---

理念

所有情报需求 —— AI 资讯、股票异动、补货监控、显卡行情、羊毛优惠 —— 形状都一样: 盯住一些源,留下重要的,忽略其余的,该叫你的时候叫你。 现有工具各给一片: 一个爬虫、一个订阅源、一个 diff 监视器。世事 给的是全链路,配置驱动,并且天生 就是给 AI agent 开的。

任何情报品类 = 一份 YAML 文件。世事 负责抓取(六级引擎自动降级)、分类去重、 打分(先关键词粗筛 —— 零 token;可选 LLM 精评),把重要的推到你的即时通讯。 YAML 本身由你的编码 agent 照规范现场生成:读 schema、写配置、myssia test 试抓 验证、凭 myssia doctor 输出自行修复失效源。

人做决策,AI 做其余。

市面空白:为什么是 世事

现有工具	给你什么	缺什么
Crawlab / Kestra	爬虫/工作流编排	只是载具 —— 没有分类、去重、推送
changedetection.io	网站变更监控	只盯 diff —— 没有采集管线、分析与打分
RSSHub	把站点转成 RSS	只做源转换 —— 没有过滤、打分与投递
单一用途盯盘工具(凭证/羊毛各一款)	各覆盖一个细分品类	品类写死;换目标 = 换工具
世事	品类无关的采集→分类→分析→推送全链,配置驱动	—

核心亮点

🧠 是情报,不只是爬虫 七大类关键词分类器(零 token,以独立包 myssia-classifier 发行)+ 可选 LLM 精评(价值/相关性/ 可信度 0–10)。阈值分级路由:score ≥ 8 立即推,≥ 5 进早晚摘要,其余归档。	🪜 六级采集降级梯 direct_api → static_html → crawl4ai / firecrawl → Scrapling → 隐身浏览器 → LLM 浏览器。源永不写死引擎:某一级失灵,下一级顶上,胜出引擎 按源记忆(存 SQLite —— 永不回写你的 YAML)。
📬 推送尊重你的注意力 URL 键去重注册表 + 早/晚摘要槽位 —— 生产验证过的语义;同一条情报你永远不会 收到第二遍。通道:飞书卡片、Telegram、webhook、stdout。	🔐 凭据永不落明文 凭据永不进 YAML —— 只允许 env:VAR / keychain:myia/<scope>/<name> 引用; 配置文件里出现明文凭据,加载即拒。落 macOS Keychain / Windows DPAPI。
🤖 天生 Agent 友好 全量文档化的 12 节 YAML schema、处处有缺省值、每条命令都有 --json (stdout 恒为恰好一份 JSON 文档)、结构化诊断 —— 每个接口都为 agent 可驱动 而设计,不只是给人用的。	🔁 反馈闭环自我调优 对推送标记有价值 / 无价值(CLI 现已可用;桌面卡片内按钮已随桌面对齐批次 落地 main、随下个发布版交付;Telegram/飞书回调接收已就绪)—— 负反馈持续 回写,调优盯盘权重与阈值。
⏰ 定时任务:到点跑一遍 + 发摘要 myssia cron serve 常驻、桌面端内置 ticker、外接 crontab 手动 cron tick —— 同一底座(tick 文件锁 + fire 认领互斥)多宿主共存;到点自动跑一遍品类 管线,运行摘要定向投递(本地留档 / 飞书 / Telegram 等平台 spec)。schedule 吃自然语言(every monday 9am)也吃 5 段 cron(POSIX 周几)。

开箱即用:SQLite 单文件存储(无 Redis、无 Postgres、无常驻守护),保留期 自动清理 + 定期 VACUUM,进程内调度。

快速开始

git clone https://github.com/xinzhuzi/myia
cd myssia
uv sync                     # uv workspace(主口径):一并装好 myssia 与 myssia-classifier
uv run myssia --version       # myssia 0.0.1

本仓是 uv workspace(myssia-classifier 是 workspace 成员):裸 pip install -e . 装不齐依赖,源码安装请用 uv sync;pip install myssia 待 PyPI 上架后可用。

重型采集引擎是可选 extras;缺引擎时沿梯子优雅降级,给出结构化 dependency_missing 错误而不是崩溃:

uv sync --extra crawl4ai    # L3 JS 渲染引擎
uv sync --extra llm         # LLM 精评 / 事件聚合

跑第一个品类 —— 零凭据的完整配置就一个小文件:

cat > plugins/demo-min.yaml <<'YAML'
id: demo-min
name: Minimal demo
schedule: "0 9 * * *"
sources:
  - name: example-news
    engine: static_html
    url: "https://example.com/news"
    extract:
      type: list
      item: "article"
      fields:
        title: "h2 a"
        url: "h2 a@href"
classify:
  builtin: false                 # demo items match no category; disable the filter
push:
  - channel: stdout              # zero-credential local verification
YAML

uv run myssia test plugins/demo-min.yaml --json            # 试抓:不入库、不推送
uv run myssia run plugins/demo-min.yaml --dry-run --json   # 全链演练,不推送
uv run myssia run plugins/demo-min.yaml                    # 正式跑;--loop 常驻调度

完整走读:docs/zh/getting-started.md。

PyPI 包(myssia、myssia-classifier)将走手动发布流程;在那之前请如上从源码安装, 桌面用户可直接用下方安装包。

下载安装(桌面应用)

macOS(Apple Silicon)

安装包随 GitHub Releases 发布:

1. 从 Releases 下载 myssia_<版本>_aarch64.dmg(版本号随发布更替,以 Releases 页面实际资产为准), 把 世事 拖入「应用程序」;
2. 首次打开:在「应用程序」里右键 世事 →「打开」→ 再点「打开」 (或双击被拦后到 系统设置 → 隐私与安全性 → 点「仍要打开」);
3. 安装包未做 Apple 公证(公证需付费开发者账号)—— 代码完全开源可审计, 每个安装包由 GitHub Actions 公开构建、日志可溯;右键打开一次即完成 Gatekeeper 放行,之后正常双击启动。

Windows(x64)

安装包(msi)将随 GitHub Releases 发布——Windows 构建流水线已就绪,首个 Windows 版本随下个 Release 交付;交付前 Releases 页暂无 msi(见路线图注记):

1. 从 Releases 下载 myssia_<版本>_x64.msi,双击安装(版本号随发布更替,以 Releases 页面 实际资产为准);
2. 首次运行弹 SmartScreen「Windows 已保护你的电脑」时,点 「更多信息」→「仍要运行」(未购买代码签名证书的如实代价,与 macOS 右键打开同一口径);
3. Windows Defender 误报可能:安装包内含未签名的 PyInstaller sidecar (myssia-core.exe),SmartScreen/Defender 可能告警——代码完全开源可审计, 每个安装包由 GitHub Actions 公开构建、日志可溯;如遇拦截,同样 「更多信息」→「仍要运行」,必要时在 Defender 提示里选「允许」;
4. 数据根在 %APPDATA%\MYIA(资源管理器地址栏粘贴即达),装机首跑自动 种子官方插件,首屏点「运行第一个插件」;
5. 升级:设置页「检查更新」,msi 静默(passive)安装后自动重启 (见 desktop/UPDATER.md)。

装机首跑自动种子官方插件(含零凭据演示件 myssia-demo:GitHub 新星榜), 第一次点「运行第一个插件」就出真数据;设置页「检查更新」走签名更新通道。

桌面五屏(截图为 demo 插件真实抓取数据):

AI 原生闭环

把闭环交给任何编码 agent(Claude Code、Cursor……)。 Agent Skill 是自包含速查表,一条命令安装 (myssia skill install --agent claude;也支持 cursor / zcode, --path 自定义目录,--link 以链接代替复制):

myssia skill install --agent claude         # 一次性:速查表 → ~/.claude/skills/myia/
myssia init --json                          # 拿结构化信息清单
( agent 现场写 <id>.yaml )                # 对照 12 节 schema
myssia test plugins/<id>.yaml --json        # 试抓,核对字段与去重键
myssia run plugins/<id>.yaml --dry-run      # 演练
myssia run plugins/<id>.yaml --loop         # 常驻调度
myssia doctor --json                        # 拿 findings;agent 自修后复查

架构

用户层          myssia CLI · Agent Skill · 桌面应用(Tauri,v1.1)· Web UI(规划中)
                     │
编排层          流水线:fetch → classify → dedup → analyze → enrich → push
                (进程内 APScheduler + asyncio;无外部编排器、无守护进程)
                     │
插件层          一个品类一份 YAML · 6 个官方品类 · 市场插件
                desktop 级 = 进程内 adapter(零 docker)· remote/server 级走端点
                (代理池 · 变更监控 · OSINT · 抖音 · maxun · …)
                     │
采集引擎        L1 direct_api → L2 static_html → L3 crawl4ai ⇄ firecrawl
                → L4 scrapling → L5 stealth_browser → L6 llm_browser
                (自动降级链;胜出引擎按源持久化)
                     │
分析            内置七类关键词分类器(myssia-classifier)
                + 可选 LLM 精评(价值/相关性/可信度,0–10)
                     │
存储            SQLite 单文件 · 保留期 + VACUUM · 变更基线
                     │
推送            飞书卡片 · Telegram · webhook · stdout,阈值分级路由
                (立即 / 早晚摘要 / 归档)+ 反馈闭环

文档

双语文档随仓库发行,由测试(tests/test_docs.py)锁住与代码一致 —— 每个 示例 YAML 都过真实 schema 入口加载,zh/en 两棵树结构上不许漂移:

	English	中文
Getting started	docs/en/getting-started.md	docs/zh/getting-started.md
Write a plugin	docs/en/write-a-plugin.md	docs/zh/write-a-plugin.md
Schema reference	docs/en/schema.md	docs/zh/schema.md
FAQ(伦理与边界)	docs/en/faq.md	docs/zh/faq.md
Scheduled jobs (cron)	docs/en/cron.md	docs/zh/cron.md

面向 agent 的浓缩参考:skill/SKILL.md。

路线图

核心流水线已实现且有测试覆盖 —— 1300+ 测试跑在 CI 里,无一条碰真实网络。 桌面端已可日常使用(数据通路统一、官方插件随包、开箱 demo、签名更新通道, v0.0.1 起随 Releases 交付);桌面卡片内反馈按钮、设置反馈开关、采集量趋势 已随桌面对齐批次落地 main、随下个发布版交付(反馈闭环 CLI 现已可用); Windows 构建这版未通过,Release 暂无 Windows 安装包(后续批次计划补上)。

里程碑	范围	状态
v0.1 骨架	核心流水线、direct_api/static/firecrawl 引擎、12 节 schema、变更指纹、分类器、飞书路由	✅ 已交付
v0.2 可用	SQLite 存储 + 注册表 + 保留期、完整 CLI(init/test/list/doctor)、crawl4ai L3、LLM 精评 + 预算护栏、docker compose、Telegram、钥匙链凭据	✅ 已交付
v0.3 生态	Agent Skill、插件市场(本地/远端双模)、Scrapling L4、反馈闭环(CLI + 回调接收)	✅ 已交付
v0.4 深水区	stealth_browser L5、llm_browser L6、趋势基线、事件聚合	✅ 已交付
v1.0 发布	双语文档、演示物料、GitHub 门面、公开交付	✅ 已交付
v1.1 桌面优先	Tauri 桌面壳(Python 核心以 sidecar 嵌入)、五屏 UI、进程内插件级;卡片内反馈按钮/settings 反馈开关/采集量趋势已随桌面对齐批次落地、随下个发布版交付(反馈 CLI 已可用)	✅ 已交付
v1.1.1 通路修复	桌面数据通路统一(MYIA_HOME/官方插件随包/首跑种子)、开箱 demo 插件、签名更新通道(检查更新 + 自动安装)	✅ 已交付
Web UI	同一核心上的浏览器前端	📋 规划中

伦理与边界

世事 默认讲礼貌:尊重 robots.txt、限速采集、凭据永不落明文。要求真人验证 (验证码、手机号)的源会以结构化错误失败 —— 世事 不做绕过。完整表述见 中文 FAQ · English FAQ。

社区

• Bug 与需求:issue 模板
• 参与贡献:CONTRIBUTING.md
• 安全策略与凭据处理设计:SECURITY.md

致谢

世事 自有代码全部原创(MIT),但站在巨人的肩膀上 —— 以依赖、插件后端与设计 参考的形式接入:

• crawl4ai —— L3 采集引擎(可选依赖)
• Scrapling —— L4 自适应反爬引擎(可选依赖)
• Firecrawl —— L3 云端/自建渲染后端(可选依赖,以 API 调用)
• Skyvern —— L6 LLM 浏览器兜底(可选依赖)
• changedetection.io —— 源管理与 diff 交互参考,myssia-monitor 插件后端
• RSSHub —— 「一切皆源」的哲学参考
• jhao104/proxy_pool —— myssia-proxy 插件后端
• Photon —— myssia-osint 插件后端(以 git 子模块引入)
• Douyin_TikTok_Download_API —— myssia-douyin 插件后端
• Maxun —— myssia-maxun 插件后端
• Tauri —— 桌面壳(Python 核心以 sidecar 嵌入)

除明确标注的 git 子模块外,不复制任何上游源码进本仓库;依赖接入策略见 CONTRIBUTING.md。

许可证

MIT © 2026 xinzhuzi

---

The idea

Every intelligence need — AI news, stock moves, restocks, GPU prices, freebies and deals — has the same shape: watch some sources, keep what matters, ignore the rest, tell me when it counts. Existing tools each give you one slice: a crawler, a feed, a diff watcher. 世事 is the whole chain, config-driven, and built to be driven by your AI agent.

Describe any category as one YAML file. 世事 fetches it (six engines on an auto-degrading ladder), classifies and dedups it, scores it (keywords first — zero tokens; optional LLM for precision), and pushes what matters to your messaging apps. The YAML itself is written by your coding agent: it reads the schema, generates the config, trial-fetches with myssia test, and repairs broken sources on its own from myssia doctor output.

Weixin outbound is a bridge via a local Hermes-Agent install — without one, the weixin channel is unavailable (myssia itself holds zero WeChat credentials).

Humans decide; AI does the rest.

Why 世事

Existing tools	What they give you	What they miss
Crawlab / Kestra	crawler & workflow orchestration	a vehicle only — no classification, dedup or push
changedetection.io	website change monitoring	watches diffs — no collection pipeline, no analysis, no LLM scoring
RSSHub	turns sites into RSS feeds	source conversion only — no filtering, scoring or delivery
Single-purpose watchers (credential / deal trackers)	one niche category each	category-locked; every new target means new tooling
世事	category-agnostic fetch → classify → analyze → push, all config-driven	—

Highlights

🧠 Intelligence, not just crawling A seven-category keyword classifier (zero tokens, shipped as the standalone myssia-classifier package) plus optional LLM enrichment scoring value / relevance / credibility 0–10. Thresholds route the result: score ≥ 8 pushes immediately, ≥ 5 waits for the AM/PM digest, the rest is archived.	🪜 A six-engine fetch ladder direct_api → static_html → crawl4ai / firecrawl → Scrapling → stealth browser → LLM browser. Sources never hard-code an engine: when one rung fails, the next takes over, and the winning engine is remembered per source (in SQLite — never written back into your YAML).
📬 Push that respects your attention URL-key dedup registry + AM/PM digest slots — production-proven semantics; you never receive the same item twice. Feishu card, Telegram, webhook and stdout channels.	🔐 Secrets stay secret Credentials never live in YAML — only env:VAR / keychain:myia/<scope>/<name> references. A plaintext credential in a config file is rejected at load time. macOS Keychain / Windows DPAPI backed.
🤖 Agent-native by design A fully documented 12-section YAML schema with defaults everywhere, --json on every command (stdout is always exactly one JSON document), structured diagnostics. Every interface is built so an agent can drive it — not just a human.	🔁 A feedback loop that tunes itself Mark pushed items valuable / not valuable (CLI today; in-card desktop buttons have landed on main in the desktop-parity batch and ship with the next release; Telegram/Feishu callback receivers are ready) — negative feedback retunes watchlist weights and thresholds over time.
⏰ Scheduled jobs: run it on a clock, deliver the summary myssia cron serve daemon, the desktop app's built-in ticker, or a manual cron tick from system crontab — one substrate coexisting under a tick file lock + fire claims. Each due fire runs the category pipeline once and delivers the run summary (local archive / Feishu / Telegram and other platform specs). Schedules take natural language (every monday 9am) or 5-field cron (POSIX dow).

Batteries included: SQLite single-file storage (no Redis, no Postgres, no daemon), retention + auto-VACUUM, in-process scheduling.

Quickstart

git clone https://github.com/xinzhuzi/myia
cd myssia
uv sync                     # uv workspace (primary): installs myssia + myssia-classifier
uv run myssia --version       # myssia 0.0.1

This repo is a uv workspace (myssia-classifier is a workspace member): a bare pip install -e . won't pull it in — install from source with uv sync; pip install myssia becomes available once the package lands on PyPI.

Heavy fetch engines are optional extras; a missing engine degrades gracefully down the ladder with a structured dependency_missing error instead of crashing:

uv sync --extra crawl4ai    # L3 JS-rendered page engine
uv sync --extra llm         # LLM enrich scoring / event aggregation

Run your first category — a complete, zero-credential config in one small file:

cat > plugins/demo-min.yaml <<'YAML'
id: demo-min
name: Minimal demo
schedule: "0 9 * * *"
sources:
  - name: example-news
    engine: static_html
    url: "https://example.com/news"
    extract:
      type: list
      item: "article"
      fields:
        title: "h2 a"
        url: "h2 a@href"
classify:
  builtin: false                 # demo items match no category; disable the filter
push:
  - channel: stdout              # zero-credential local verification
YAML

uv run myssia test plugins/demo-min.yaml --json            # trial fetch: no DB, no push
uv run myssia run plugins/demo-min.yaml --dry-run --json   # full rehearsal, no push
uv run myssia run plugins/demo-min.yaml                    # real run; add --loop for scheduling

Full walk-through: docs/en/getting-started.md.

PyPI packages (myssia, myssia-classifier) will publish via a manual release workflow; until then, install from source as above — or grab the desktop installer below.

Download & install (desktop app)

The macOS (Apple Silicon) installer ships via GitHub Releases:

1. Download myssia_<version>_aarch64.dmg from Releases (the version token rotates per release — the Releases page is authoritative) and drag 世事 into Applications;
2. On first launch: right-click 世事 in Applications → Open → Open (or, after a blocked double-click: System Settings → Privacy & Security → Open Anyway);
3. The package is not Apple-notarized (notarization requires a paid developer account) — the code is fully open-source and auditable, and every installer is built in public by GitHub Actions with traceable logs. One right-click open clears Gatekeeper; subsequent launches open normally.

A fresh install auto-seeds the official plugins (including the zero-credential demo myssia-demo: GitHub's new-star board), so the first click of "run your first plugin" shows real data; the settings screen offers "Check for updates" over a signed update channel.

The five desktop screens (fed by real demo-plugin data):

The AI-native loop

Hand the loop to any coding agent (Claude Code, Cursor, …). The Agent Skill is a self-contained cheat sheet, installed with one command (myssia skill install --agent claude; also cursor / zcode, --path for a custom dir, --link to symlink instead of copy):

myssia skill install --agent claude         # one-time: skill sheet → ~/.claude/skills/myia/
myssia init --json                          # structured checklist of what to collect
( agent writes <id>.yaml )                # against the 12-section schema
myssia test plugins/<id>.yaml --json        # trial fetch, inspect fields + dedup keys
myssia run plugins/<id>.yaml --dry-run      # rehearsal
myssia run plugins/<id>.yaml --loop         # scheduled operation
myssia doctor --json                        # findings; the agent repairs and re-checks

Architecture

User layer      myssia CLI · Agent Skill · desktop app (Tauri, v1.1) · Web UI (planned)
                     │
Orchestration   Pipeline: fetch → classify → dedup → analyze → enrich → push
                (in-process APScheduler + asyncio; no external orchestrator, no daemon)
                     │
Plugin layer     One YAML per category · 6 official categories · market plugins
                desktop tier = in-process adapter (zero docker) · remote/server tier
                (proxy pool · changedetection · OSINT · douyin · maxun · …)
                     │
Fetch engines   L1 direct_api → L2 static_html → L3 crawl4ai ⇄ firecrawl
                → L4 scrapling → L5 stealth_browser → L6 llm_browser
                (auto degrade chain; winning engine persisted per source)
                     │
Analysis        builtin 7-category keyword classifier (myssia-classifier)
                + optional LLM enrich (value / relevance / credibility, 0–10)
                     │
Storage         SQLite single file · retention + VACUUM · change baselines
                     │
Push            Feishu card · Telegram · webhook · stdout, threshold-routed
                (immediate / digest AM-PM / archive) + feedback loop

Documentation

Bilingual docs ship in-repo, kept consistent with the code by tests (tests/test_docs.py) — every example YAML loads through the real schema entry point, and zh/en trees cannot drift apart:

	English	中文
Getting started	docs/en/getting-started.md	docs/zh/getting-started.md
Write a plugin	docs/en/write-a-plugin.md	docs/zh/write-a-plugin.md
Schema reference	docs/en/schema.md	docs/zh/schema.md
FAQ (ethics & boundaries)	docs/en/faq.md	docs/zh/faq.md
Scheduled jobs (cron)	docs/en/cron.md	docs/zh/cron.md

Agent-facing condensed reference: skill/SKILL.md.

Roadmap

The core pipeline is implemented and tested — 1300+ tests run in CI, none of them touch the real network. The desktop app is ready for daily use (unified data paths, bundled official plugins, an out-of-the-box demo, a signed update channel — shipping in Releases since v0.0.1); in-card feedback buttons, the settings feedback toggle and collection trends have landed on main in the desktop-parity batch and ship with the next release (the feedback loop already works via CLI); the Windows build did not ship in this release (no Windows installer in Releases yet; planned for a follow-up batch).

Milestone	Scope	Status
v0.1 skeleton	Core pipeline, direct_api/static/firecrawl engines, 12-section schema, change fingerprint, classifier, Feishu routing	✅ shipped
v0.2 usable	SQLite store + registry + retention, full CLI (init/test/list/doctor), crawl4ai L3, LLM enrich + budget guardrails, docker compose, Telegram, keychain secrets	✅ shipped
v0.3 ecosystem	Agent Skill, plugin market (local/remote dual-mode), Scrapling L4, feedback loop (CLI + callback receivers)	✅ shipped
v0.4 deep water	stealth_browser L5, llm_browser L6, trend baselines, event aggregation	✅ shipped
v1.0 launch	Bilingual docs, demo assets, GitHub facade, public delivery	✅ shipped
v1.1 desktop-first	Tauri desktop shell (Python core as sidecar), five-screen UI, in-process plugin tier; in-card feedback buttons / settings feedback toggle / collection trends have landed in the desktop-parity batch and ship with the next release (feedback works via CLI today)	✅ shipped
v1.1.1 data paths	Desktop data-path unification (MYIA_HOME / bundled official plugins / first-run seed), out-of-the-box demo plugin, signed update channel (check + install)	✅ shipped
Web UI	browser front-end on the same core	📋 planned

Ethics & boundaries

世事 is polite by default: robots.txt respected, rate-limited fetching, credentials never in plaintext. Sources that demand human verification (CAPTCHA, phone numbers) fail with a structured error — 世事 does not attempt to bypass them. Full statement in the FAQ · 中文 FAQ.

Community

• Bug reports & feature requests: issue templates
• Contributing: CONTRIBUTING.md
• Security policy & credential-handling design: SECURITY.md

Acknowledgments

世事's own code is original (MIT), but it stands on giants — consumed as dependencies, plugin backends and design references:

• crawl4ai — L3 fetch engine (optional dependency)
• Scrapling — L4 adaptive anti-bot engine (optional dependency)
• Firecrawl — L3 cloud/self-hosted rendering backend (optional dependency, called as an API)
• Skyvern — L6 LLM-browser fallback (optional dependency)
• changedetection.io — source-management & diff UX reference; myssia-monitor plugin backend
• RSSHub — the "everything is a feed" philosophy
• jhao104/proxy_pool — myssia-proxy plugin backend
• Photon — myssia-osint plugin backend (vendored via git submodule)
• Douyin_TikTok_Download_API — myssia-douyin plugin backend
• Maxun — myssia-maxun plugin backend
• Tauri — desktop shell (Python core embedded as a sidecar)

No upstream source is copied into this repository except clearly-marked git submodules; dependency policy in CONTRIBUTING.md.

License

MIT © 2026 xinzhuzi