wechat-feeds

name	wechat-feeds
description	Maintain a local, incrementally-updated library of WeChat Official Account (公众号) articles as Markdown + a SQLite index, and pull new posts on demand. Use when the user wants to subscribe to 公众号, fetch/update 公众号 articles, build a searchable feed of specific accounts, or says "update my feeds" / "拉一下公众号" / "增量更新公众号文章". Bundles a self-contained Bun scraper (tool/) that drives a logged-in Chrome via a tiny CDP proxy.
allowed-tools	Bash, Read

wechat-feeds

Build and maintain a local library of WeChat Official Account (微信公众号) articles. The bundled tool (tool/) logs into the 公众号 admin platform in your own Chrome, then pulls each subscribed account's article list and full text into data/<account>/<date>-<title>.md, with a SQLite index (feeds.db) for querying.

Two things this skill does:

1. Set up / maintain the library — install, log in, subscribe to accounts.
2. Incrementally update — pull new articles up to now (the update-feeds workflow below), including automatic session re-login handling.

How it works (architecture)

CLI (bun, tool/src)  ──HTTP──▶  cdp-proxy.mjs  ──WebSocket(CDP)──▶  your Chrome (logged into mp.weixin.qq.com)

• The scraper never talks to WeChat directly. It asks the CDP proxy to run fetch() inside a logged-in mp.weixin.qq.com tab, so every request is same-origin and carries the real login cookie — no separate auth, no easy anti-crawl trip.
• Login state is a token (in session.json) plus the browser's cookie. The token expires often; the cookie lasts a few days. refresh re-derives the token from a still-valid cookie; when the cookie itself dies you re-login in the browser once.

Prerequisites

• Bun (runs the CLI; uses bun:sqlite) — https://bun.sh
• Node.js 22+ or Bun (runs the CDP proxy; needs a global WebSocket)
• Google Chrome you can start with remote debugging
• A WeChat Official Account you can log into mp.weixin.qq.com with (any account works — the 图文素材 search reads other public accounts' published articles)

One-time setup

All commands run from the tool/ directory.

cd tool
bun install

# 1. Start Chrome with remote debugging (keep it running).
#    macOS:
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222
#    Linux:   google-chrome --remote-debugging-port=9222
#    Windows: chrome.exe --remote-debugging-port=9222

# 2. Start the CDP proxy (separate terminal; keep it running).
node cdp-proxy.mjs        # or: bun cdp-proxy.mjs

# 3. Log in — opens the 公众号 login page; scan with WeChat.
bun run src/cli.ts login

# 4. Subscribe to accounts (by name or by a share link to one of their articles).
bun run src/cli.ts add 公众号名称
bun run src/cli.ts add https://mp.weixin.qq.com/s/xxxxx
bun run src/cli.ts list

Optional — wire login-expiry alerts to a channel of your choice (Feishu / Slack / email / desktop). When unset, the alert just prints to the console:

# macOS desktop notification example:
export WECHAT_FEEDS_NOTIFY_CMD='osascript -e "display notification \"$WECHAT_FEEDS_MSG\""'

The update-feeds workflow

Incrementally update all subscribed accounts to now. Run the steps in order; never do anything destructive (don't clear cookies, don't log out).

1. Current time & pre-update baseline

• TZ=Asia/Shanghai date '+%Y-%m-%d %H:%M' — note current time.
• Latest article date already in the DB (the "before" baseline):

  cd tool && sqlite3 -readonly feeds.db "SELECT date(MAX(publish_time),'unixepoch','+8 hours') FROM articles WHERE fetched_at IS NOT NULL;"
  

2. Check the runtime

• CDP proxy healthy: curl -s http://localhost:3456/health → expect "connected":true. If no response, the proxy isn't running — start it (node cdp-proxy.mjs) and make sure Chrome is up with remote debugging.
• An mp tab exists: curl -s http://localhost:3456/targets | grep -o 'mp.weixin.qq.com' | head -1. If empty, open one and wait:

  curl -s "http://localhost:3456/new?url=https://mp.weixin.qq.com/" ; sleep 6
  

3. Run the incremental pull (background)

Pulls take a few minutes — run in the background from tool/:

• If the user passed args, pass them straight through: bun run src/cli.ts pull <args> (e.g. --since=2026-05-01).
• Otherwise default to the last ~two weeks (efficient; paging stops automatically once it reaches older articles):

  SINCE=$(TZ=Asia/Shanghai date -v-14d '+%Y-%m-%d'); bun run src/cli.ts pull --since=$SINCE
  

  (GNU date: SINCE=$(date -d '14 days ago' '+%Y-%m-%d').)

4. Monitor output; handle re-login (the key step)

Track the background command's output:

• If you see 🔗 登录链接已通知: the browser login has fully expired (the cookie died — happens roughly every few days; normal, not an error). This pull round skips every account. Re-login and re-run, in this order:
  1. Tell the user: open https://mp.weixin.qq.com/ in the CDP-controlled Chrome (the one started with --remote-debugging-port) and log in (the browser usually remembers the account — often no QR scan needed).
  2. After they confirm login, run bun run src/cli.ts refresh first to sync the token (required — pulling with the old token gives 200040 invalid csrf token). Wait for ✅ 刷新成功.
  3. Then re-run step 3's pull with the same --since window. This time the session is valid and it pulls in full.
• If no re-login notice appears, the session is valid — just wait for the pull to finish (exit 0). Don't re-trigger pull while it's running.

5. Report

After the pull completes:

• Latest date & total now:

  sqlite3 -readonly feeds.db "SELECT date(MAX(publish_time),'unixepoch','+8 hours'), COUNT(*) FROM articles WHERE fetched_at IS NOT NULL;"
  

• Summarize new-article counts per account from the output.
• Report to the user: updated from [before date] to [latest date], N new articles, and list the accounts with the most additions.

Reading the library

• Markdown body: data/<account nickname>/<YYYY-MM-DD>-<title>.md — YAML frontmatter (title/author/date/url/digest) + Markdown text.
• SQLite index (feeds.db): query what exists, who posted, when, and where the file is.

accounts(fakeid PK, nickname, alias, added_at)
articles(id PK, fakeid FK, title, url, author,
         publish_time,   -- Unix SECONDS  → datetime(publish_time,'unixepoch','+8 hours')
         fetched_at,     -- Unix MILLIS   → datetime(fetched_at/1000,'unixepoch','+8 hours'); NULL = metadata only, body not fetched
         file_path)

Use fetched_at IS NOT NULL to mean "body actually fetched" (more reliable than file_path).

CLI reference

bun run src/cli.ts login              log into the 公众号 platform (QR)
bun run src/cli.ts add <name|url>     subscribe to an account
bun run src/cli.ts list               list subscriptions + article counts
bun run src/cli.ts pull               pull new articles (default max 100/account)
bun run src/cli.ts pull --today       only today's
bun run src/cli.ts pull --date=MM-DD  a specific date
bun run src/cli.ts pull --since=YYYY-MM-DD   from a date to now
bun run src/cli.ts pull --max=50      cap per account
bun run src/cli.ts pull --concurrency=6   accounts pulled in parallel (default 4; higher = faster but more rate-limit risk)
bun run src/cli.ts refresh            re-derive token from a live cookie
bun run src/cli.ts relogin            send a re-login alert (cookie fully dead)
bun run src/cli.ts keepalive          for a scheduler (cron/launchd): refresh, alert if dead
bun run src/cli.ts status             show login status

Notes & red flags

• Rate limits: WeChat's appmsgpublish endpoint is frequency-controlled. The tool paces itself (per-article + per-page sleeps) and defaults to concurrency 4. Pushing concurrency high risks anti-crawl. If accounts start failing, lower it and re-run.
• Accounts >14 days behind may be missed by the default window — pass a wider --since=YYYY-MM-DD.
• Keep the Chrome mp tab open (can be minimized) — every API call runs inside it.
• Nothing secret is committed: feeds.db, session.json, config.json, and data/ are gitignored. Your subscription list lives in feeds.db, not in source.
• Use responsibly: this drives your own logged-in session against WeChat's platform. Respect WeChat's Terms of Service, article copyright, and reasonable request rates. For personal reading/archival of accounts you follow.