AI Trends Digest · dsh-ai-trends-digest

A plugin for DeepSeek Harness (DSH) that turns "what happened in AI today?" into one clean daily digest: embodied AI, large models, industry, and papers, written to Markdown + HTML.

New to DSH? It is the desktop agent you are chatting with. A plugin adds one capability to it. This one adds a tool named ai_trends_digest.

中文说明：README.zh.md

What you get

Ask the model in plain language:

今天具身智能和大模型有什么值得看的？

and it calls ai_trends_digest, which:

1. fetches 23 RSS / Atom / arXiv feeds (Chinese + English),
2. keeps only the last 42 hours, removes duplicates,
3. classifies each story and guarantees slots per topic so robotics is never buried by funding news,
4. writes <date>.md, <date>.html, latest.md, latest.html and a JSON snapshot.

A finished digest looks like this:

# AI 趋势日报 · 2026-11-03
> 具身智能 & 大模型 · 22/23 个订阅源 · 时间窗内 210 条 · 精选 20 条

## 🔥 今日要闻
1. **【示例】人形机器人量产线落地** — 量子位
2. **【示例】某开源模型发布 70B 版本** — 新智元
3. **【示例】机器人公司完成新一轮融资** — 雷峰网

## 🦾 具身智能
- **【示例】灵巧手触觉传感器开源** — IEEE Spectrum …

## 🧠 大模型与智能体
- **【示例】推理模型在数学基准上刷新纪录** — OpenAI …

## 🏭 产业动态
- …

## 📄 论文速递
- **【示例】arXiv:2601.01234 — 视觉-语言-动作策略** …

Why this exists

Generic AI newsletters drown robotics in funding rounds, and paper feeds drown everything in titles. This plugin fixes that with three rules:

• Topic floors — embodied: 3, industry: 2, research: 2 slots are reserved before score-based filling.
• Dedupe — same story from 量子位 and 新智元 is kept once (link normalisation + title similarity ≥ 0.62).
• Scoring — sourceWeight × 2 + title hits × 3 + body hits + recency, with a 4.5 floor.

Requirements

Thing	Version	Notes
DeepSeek Harness	>= 0.1.7	needed only for the plugin mode
Node.js	>= 18	already bundled with DSH
Network	—	plain HTTPS, no API keys

Install

Option 1 — Plugin market (fastest)

1. Open DSH → 插件 (Plugins) → 自定义插件市场 (Custom market) → 添加索引源 (Add index source).

2. Paste:

   https://wa2124184.github.io/dsh-ai-trends-digest/index.json
   

3. Refresh. AI 趋势日报 / AI Trends Digest appears under 发现 (Discover) — click install.

The market entry installs dsh-ai-trends-digest from npm. If npm reports "package not found", the release is still being published — use Option 2 meanwhile.

Option 2 — From a local folder

Clone this repo, then ask DSH:

plugin_manager action=install_bundle target="D:/path/to/dsh-ai-trends-digest"

Approve the prompt. The plugin shows up in the installed list, and the tool is available in the next session.

Option 3 — Manual patch

Add this row to your DSH profile's cordis.patch.yml:

- insert:
    - id: ai-trends-digest
      name: dsh-ai-trends-digest
      config:
        timezone: Asia/Shanghai
        windowHours: 42

Use it

Just talk:

You say	What happens
"今天 AI 圈有什么大事？"	full digest with default settings
"总结最近三天的大模型新闻"	window_hours=72
"只要 10 条，存到 D:/日报"	max_items=10, output_dir="D:/日报"
"生成 11 月 1 日的日报"	date="2026-11-01"

Direct tool call, if you prefer:

ai_trends_digest(output_dir="D:/AI日报", max_items=20)

Parameter	Type	Default	Meaning
date	string	today	archive label YYYY-MM-DD
window_hours	integer	42	only keep items newer than this
max_items	integer	20	digest size
output_dir	string	~/.dsh/ai-trends-digest	where archives are written
min_score	number	4.5	score threshold

The tool returns the item count, source health (sourcesOk/sourcesTotal), failed feeds, top-3 headlines and the Markdown path. The full text lives in the generated file.

Where the files go

<output_dir>/
├─ 2026-11-03.md        # today's digest
├─ 2026-11-03.html      # same, styled for mail/浏览器
├─ latest.md            # newest digest, stable name
├─ latest.html
├─ data/2026-11-03.json # machine-readable snapshot
└─ index.html           # archive browser

Archives older than keepDays (90 by default) are pruned automatically; set keepDays: 0 to keep everything.

Configuration

Edit the config block of the patch row (Option 3) or the plugin's settings:

Key	Default	Meaning
timezone	system	timezone used to decide "today"
windowHours	42	recency window
maxItems	20	digest size
maxPerSource	6	max items per feed
minScore	4.5	global score floor
topNewsCount	3	pinned headlines
topicFloors	{embodied:3, industry:2, research:2}	guaranteed slots
keepDays	90	archive retention (0 = never prune)
outputDir	~/.dsh/ai-trends-digest	default archive directory
feeds	built-in 23	replace the source list

Example — a bigger morning report, robotics first:

config:
  timezone: Asia/Shanghai
  windowHours: 72
  maxItems: 30
  topicFloors:
    embodied: 5
    industry: 3
    research: 3

Customize sources

feeds.js exports DEFAULT_FEEDS. A feed is one line:

{ name: '量子位', url: 'https://www.qbitai.com/feed', weight: 2.0, lang: 'zh', note: '机器人 + 大模型' }

• weight (0.8–2.0) is the ranking multiplier — raise it for sources you trust.
• Set enabled: false to mute a feed without deleting it.
• To replace the list entirely, pass feeds: [...] in the plugin config.

Built-in sources (23)

中文：量子位、新智元、雷峰网、爱范儿、极客公园、IT之家、钛媒体、InfoQ 中文、Solidot

English：OpenAI News、Google AI Blog、Google DeepMind、IEEE Spectrum (AI / Robotics)、TechCrunch AI、MIT Technology Review (AI)、MarkTechPost、NVIDIA Blog、BAIR、Microsoft Research、The Verge (AI)、The Decoder、arXiv cs.RO

Run it without DSH

core.js is plain Node.js with zero dependencies. A ready-made CLI lives in examples/run.mjs:

git clone https://github.com/WA2124184/dsh-ai-trends-digest.git
cd dsh-ai-trends-digest
node examples/run.mjs --output ./out --window 48 --max 20

It prints progress per feed, then the number of items and the paths written. Use it with cron / Windows Task Scheduler if you want a file every morning:

node "D:\tools\dsh-ai-trends-digest\examples\run.mjs" --output "D:\AI日报"

Smoke test

node test/smoke.mjs --output ./.tmp-out --window 48

This really hits the network for all 23 feeds and prints item count, source health, per-topic distribution and paths. Exit code 1 means nothing was selected (usually a quiet weekend or no network).

How it works

23 feeds ──► fetch (30s timeout, 2 retries)
          ──► parse RSS / Atom / JSON Feed
          ──► 42h window ──► dedupe (link + title similarity)
          ──► classify (embodied / llm / industry / research)
          ──► score ──► topic floors ──► fill by score
          ──► Markdown + HTML + JSON archives

Troubleshooting

Symptom	Why	Fix
First run takes 20–40 s	23 feeds are fetched in parallel	normal
A few feeds show as failed	CDN 502s / timeouts on the source side	retry later; it never aborts the digest
论文 section is empty	arXiv has no fresh cs.RO submissions (weekends)	expected
Same story appears twice	titles differ more than the 0.62 threshold	lower the threshold in core.js if it bothers you
Titles contain typos	they come from the source	plugin never rewrites titles
Market install says package not found	npm release still rolling out	install from the local folder (Option 2)

Project layout

dsh-ai-trends-digest/
├─ index.js            # plugin entry: registers ai_trends_digest
├─ core.js             # fetch / parse / dedupe / score / render / archive
├─ feeds.js            # 23 default sources
├─ cordis.patch.yml    # config row inserted on install
├─ locale/{en,zh}.json # market metadata
├─ icon.svg            # plugin icon
├─ market/index.json   # format-1 market index (single entry)
├─ examples/run.mjs    # standalone CLI
└─ test/smoke.mjs      # networked smoke test

Contributing

Issues and PRs are welcome — especially new feed URLs (must be RSS/Atom JSON, not HTML pages) and keyword improvements for classification. Please run node test/smoke.mjs before opening a PR.

License

MIT