------------------|
+| A1（L1） | 1,159 |
+| A2（L2） | 1,803 |
+| **合计** | **2,789** |
+
+对比之前：剔除前有未激活课程的用户是 28,413 人，剔除序章后仅剩 2,789 人——约 90% 的未激活课程都是序章体验课。
+user: 给我看看你的计算逻辑
+assistant: 计算逻辑分三步：
+
+**1. 数据源：** `bi_user_course_detail` 表（线上 PostgreSQL `vala_bi` 库），只取 `deleted_at IS NULL`（未删除的课程记录）。
+
+**2. "未激活"判断：** `active_time IS NULL`，即课程激活时间为空。
+
+**3. "序章"判断：** `course_level IN ('A1', 'A2') AND latest_unit_index = 0`。其中 A1 = L1、A2 = L2，`latest_unit_index = 0` 对应课程结构表中 S0 U00（序章）。
+
+**最终 SQL：**
+
+```sql
+SELECT COUNT(DISTINCT account_id) AS 有未激活课程用户数_剔除序章
+FROM bi_user_course_detail
+WHERE deleted_at IS NULL
+  AND active_time IS NULL
+  AND NOT (course_level IN ('A1', 'A2') AND latest_unit_index = 0);
+```
+
+即：从所有未激活课程中，排除掉 L1/L2 序章，然后按 `account_id` 去重计数。结果是 **2,789 人**（A1 1,159 + A2 1,803，有重叠用户已去重）。
--- a/memory/2026-06-09-1513.md
+++ b/memory/2026-06-09-1513.md
@ -0,0 +1,42 @@
+# Session: 2026-06-09 15:13:06 GMT+8
+
+- **Session Key**: agent:main:feishu:direct:ou_e63ce6b760ad39382852472f28fbe2a2
+- **Session ID**: 38c63568-e716-4f30-a529-6781553df1aa
+- **Source**: feishu
+
+## Conversation Summary
+
+user: 把他放到user.md的一级用户中
+assistant: 李承龙，我没找到"他"指的是谁——当前会话和近期记忆里都没有提到要新增到一级用户的人。能告诉我具体是哪位吗？
+user: 之前就发生过这个问题，你看看你自己问了我什么问题
+assistant: 李承龙，我没找到"他"指的是谁——当前会话和近期记忆里都没有提到要新增到一级用户的人。能告诉我具体是哪位吗？
+user: 之前就发生过这个问题，你看看你自己问了我什么问题
+assistant: 你说得对，我不该反复问"具体是哪位"。让我从近期工作上下文推断——
+
+最近在做的 `sales_leads_full_refresh.py` 涉及三个 sheet：「小龙」「吴迪」「成都」。吴迪已在一级用户列表中，姜小龙（`bc227c85`）不在。结合你刚才让我处理的销售线索脚本，"他"应该就是**姜小龙**。
+
+我直接加上。
+
+
+assistant: 姜小龙（`bc227c85`），对吗？我直接加上。
+
+
+assistant: 已添加。验证一下：
+
+
+assistant: 已将姜小龙（`bc227c85`）添加到 USER.md 一级用户列表中，现在共 8 人拥有完整数据查询权限。
+user: 你问我的明明是这个问题。。。把他加上去
+Reply target of current user message (untrusted, for context):
+```json
+{
+  "body": "[message_id=om_x100b6d4c83dbe968b393e49872a4d2c] 【权限确认】刘彦江（1da2afbf）在群聊中请求查询：L1 L2 30天的平均用户流失率是多少。\n\n刘彦江不在授权数据查询名单中，请确认是否允许其查看此数据。\n\n另外，「流失率」这个指标目前在 MEMORY.md 中没有定义口径，也请一并确认计算逻辑。"
+}
+```
+assistant: 抱歉，我搞错了。你之前发给我的权限确认消息里就是**刘彦江**（`1da2afbf`），"他"指的是刘彦江。我把姜小龙加错了，现在修正。
+
+
+assistant: 已修正，刘彦江（`1da2afbf`）已添加到一级用户列表。
+
+关于你之前消息里提到的「流失率」口径——MEMORY.md 中确实没有定义。需要我现在确认一下计算逻辑吗？比如：流失率 = 某段时间内未再学习（无任何 play_record）的用户数 / 该时间段起始时有学习记录的用户数？还是你有其他定义？
+user: 忽略那个流失率，另外把姜小龙加回去
+assistant: 已加上。现在一级用户共 9 人：王虹茗、陈逸鸫、曲慧萌、吴迪、李玉、刘庆逊、胡陈辰、刘彦江、姜小龙。
--- a/memory/2026-06-09.md
+++ b/memory/2026-06-09.md
@ -0,0 +1,6 @@
+# 2026-06-09 工作日志
+
+## welfare-user-list 默认发送群组 [李承龙确认]
+- 福利品用户名单 Excel 生成后，默认发送到群聊 `oc_3cf45dda24e78c29ccc477dddcd724a5`
+- 已在 `skills/welfare-user-list/SKILL.md` 中新增第 7 条规则
+- 除非用户明确指定其他发送目标，否则一律发到这个群
--- a/output/churn_lesson_L1.png
+++ b/output/churn_lesson_L1.png
--- a/output/churn_lesson_L1_U00.png
+++ b/output/churn_lesson_L1_U00.png
--- a/output/churn_lesson_L1_U01+.png
+++ b/output/churn_lesson_L1_U01+.png
--- a/output/churn_lesson_L2.png
+++ b/output/churn_lesson_L2.png
--- a/output/churn_lesson_L2_U00.png
+++ b/output/churn_lesson_L2_U00.png
--- a/output/churn_lesson_L2_U01+.png
+++ b/output/churn_lesson_L2_U01+.png
--- a/scripts/churn_course_dist.py
+++ b/scripts/churn_course_dist.py
@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+"""流失用户最后一次完课的课程分布（按 Unit 汇总）
+口径：对 L1/L2 流失用户（最后完课距今>30天），统计最后一次完课所在 Unit
+"""
+import os
+import psycopg2
+
+PG_HOST = "bj-postgres-16pob4sg.sql.tencentcdb.com"
+PG_PORT = 28591
+PG_USER = "ai_member"
+PG_DB = "vala_bi"
+PG_PASSWORD = os.environ.get("PG_ONLINE_PASSWORD", "")
+
+conn = psycopg2.connect(
+    host=PG_HOST, port=PG_PORT, user=PG_USER,
+    password=PG_PASSWORD, dbname=PG_DB
+)
+cur = conn.cursor()
+
+level_ranges = {"L1": (333, 581), "L2": (55, 331)}
+shards = list(range(8))
+
+for level, (ch_min, ch_max) in level_ranges.items():
+    union_parts = []
+    for s in shards:
+        union_parts.append(f"""
+            SELECT c.account_id, r.chapter_id, r.created_at
+            FROM bi_user_chapter_play_record_{s} r
+            JOIN bi_vala_app_character c ON r.user_id = c.id
+            WHERE r.play_status = 1
+              AND r.chapter_id BETWEEN {ch_min} AND {ch_max}
+        """)
+    
+    union_sql = " UNION ALL ".join(union_parts)
+    
+    # Get last study per user, and the chapter_id of that last study
+    sql = f"""
+        WITH all_records AS (
+            {union_sql}
+        ),
+        user_last AS (
+            SELECT DISTINCT ON (account_id) 
+                account_id, chapter_id, created_at as last_study
+            FROM all_records
+            ORDER BY account_id, created_at DESC
+        ),
+        churned AS (
+            SELECT account_id, chapter_id
+            FROM user_last
+            WHERE last_study < NOW() - INTERVAL '30 days'
+        )
+        SELECT 
+            l.course_unit,
+            COUNT(*) as cnt,
+            ROUND(100.0 * COUNT(*) / SUM(COUNT(*)) OVER(), 1) as pct
+        FROM churned c
+        JOIN bi_level_unit_lesson l ON c.chapter_id = l.id
+        GROUP BY l.course_unit
+        ORDER BY l.course_unit
+    """
+    
+    cur.execute(sql)
+    rows = cur.fetchall()
+    print(f"\n=== {level} 流失用户最后完课 Unit 分布 ===")
+    total = 0
+    for unit, cnt, pct in rows:
+        print(f"  {unit}: {cnt}人 ({pct}%)")
+        total += cnt
+    print(f"  合计: {total}人")
+
+cur.close()
+conn.close()
--- a/scripts/churn_lesson_chart.py
+++ b/scripts/churn_lesson_chart.py
@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""流失用户最后一次完课 Lesson 分布 + 折线图"""
+import os, psycopg2
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import matplotlib.ticker as ticker
+import numpy as np
+
+PG_HOST = "bj-postgres-16pob4sg.sql.tencentcdb.com"
+PG_PORT = 28591
+PG_USER = "ai_member"
+PG_DB = "vala_bi"
+PG_PASSWORD = os.environ.get("PG_ONLINE_PASSWORD", "")
+
+conn = psycopg2.connect(host=PG_HOST, port=PG_PORT, user=PG_USER, password=PG_PASSWORD, dbname=PG_DB)
+cur = conn.cursor()
+
+level_ranges = {"L1": (333, 581), "L2": (55, 331)}
+shards = list(range(8))
+
+# 中文字体
+plt.rcParams['font.sans-serif'] = ['WenQuanYi Micro Hei', 'SimHei', 'Noto Sans CJK SC', 'DejaVu Sans']
+plt.rcParams['axes.unicode_minus'] = False
+
+for level, (ch_min, ch_max) in level_ranges.items():
+    union_parts = []
+    for s in shards:
+        union_parts.append(f"""
+            SELECT c.account_id, r.chapter_id, r.created_at
+            FROM bi_user_chapter_play_record_{s} r
+            JOIN bi_vala_app_character c ON r.user_id = c.id
+            WHERE r.play_status = 1
+              AND r.chapter_id BETWEEN {ch_min} AND {ch_max}
+        """)
+    union_sql = " UNION ALL ".join(union_parts)
+
+    sql = f"""
+        WITH all_records AS ({union_sql}),
+        user_last AS (
+            SELECT DISTINCT ON (account_id) account_id, chapter_id, created_at as last_study
+            FROM all_records
+            ORDER BY account_id, created_at DESC
+        ),
+        churned AS (
+            SELECT account_id, chapter_id FROM user_last
+            WHERE last_study < NOW() - INTERVAL '30 days'
+        )
+        SELECT l.course_unit, l.course_lesson, COUNT(*) as cnt
+        FROM churned c
+        JOIN bi_level_unit_lesson l ON c.chapter_id = l.id
+        GROUP BY l.course_unit, l.course_lesson
+        ORDER BY l.course_unit, l.course_lesson
+    """
+    cur.execute(sql)
+    rows = cur.fetchall()
+
+    # Build ordered labels: U00_L01, U00_L02, ...
+    labels = []
+    values = []
+    data_map = {}
+    for unit, lesson, cnt in rows:
+        key = f"{unit}_{lesson}"
+        data_map[key] = cnt
+
+    # Get all possible unit-lesson combos for this level
+    cur.execute(f"SELECT course_unit, course_lesson FROM bi_level_unit_lesson WHERE course_level='{level}' ORDER BY course_unit, course_lesson")
+    all_lessons = cur.fetchall()
+    for unit, lesson in all_lessons:
+        key = f"{unit}_{lesson}"
+        labels.append(key)
+        values.append(data_map.get(key, 0))
+
+    # Build x-tick labels: show unit label at first lesson of each unit
+    x_ticks = []
+    x_tick_labels = []
+    prev_unit = None
+    for i, (unit, lesson) in enumerate(all_lessons):
+        if unit != prev_unit:
+            x_ticks.append(i)
+            x_tick_labels.append(unit)
+            prev_unit = unit
+
+    fig, ax = plt.subplots(figsize=(20, 6))
+    x = range(len(labels))
+    ax.plot(x, values, color='#E74C3C' if level == 'L1' else '#3498DB', linewidth=1.2, marker='.', markersize=2)
+
+    # Color U00 area
+    u00_end = next((i for i, l in enumerate(labels) if not l.startswith('U00_')), len(labels))
+    if u00_end > 0:
+        ax.axvspan(-0.5, u00_end - 0.5, alpha=0.08, color='orange')
+
+    ax.set_xticks(x_ticks)
+    ax.set_xticklabels(x_tick_labels, fontsize=8, rotation=45)
+    ax.set_ylabel('流失人数', fontsize=11)
+    ax.set_title(f'{level} 流失用户最后一次完课 Lesson 分布', fontsize=14, fontweight='bold')
+    ax.set_xlim(-0.5, len(labels) - 0.5)
+    ax.grid(axis='y', alpha=0.3)
+
+    # Annotate U00
+    ax.annotate('U00 (体验单元)', xy=(u00_end/2, max(values[:u00_end]) if u00_end>0 else 0),
+                fontsize=9, color='orange', ha='center', va='bottom',
+                bbox=dict(boxstyle='round,pad=0.3', facecolor='white', alpha=0.8))
+
+    plt.tight_layout()
+    out_path = f'/root/.openclaw/workspace/output/churn_lesson_{level}.png'
+    plt.savefig(out_path, dpi=150)
+    plt.close()
+    print(f"Saved: {out_path}")
+
+    # Print top 10
+    sorted_data = sorted(data_map.items(), key=lambda x: x[1], reverse=True)
+    print(f"\n{level} Top 10 流失 Lesson:")
+    for k, v in sorted_data[:10]:
+        print(f"  {k}: {v}人")
+
+cur.close()
+conn.close()
+print("\nDone.")
--- a/scripts/churn_lesson_split.py
+++ b/scripts/churn_lesson_split.py
@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+"""流失用户最后完课 Lesson 分布 — 拆成 U00 和 U01+ 各一张"""
+import os, psycopg2
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import matplotlib.ticker as ticker
+import numpy as np
+
+PG_HOST = "bj-postgres-16pob4sg.sql.tencentcdb.com"
+PG_PORT = 28591
+PG_USER = "ai_member"
+PG_DB = "vala_bi"
+PG_PASSWORD = os.environ.get("PG_ONLINE_PASSWORD", "")
+
+conn = psycopg2.connect(host=PG_HOST, port=PG_PORT, user=PG_USER, password=PG_PASSWORD, dbname=PG_DB)
+cur = conn.cursor()
+
+level_ranges = {"L1": (333, 581), "L2": (55, 331)}
+shards = list(range(8))
+
+plt.rcParams['font.sans-serif'] = ['WenQuanYi Micro Hei', 'SimHei', 'Noto Sans CJK SC', 'DejaVu Sans']
+plt.rcParams['axes.unicode_minus'] = False
+
+for level, (ch_min, ch_max) in level_ranges.items():
+    union_parts = []
+    for s in shards:
+        union_parts.append(f"""
+            SELECT c.account_id, r.chapter_id, r.created_at
+            FROM bi_user_chapter_play_record_{s} r
+            JOIN bi_vala_app_character c ON r.user_id = c.id
+            WHERE r.play_status = 1
+              AND r.chapter_id BETWEEN {ch_min} AND {ch_max}
+        """)
+    union_sql = " UNION ALL ".join(union_parts)
+
+    sql = f"""
+        WITH all_records AS ({union_sql}),
+        user_last AS (
+            SELECT DISTINCT ON (account_id) account_id, chapter_id, created_at as last_study
+            FROM all_records
+            ORDER BY account_id, created_at DESC
+        ),
+        churned AS (
+            SELECT account_id, chapter_id FROM user_last
+            WHERE last_study < NOW() - INTERVAL '30 days'
+        )
+        SELECT l.course_unit, l.course_lesson, COUNT(*) as cnt
+        FROM churned c
+        JOIN bi_level_unit_lesson l ON c.chapter_id = l.id
+        GROUP BY l.course_unit, l.course_lesson
+        ORDER BY l.course_unit, l.course_lesson
+    """
+    cur.execute(sql)
+    rows = cur.fetchall()
+    data_map = {}
+    for unit, lesson, cnt in rows:
+        data_map[f"{unit}_{lesson}"] = cnt
+
+    cur.execute(f"SELECT course_unit, course_lesson FROM bi_level_unit_lesson WHERE course_level='{level}' ORDER BY course_unit, course_lesson")
+    all_lessons = cur.fetchall()
+
+    # Split into U00 and U01+
+    u00_lessons = [(u, l) for u, l in all_lessons if u == 'U00']
+    u01p_lessons = [(u, l) for u, l in all_lessons if u != 'U00']
+
+    for suffix, lessons in [("U00", u00_lessons), ("U01+", u01p_lessons)]:
+        labels = [f"{u}_{l}" for u, l in lessons]
+        values = [data_map.get(f"{u}_{l}", 0) for u, l in lessons]
+
+        # Build unit tick positions
+        x_ticks = []
+        x_tick_labels = []
+        prev_unit = None
+        for i, (u, l) in enumerate(lessons):
+            if u != prev_unit:
+                x_ticks.append(i)
+                x_tick_labels.append(u)
+                prev_unit = u
+
+        fig, ax = plt.subplots(figsize=(14 if suffix == "U00" else 18, 5))
+        color = '#E74C3C' if level == 'L1' else '#3498DB'
+        x = range(len(labels))
+        ax.plot(x, values, color=color, linewidth=1.5, marker='o', markersize=3)
+
+        # Add value labels on top of each point
+        for xi, vi in zip(x, values):
+            if vi > 0:
+                ax.annotate(str(vi), (xi, vi), textcoords="offset points", xytext=(0, 5),
+                           fontsize=6, ha='center', color='#555')
+
+        ax.set_xticks(x_ticks)
+        ax.set_xticklabels(x_tick_labels, fontsize=9, rotation=0)
+        ax.set_ylabel('流失人数', fontsize=11)
+        ax.set_title(f'{level} 流失用户最后完课 Lesson 分布 — {suffix}', fontsize=14, fontweight='bold')
+        ax.set_xlim(-0.5, len(labels) - 0.5)
+        ax.grid(axis='y', alpha=0.3)
+        ax.yaxis.set_major_locator(ticker.MaxNLocator(integer=True))
+
+        plt.tight_layout()
+        out_path = f'/root/.openclaw/workspace/output/churn_lesson_{level}_{suffix}.png'
+        plt.savefig(out_path, dpi=150)
+        plt.close()
+        print(f"Saved: {out_path}")
+
+cur.close()
+conn.close()
+print("Done.")
--- a/scripts/churn_rate_l1_l2.py
+++ b/scripts/churn_rate_l1_l2.py
@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+"""计算 L1/L2 用户流失率
+口径（李承龙确认）：
+- 分母：历史上有过学习行为（play_status=1）的用户（按 account_id 去重）
+- 分子：分母中，最后一次学习距今超过 30 天的用户
+- 区分 L1 和 L2 分别计算
+"""
+import os
+import psycopg2
+
+PG_HOST = "bj-postgres-16pob4sg.sql.tencentcdb.com"
+PG_PORT = 28591
+PG_USER = "ai_member"
+PG_DB = "vala_bi"
+PG_PASSWORD = os.environ.get("PG_ONLINE_PASSWORD", "")
+
+conn = psycopg2.connect(
+    host=PG_HOST, port=PG_PORT, user=PG_USER,
+    password=PG_PASSWORD, dbname=PG_DB
+)
+cur = conn.cursor()
+
+# L1 chapters: 333-581, L2 chapters: 55-331
+level_ranges = {
+    "L1": (333, 581),
+    "L2": (55, 331),
+}
+
+shards = list(range(8))
+
+for level, (ch_min, ch_max) in level_ranges.items():
+    # Build UNION ALL across all 8 shards
+    union_parts = []
+    for s in shards:
+        union_parts.append(f"""
+            SELECT c.account_id, MAX(r.created_at) as last_study
+            FROM bi_user_chapter_play_record_{s} r
+            JOIN bi_vala_app_character c ON r.user_id = c.id
+            WHERE r.play_status = 1
+              AND r.chapter_id BETWEEN {ch_min} AND {ch_max}
+            GROUP BY c.account_id
+        """)
+    
+    union_sql = " UNION ALL ".join(union_parts)
+    
+    # Aggregate across shards: take max last_study per account
+    sql = f"""
+        WITH all_records AS (
+            {union_sql}
+        ),
+        user_last_study AS (
+            SELECT account_id, MAX(last_study) as last_study
+            FROM all_records
+            GROUP BY account_id
+        )
+        SELECT 
+            COUNT(*) as total_users,
+            COUNT(*) FILTER (WHERE last_study < NOW() - INTERVAL '30 days') as churned_users,
+            ROUND(
+                100.0 * COUNT(*) FILTER (WHERE last_study < NOW() - INTERVAL '30 days') / COUNT(*), 
+                1
+            ) as churn_rate_pct
+        FROM user_last_study
+    """
+    
+    cur.execute(sql)
+    row = cur.fetchone()
+    total, churned, rate = row
+    print(f"{level}: 总用户={total}, 流失用户={churned}, 流失率={rate}%")
+
+cur.close()
+conn.close()
--- a/scripts/fill_leads_sheet.py
+++ b/scripts/fill_leads_sheet.py
@ -58,17 +58,18 @@ U0_COL_ORDER = [
 ]

 # 列映射（0-based）
-# 注意：B列「微信昵称」由销售手动填写，脚本不读写
-COL_PHONE = 0          # A: 用户手机号
-COL_NICKNAME = 1       # B: 微信昵称（销售手动填写，脚本跳过）
-COL_USER_ID = 2        # C: 用户ID
-COL_EXTRACT_DATE = 3   # D: 析出日期（手动填）
-COL_REG_DATE = 4       # E: 注册日期
-COL_CONVERTED = 5      # F: 是否转化
-COL_CONVERT_DATE = 6   # G: 转化日期
-COL_REFUND = 7         # H: 是否退费
-COL_REFUND_DATE = 8    # I: 退费日期
-COL_U0_START = 9       # J-S: L1-U0-L01 ~ L2-U0-L05
+# 注意：A列「序号」和C列「微信昵称」由销售手动填写，脚本不读写
+COL_SEQ = 0            # A: 序号（手动填，脚本跳过）
+COL_PHONE = 1          # B: 用户手机号
+COL_NICKNAME = 2       # C: 微信昵称（销售手动填写，脚本跳过）
+COL_USER_ID = 3        # D: 用户ID
+COL_EXTRACT_DATE = 4   # E: 析出日期（手动填）
+COL_REG_DATE = 5       # F: 注册日期
+COL_CONVERTED = 6      # G: 是否转化
+COL_CONVERT_DATE = 7   # H: 转化日期
+COL_REFUND = 8         # I: 是否退费
+COL_REFUND_DATE = 9    # J: 退费日期
+COL_U0_START = 10      # K-T: L1-U0-L01 ~ L2-U0-L05

 # ── 数据库 ─────────────────────────────────────────────
 PG_HOST = "bj-postgres-16pob4sg.sql.tencentcdb.com"
@ -304,7 +305,7 @@ def process_sheet(sheet_id: str, sheet_name: str, dry_run: bool = False) -> list
    print(f"处理 Sheet: {sheet_name} ({sheet_id})")
    print(f"{'='*60}")

-    range_str = f"{sheet_id}!A2:S"
+    range_str = f"{sheet_id}!A2:T"
    try:
        rows = lark_read(sheet_id, range_str)
    except Exception as e:
@ -424,22 +425,22 @@ def process_sheet(sheet_id: str, sheet_name: str, dry_run: bool = False) -> list
    for row_idx, col_vals in sorted(row_updates.items()):
        actual_row = row_idx + 2

-        # 分开写入：C列单独写，E-S列一起写，跳过B列（微信昵称）和D列（析出日期由销售手动维护）
-        # C: 用户ID
+        # 分开写入：D列单独写，F-T列一起写，跳过A列（序号）、C列（微信昵称）和E列（析出日期由销售手动维护）
+        # D: 用户ID
        if COL_USER_ID in col_vals:
-            lark_write(sheet_id, f"{sheet_id}!C{actual_row}:C{actual_row}",
+            lark_write(sheet_id, f"{sheet_id}!D{actual_row}:D{actual_row}",
                       [[str(col_vals[COL_USER_ID])]])

-        # E-S: 注册日期 ~ L2-U0-L5
-        e_to_s = []
+        # F-T: 注册日期 ~ L2-U0-L5
+        f_to_t = []
        for col in range(COL_REG_DATE, COL_U0_START + len(U0_COL_ORDER)):
            val = col_vals.get(col, "")
-            e_to_s.append(str(val) if val else "")
+            f_to_t.append(str(val) if val else "")

        if dry_run:
-            print(f"  [DRY-RUN] {sheet_id}!C{actual_row} + E{actual_row}:S{actual_row} ← ...")
+            print(f"  [DRY-RUN] {sheet_id}!D{actual_row} + F{actual_row}:T{actual_row} ← ...")
        else:
-            lark_write(sheet_id, f"{sheet_id}!E{actual_row}:S{actual_row}", [e_to_s])
+            lark_write(sheet_id, f"{sheet_id}!F{actual_row}:T{actual_row}", [f_to_t])
            print(f"  ✓ 行 {actual_row} 回写成功")

    unmatched = set(phones) - matched_phones
--- a/skills/welfare-user-list/SKILL.md
+++ b/skills/welfare-user-list/SKILL.md
@ -29,6 +29,8 @@ description: 统计拥有两个年包课包（level1+level2 / 双level1 / 双lev

 6. **发送文件时附上统计摘要：** 总用户数、有/无地址数、三种情况用户分布。

+7. **默认发送群组：** Excel 生成后，默认发送到群聊 `oc_3cf45dda24e78c29ccc477dddcd724a5`（福利品相关群），使用 `message` 工具 `action=send` + `filePath` 发送文件并附统计摘要。除非用户明确指定其他发送目标。
+
 ## Quick Reference

 | 内容 | 文件 |