baoyu-wechat-summary v1.118.0: 归因校验、群主查证、memory 强制留痕、机械计数

2026-07-12 05:51:44 +08:00 · 2026-07-01 21:34:41 -05:00
parent 348dc82612
commit 1378544ea2
2 changed files with 54 additions and 8 deletions
@@ -1,7 +1,7 @@
 ---
 name: baoyu-wechat-summary
 description: Summarizes WeChat group chat highlights into a structured digest using the local wx-cli binary (https://github.com/jackwener/wx-cli). Generates a normal digest by default; a roast (毒舌) version is opt-in. Maintains per-group history (history.json + history-digests.jsonl), per-user profiles, and per-group fact memory (memory.md) across runs, with privacy guardrails baked in. Use when the user asks to "总结群聊", "群聊精华", "群聊摘要", "summarize group chat", "group chat digest", mentions a WeChat group name with a time range, says "帮我看看 XX 群最近聊了什么", "XX 群有什么值得看的", or asks to "回溯画像" / "初始化画像" / "backfill profiles". Adds the roast version when the user says "毒舌版", "roast 版", "再来个毒舌的", or similar.
-version: 1.117.4
+version: 1.118.0
 metadata:
  openclaw:
    homepage: https://github.com/JimLiu/baoyu-skills#baoyu-wechat-summary
@@ -146,15 +146,29 @@ where `data_root` is from EXTEND.md (default `{project_root}/wechat`).

 **Group-rename detection**: list existing folders under `{data_root}/` and find any folder whose name starts with `{group_id}-`. If one exists but the suffix differs (group was renamed), rename the existing folder to the new `{group_id}-{sanitized_new_name}` form. If a target with the new name already exists (rare), keep both and prefer the existing one for this run.

+### Step 2.5: Look up the group owner（群主）
+
+群主是谁**必须有据可查**，不能凭历史摘要、群友玩笑或印象推断（群主可能换届，历史摘要里的说法会过期）：
+
+```bash
+wx members "<group_name_or_id>" --json
+```
+
+- 检查输出中是否有 owner / role 字段标识群主；有则以此为准
+- 如果 wx-cli 版本不暴露群主信息，则查 memory.md「群基本档案」里有出处的记录；两处都没有 → **摘要里不要断言谁是群主**
+- 查到的结果与「群基本档案」不一致时以本次查询为准，更新档案并追加修订记录（注明查询日期）
+
 ### Step 3: Fetch messages

-For small batches (single-day digest, typically < 200 messages), pipe JSON into the agent directly:
+**Always redirect the fetch to a `$TMPDIR` file** — this file is the single source of truth for the whole run: Round 3's attribution audit greps it, and the statistics are computed from it. Never write the digest purely from conversation memory.
+
+For small batches (single-day digest, typically < 200 messages), you may additionally pipe JSON into the agent directly for reading:

 ```bash
 wx history "<group_name_or_id>" --since YYYY-MM-DD --until YYYY-MM-DD -n 5000 --json
 ```

-For **large batches** (weekly / monthly digests, > 200 messages), redirect to `$TMPDIR` first so the raw payload never sits in conversation context:
+For **large batches** (weekly / monthly digests, > 200 messages), the `$TMPDIR` redirect also keeps the raw payload out of conversation context:

 ```bash
 wx history "<group_name_or_id>" --since YYYY-MM-DD --until YYYY-MM-DD -n 5000 --json > "$TMPDIR/wx-messages.json"
@@ -171,7 +185,7 @@ Notes:
 - Filter the returned messages by their `timestamp` to be safe (some daemons may return adjacent days).
 - **Range splitting**: for ranges > 7 days OR > 500 messages, prefer generating per-3-day digests and then a meta-summary over forcing one giant digest — the categorization quality degrades sharply past a week's worth of unrelated topics.

-**Incremental mode**: after the fetch, drop any message whose `timestamp` is `<=` the `last_message_time` from `history.json`. If zero messages remain, tell the user "上次摘要后没有新消息，已跳过生成" and exit.
+**Incremental mode**: after the fetch, drop any message whose `timestamp` is `<=` the `last_message_time` from `history.json`, and write the filtered set back to the `$TMPDIR` file (so audits and stats run on exactly what the digest covers). Caution: `last_message_time` is `MM-DD HH:MM` — plain string comparison breaks across a year boundary (12-31 vs 01-01); compare by date semantics there. If zero messages remain, tell the user "上次摘要后没有新消息，已跳过生成" and exit.

 ### Step 3.5: Parse the message schema

@@ -195,6 +209,7 @@ Notes:

 - Substitute `self_display` for every message whose `from_wxid` matches `self_wxid` (from EXTEND.md). Apply this in the leaderboard, portraits, and body text. The user MUST appear under their real display name and count toward stats — never skip them.
 - Scan all unique senders for ambiguous handles: ≤2 characters, common programming words (`nil`, `null`, `test`, `admin`, `user`, `undefined`), single emoji, or otherwise low-information. For each, run `wx contacts --query "<nick>" --json --limit 5` and pick a meaningful name in this priority: remark > nickname > wxid. Apply the substitution everywhere in the digest.
+- **硬规则**：`nil`、空白、单标点这类占位符样式的名字**绝不允许原样出现在摘要里**。contacts 查不到 remark 时，用「昵称（wxid 后 4 位）」形式区分（如 `nil（…n77g）`），确保读者知道这是谁、且与其他人不混淆。已解析过的映射写入 memory.md「群基本档案」，下期直接复用不再重查。

 ### Step 3.7: Load user profiles

@@ -224,7 +239,7 @@ See [references/profiles.md](references/profiles.md) for the full file format.

 除了按人的 profiles，每个群还有一份全局事实记忆 `{folder}/memory.md`，记录群友指正过、确认过的客观事实（如"某个报错提示的真实原因"、"某产品名的正确写法"、"某事件的实际经过"）。

-1. 如果 `memory.md` 存在，读入作为内部背景知识（不写入最终摘要）
+1. 如果 `memory.md` 存在，读入作为内部背景知识（不写入最终摘要）。「群基本档案」小节记录群主、昵称映射等长期事实，写摘要时直接引用（群主以 Step 2.5 的查证结果为最终依据）
 2. **写摘要时必须遵守其中的事实修正**——上一期摘要里说错、已被群友指正的说法，这一期绝不能再犯。例如记忆中有"『当前微信版本不支持』是 AI Agent 无法获取微信链接导致的提示，普通用户可正常打开"，就不能再把它当成"骗点击"的梗来写
 3. 记忆条目是事实约束，不是风格指令——它只纠正"说什么"，不改变 normal/roast 两个版本各自的语气和写法
 4. 标注为「群友说法（未验证）」的条目，引用时保留这个限定，不当成已证实的事实陈述
@@ -280,7 +295,7 @@ Internal working format (not written to the final file):

 ```
 == 话题清单（共 N 条消息）==
-1. [HH:MM-HH:MM] 话题名称（参与者：A, B, C）— 一句话概括（锚点 id：54052, 54055, 54063）
+1. [HH:MM-HH:MM] 话题名称（参与者：A, B, C）— 一句话概括（锚点：54052 宝玉:"原话片段" → 54063 鸭哥:"回应片段"）
 2. [HH:MM-HH:MM] 话题名称（参与者：D, E）— 一句话概括（锚点 id：54100-54112）
 ...

@@ -300,7 +315,7 @@ Topic principles:
 - Topic-switch signals: time gap > 30 min, participant change, content jump.
 - 2+ participants OR substantive content qualifies as a topic; pure emoji-banter does not.
 - **Strict attribution**: each topic must record "who said what". Don't fuse adjacent messages from different senders just because they're close in time — when minutes apart or interleaved with others, split into separate topics. Prefer two topics over one wrongly-merged topic.
- **Carry anchor IDs**: list the key message IDs for each topic. In Round 2, jump back to these IDs in the raw messages and verify content, don't guess from context. If `quote_id` / `reply_to` is present, use the ID chain — that's the most reliable attribution.
+- **Carry anchor IDs with verbatim quotes**: for key messages, record `id 发言人:"原话片段"` — sender and quote fragment **copied verbatim from the raw messages**, not paraphrased. In Round 2, jump back to these anchors and verify content, don't guess from context. If `quote_id` / `reply_to` is present, use the ID chain — that's the most reliable attribution. Pinning "who said what" at the skeleton stage is the first line of defense against misattribution (张冠李戴).

 **Flag-for-images criteria** (any one triggers): an explicit comment on an image (`看发型是X？`, `这是谁？`, `笑死`), multiple people piling onto the same image without saying what it is, an image as the core information (晒单/截图/资料), an explanatory line right after an image (`gpt-image-2`, `太可怕了`), or cross-sender ambiguity (B says "这个看着像 X" but the previous image is from A).

@@ -340,6 +355,17 @@ Walk the Round 1 skeleton against the finished digest. Check:
 - Quotes, names, product/tool names preserved verbatim?
 - Categorization makes sense — is anything in the wrong bucket?

+**Attribution audit (mandatory — never skip)**: for every direct quote (text in quotation marks) and every "X 说 / X 发 / X 分享" attribution in the draft, grep the raw `$TMPDIR` messages file and confirm the words actually came from that sender:
+
+```bash
+grep "原话片段" "$TMPDIR/wx-messages.json"   # or jq 'map(select(.content | contains("原话片段")))'
+```
+
+- Quote not found in the file → paraphrase drift or invented memory; restore the original wording or cut it
+- Quote found but sender doesn't match → misattribution; fix the name
+- Audit BOTH versions (normal + roast) if both were generated
+- Record a one-line verdict in working notes: `归因校验：共 N 处引用，通过 X 处，修正 Y 处`
+
 Fix in place. When clean, confirm and proceed.

 ### Step 7: Save the digest file(s)
@@ -409,6 +435,18 @@ Counts, frontmatter updates, append-only rules for quotes and events, and privac

 更新画像后，扫描本期消息，看是否有需要写入/修订 `{folder}/memory.md` 的事实修正。这一步要**保守**：宁可漏记，不可乱记。

+**这一步必须执行、必须留痕，不允许静默跳过。** 按以下流程扫描：
+
+1. 关键词初筛（对 `$TMPDIR` 消息文件跑一遍，圈出候选消息）：
+```bash
+grep -nE "错了|不对|纠正|搞错|其实是|不是.*是|瞎说|胡说|张冠李戴|谁是群主|群主是" "$TMPDIR/wx-messages.json"
+```
+2. 补充人工检查两类高概率位置：
+   - 所有**回复摘要消息**的 reply（Step 3.8 检测到的 in-chat digest，指向它的引用都是候选）
+   - @bot 请求里带指正性质的（Step 3.9 清单）
+3. 逐条按下方门槛判断是否写入
+4. 无论写入几条，最终报告里必须有一行结论：`memory 扫描：候选 N 条 → 写入 M 条`（0 也要写）——强制留痕是为了防止这一步被习惯性跳过
+
 #### 什么算"值得记的事实修正"

 典型场景：上一期摘要里有个说法（梗、归因、解释），群友在本期指出它不对，并给出了正确解释。例如摘要把"当前微信版本不支持"写成骗点击的链接，群友指正这其实是 AI Agent 无法获取微信链接时才出现的提示，普通人能正常打开——这就该记。
@@ -448,6 +486,11 @@ Counts, frontmatter updates, append-only rules for quotes and events, and privac
 ```markdown
 # 群级事实记忆 — {群名}

+## 群基本档案
+- 群主：{昵称}（{wxid}，查证于 YYYY-MM-DD，来源 wx members / 群友确认）
+- 昵称映射：{占位符昵称} = {remark/真名}（{wxid}）
+- {其他长期有效的群级事实：bot 的称呼、群名由来等}
+
 ## 事实修正
 - "当前微信版本不支持" 是 AI Agent/机器人无法获取微信链接时的提示，普通用户可正常打开，不是骗点击的链接。（指正：消失的大叔，2026-06-12，id 54321；另有 2 人附和）

@@ -470,7 +513,8 @@ Profile updates are easy to forget once the digest is on disk. Before reporting
 - [ ] `{folder}/history-digests.jsonl` appended one line (if `include_normal`)
 - [ ] `{folder}/profiles/{wxid}-*.md` updated for every user with 3+ messages (if `include_normal`)
 - [ ] `{folder}/profiles-roast/{wxid}-*.md` updated for every user with 3+ messages (if `include_roast`)
- [ ] `{folder}/memory.md` checked against this batch's corrections — updated if any passed the Step 8.6 threshold, untouched otherwise
+- [ ] `{folder}/memory.md` checked against this batch's corrections — updated if any passed the Step 8.6 threshold, untouched otherwise; the final report includes the `memory 扫描：候选 N 条 → 写入 M 条` verdict line
+- [ ] Round 3 attribution audit ran, with its `归因校验：…` verdict line in working notes

 If any item is unchecked, finish it before declaring success. Don't ship a digest with a stale `history.json` — incremental mode depends on it.

@@ -43,6 +43,8 @@ Example:
  - Exclude system messages and revoked messages (`[系统]`, `revokemsg`).
  - For the `self_wxid` user, substitute `self_display` from EXTEND.md before counting/displaying.
  - Resolve ambiguous nicknames (per SKILL.md Step 3.6) before tallying so the same person isn't double-counted.
+  - **Counts must be computed mechanically** from the `$TMPDIR` messages file (e.g. `jq 'group_by(.from_wxid) | map({name: .[0].from_nickname, n: length}) | sort_by(-.n)'`) — never estimated by eyeballing. Total and per-person counts both.
+- **Incremental runs must show the precise coverage window**: day-granular date ranges share their boundary day with the previous digest, which readers misread as overlap. Add a line right after the message count: `⏱ 覆盖区间: MM-DD HH:MM ~ MM-DD HH:MM` (first and last included message timestamps).

 Example: