auto-paper-collecter (skill)

A self-hosted research-literature radar that runs inside a coding agent. The Python scripts do only the deterministic work (API fetch, dedup, render, email). YOU — the assistant running this skill — do all the judgement work: query expansion, computer-science relevance filtering, Chinese summaries, and hot-topic synthesis. That means no AI API key is needed — whichever model is running this skill (Claude in Claude Code, GPT in Codex, …) is the LLM.

Layout

skill/
├── SKILL.md
├── scripts/   common.py · fetch.py · render.py · notify.py   (stdlib only)
├── state/     config.json · (queries/candidates/curated/trends/seen .json)
└── digests/   YYYY-MM-DD.md  +  .html

Run scripts from scripts/: cd skill/scripts && python3 <script>.py

Config — state/config.json

• keywords: up to ~3 topic strings to track.
• domain: the field to constrain relevance to (default computer science).
• sources: toggle arXiv / Crossref / Semantic Scholar / GitHub / HuggingFace / PapersWithCode / RSS.
• lookback_days: how far back to fetch (dedup stops repeats anyway).
• max_per_source, rss_feeds.

When the user asks to change keywords / sources / field, edit this file and confirm the change back to them.

Optional env vars (never stored in the repo): SEMANTIC_SCHOLAR_KEY (lifts S2 rate limits), GITHUB_TOKEN (lifts GitHub limits), SMTP_* / EMAIL_TO (email), and push channels — TELEGRAM_BOT_TOKEN/TELEGRAM_CHAT_ID, SLACK_WEBHOOK_URL, WECHAT_WEBHOOK (企业微信群机器人) or SERVERCHAN_KEY (Server酱).

The run pipeline — follow IN ORDER

1 · Read config & expand queries (you)

Read state/config.json. For each keyword, think of 2–3 associative English search queries — synonyms, full forms, adjacent sub-topics — so recall isn't limited to the literal term (e.g. C2Rust → ["C2Rust", "C-to-Rust translation", "migrating legacy C code to Rust"]). Write them to state/queries.json as {"<keyword>": ["q1", "q2", ...], ...}.

2 · Fetch candidates (script)

cd skill/scripts && python3 fetch.py

Fetches every enabled source for those queries, drops anything already in state/seen.json or older than lookback_days, and writes state/candidates.json. If it reports 0 candidates, tell the user "暂无新文献" and stop (nothing else to do).

3 · Filter relevance & summarize (you)

Read state/candidates.json. For each item decide: is it (a) computer-science and (b) genuinely on-topic for its topic keyword? Drop the rest (medical "translation", finance "AI", random GitHub star-lists, etc.). For every kept item write a concise Chinese summary and assemble state/curated.json — a list of objects:

{"source","topic","title","url","venue","authors","published",
 "tldr":"一句话核心 (<=60字)","method":"方法简述 (<=80字)",
 "contributions":["核心贡献1","核心贡献2"]}

Keep papers first, GitHub repos last (they are a supplementary signal). If a source gave a tldr already, you may build on it.

GitHub items are repos, not papers — don't over-summarize them. Use the repo description (its abstract) as the tldr and leave method/contributions empty. fetch.py already keeps only repos with ≥10 stars, ranked by stars, so they tend to be substantive (course / framework / awesome-list), not personal noise.

4 · Hot-topic synthesis (you, optional but recommended)

Cluster the kept items into a handful of coarse CS sub-fields (自然语言处理 / 计算机视觉 / 系统与编译 …; merge aggressively). Write state/trends.json: {"top": [{"name","delta": <count>, "summary": "<=80字方向总结", "papers": ["title", ...]}, ... up to 3]}.

5 · Render the digest (script)

cd skill/scripts && python3 render.py

Writes digests/YYYY-MM-DD.md + .html from curated.json (+ trends.json) and records everything shown into seen.json so it won't repeat.

6 · Notify (script, optional)

cd skill/scripts && python3 notify.py     # emails the HTML digest if SMTP_* env is set

7 · Report back (you)

Tell the user how many papers were kept, the top hot directions, and the digest path. Offer to open the HTML or adjust keywords.

Notes

• Scripts are pure Python stdlib — no pip install required.
• fetch.py already filters garbage future dates and de-duplicates across runs.
• This skill is the agent-driven counterpart of the project's FastAPI web dashboard; both share the same sources and pipeline philosophy.