dsh-favorites-digest

dsh-favorites-digest

中文 | English

A DeepSeek Harness plugin that exports your own Douyin and Xiaohongshu web favorites into Markdown (topic groups + optional LLM digest). It reuses a logged-in Playwright / Chrome session. Do not use it on anyone else's account.

Compliance — read first

This is a personal backup tool, not a scraper service. Keep it that way:

• Your own account, your own data. Export only favorites from an account you logged into yourself (QR scan). Never point it at other people's accounts or data.
• Respect the platforms. Automated access may violate Douyin / Xiaohongshu terms of service even for your own data. Risk control may throttle or ban your account — that risk is yours; the tool works from your own logged-in session and solves no captchas for you.
• No commercial use, no redistribution. Exported content belongs to its original authors. Keep outputs for personal reference; do not resell, republish, or bulk-redistribute them.
• Nothing leaves your machine. Collection, artifacts, and digests all run and stay locally.

Install

Requires Node.js 20 or 22 and a Harness profile that mounts the commands and sessionQuery services (the official web / agent profiles do).

dsh plugin --profile web add dsh-favorites-digest
npx playwright install chromium

Local checkout:

dsh plugin --profile web add link:/absolute/path/to/dsh-favorites-digest

Then in any session:

/favorites login
/favorites

Files land in <session cwd>/dsh-favorites/favorites-<timestamp>.md (and .json). The browser profile is stored at ~/.dsh/favorites-digest/profile.

Commands

Same as other ctx.commands handlers: results stay off the model transcript (zero tokens).

Optional model tool

Off by default. Set exposeTool: true on the plugin row to register favorites_digest for the agent.

- insert:
    - id: favorites-digest
      name: dsh-favorites-digest
      config:
        exposeTool: true

Optional LLM digest

If OPENAI_API_KEY or DEEPSEEK_API_KEY is set, /favorites run appends an AI 综述 section. Compatible OpenAI-style bases work via OPENAI_BASE_URL / OPENAI_MODEL. Topic grouping still works with no key.

Open-source project identification

run/digest also scan titles and descriptions for open-source projects — explicit github.com/<owner>/<repo> links plus dashed/camelCase names like open-code-review or Windows-MCP — and verify them against the GitHub search API, so the Markdown gains a 开源项目 section with repo, star count and which favorite mentioned it. Anonymous GitHub search is rate-limited (10 req/min), so at most 8 candidates are checked with spacing and the step degrades silently on 403. Set identifyProjects: false to turn it off; digest re-renders the stored list without re-querying.

How collection works

1. Open the logged-in 收藏 surfaces (Xiaohongshu profile → 收藏 → 笔记; Douyin /user/self?showTab=favorite_collection).
2. Listen to the page's own list JSON and scrape visible cards.
3. Scroll until the count stops growing.

Site redesigns can break selectors. Captchas must be solved in the opened window.

Troubleshooting

• collect aborts with 未登录 — login detection uses hard signals (Douyin sessionid cookie; Xiaohongshu avatar-vs-登录-button probe). An abort means the session really is not logged in: re-run /favorites login and scan both QR codes; it only reports success after both platforms are confirmed.
• Xiaohongshu shows “IP 存在风险” (error 300012) during login — XHS risk control can block QR login from automated browsers on some networks. Log in inside a Chrome you launched yourself, then attach:

  "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
    --user-data-dir="$HOME/.dsh/favorites-digest/cdp" --remote-debugging-port=9222
  # then: /favorites --cdp-url http://127.0.0.1:9222
  

  Chrome ≥136 refuses --remote-debugging-port on the default profile, so the dedicated --user-data-dir is required.
• CDP attach fails with “Browser context management is not supported” — Chrome 154 removed the browser-level CDP calls older Playwright issued at attach time; requires playwright ≥ 1.64.
• Why hard-fail when not logged in? Logged-out collection used to silently export the recommendation feed as your “favorites”. It now aborts instead of fabricating data.
• Running the TS sources with Node type stripping fails (parameter property is not supported in strip-only mode) — expected; the sources target the bundler. Use npm run bundle output (lib/), which is what the published plugin runs.

Development

npm ci
npx playwright install chromium
npm run typecheck
npm test
npm run bundle