@hadoku/jobplatform

中文文档
3.42.100发布于 2026/11/17MITTypeScriptES Module构建来源已验证近 30 天下载 2万

Job listings aggregation and tracking dashboard

原始使用文档
文档目录7 节
  1. The problem it solves
  2. Scoring
  3. Architecture
  4. How it fits together
  5. Getting access
  6. Development
  7. License

hadoku-jobplatform

An end-to-end job-search pipeline. It pulls postings off company ATS boards, scores them against role profiles you define, lets you triage what survives, and assembles a full application packet — tailored resume, cover letter, intro email, screening answers — for the ones worth applying to.

Live at hadoku.me/jobplatform (~7,600 postings in the corpus as of August 2026). The job list is public; profiles, triage state, and application packets need an account.

This is one app in the hadoku.me ecosystem — a personal platform where each app ships as an independently versioned npm package and gets mounted by a parent site. Nothing here runs standalone; see How it fits together.


The problem it solves

Job boards rank by recency and their own incentives. Searching them means re-reading the same 200 postings weekly to find the four that changed. This inverts that: describe the role you want once, and every scrape run re-ranks the whole corpus against it.

Three things make that non-trivial, and most of the interesting code is in the answers:

No ATS publishes seniority. Not Greenhouse, not Lever, not Ashby. "Staff Engineer" vs "Engineering Manager" exists only in the title string. So role is inferred at ingest (worker/src/roleClassify.ts) along two orthogonal axes rather than one list:

  • track — ic or manager. Answers "does this job have direct reports?"
  • level — the rung on that track's ladder.

They're orthogonal because a flat list makes senior and manager look like alternatives to each other, which is nonsense — they're answers to different questions. Classification reads the head of the title, before the first comma, so "Software Engineer, Ads Manager" correctly lands as an IC. Only the genuinely ambiguous lead family falls back to probing the description body.

Filtering and ranking are different requests. "Only show me IC roles" is not "rank IC roles higher." Company, track, and salary floor are hard SQL filters; everything else is a weighted score. Conflating the two produces a feed that's technically sorted and practically useless.

Salary is mostly absent. salary_min is NULL on most postings, so scoring it returned a neutral ~0.5 for nearly everything — pure noise dressed as signal. It was removed as a score factor and demoted to a view filter. Postings with no listed salary survive the filter, because hiding them would empty the feed rather than narrow it.

Scoring

Runs on read, per request — so a score always reflects the profile as it is now rather than as it was when some batch job last ran. (The precomputed match table was dropped in migration 0008 for exactly this reason.)

SignalWeightWhat it measures
keyword_match0.40profile keywords against the description
title_match0.30profile keywords against the title
level_match0.15distance on the seniority ladder
remote_match0.15remote / hybrid / onsite against preference

level_match grades by distance, not equality: exact rung 1.0, one rung off 0.7, further or off-track 0.2 — and an unclassified job scores 0.5, because no signal is not the same as a bad match. Full algorithm and data model: ARCHITECTURE.md.


Architecture

Two independently published packages live in this repo:

PackagePathWhat it is
@wolffm/jobplatformsrc/React micro-frontend, exports mount/unmount
@wolffm/jobplatform-workerworker/Cloudflare Worker API (Hono + D1)

The pipeline, end to end:

  Profile: keywords, track, levels, remote pref, companies
    │
    ▼
  GET /directives ──────────► hadoku-scrape  (Python, separate repo)
    the union of every            │  resolves company → (ats, slug)
    profile's companies +         │  enumerates Greenhouse / Lever / LinkedIn
    keywords. The scraper         │  archives raw postings to Cloudflare KV
    PULLS this each run.          │
                                  ▼
                          POST /ingest   (full postings inline, batches of 25)
                                  │
                                  ▼
                      ┌───────────────────────────┐
                      │  jobplatform-worker       │
                      │  • dedup by URL           │
                      │  • classify track + level │
                      │  • parse (ats, slug)      │
                      │  • store in D1            │
                      └───────────┬───────────────┘
                                  │  score on read, per request
                                  ▼
                      ┌───────────────────────────┐
                      │  jobplatform (React)      │
                      │  • ranked feed + filters  │
                      │  • triage: interested /   │
                      │    saved / applied / …    │
                      │  • job detail + breakdown │
                      └───────────┬───────────────┘
                                  │  service binding
                                  ▼
                            resume-api  →  resume, cover letter, intro email,
                                           screening answers → one shareable packet

Companies are a directive the scraper pulls, not targets this app pushes. The worker publishes what it wants scraped at GET /directives; the scraper reads that on its next run and registers targets itself. That keeps registry ownership in one place — the scraper owns how to scrape, this app owns what is worth scraping.

Full API table and design rationale: CLAUDE.md.


How it fits together

hadoku_site  (parent)
  ├── mounts @wolffm/jobplatform          at  hadoku.me/jobplatform
  └── hosts  @wolffm/jobplatform-worker   at  hadoku.me/jobplatform/api
                                                │
                            ┌───────────────────┼───────────────────┐
                            ▼                   ▼                   ▼
                       hadoku-scrape        resume-bot         Cloudflare D1
                       (postings)        (resume/cover)      (scored corpus)

Push to main publishes both packages to GitHub Packages and fires a packages_updated dispatch; the parent site pulls the new versions and redeploys.

Auth is tier-ranked (public < friend < service < admin), gated in-worker. The parent's edge router resolves a session cookie into a user key and stamps the tier; the worker trusts that stamp only when its provenance header verifies, so a request that bypasses the edge degrades to public rather than being trusted.


Getting access

A fresh clone installs, builds and passes every check with no credential and no .npmrc: the @wolffm/* dev dependencies are npm aliases of their public npmjs copies (npm:@hadoku/<name>). The dev server and the signed-in e2e suite talk to the real hadoku.me API, so those need your own credential. Getting one is the same for every hadoku repo, so the steps live on one page: hadoku.me/docs/contributing/#jobplatform. The identity to ask for is jobplatform-service-key. Once you have it, pnpm dev and pnpm test:e2e work as they are: they run through dev-vault (a dev dependency).


Development

Requires Node 22+ and pnpm. Installing and every check need nothing else; the dev server needs a credential (see Getting access).

pnpm install

pnpm dev          # dev server on :5173
pnpm build        # build both packages
pnpm lint         # eslint
pnpm lint:css     # stylelint + theme-token check
pnpm typecheck

cd worker && pnpm test   # node:test, no framework

To act as an authenticated user in dev, pass ?apiKey=<key> once — the harness in index.html exchanges it for a session id through the same code path production uses. The URL parameter is dev-only and never accepted in production.

Colors come from @wolffm/themes as semantic CSS custom properties (var(--color-primary), var(--color-on-primary-bg), …). Light and dark are automatic; never branch on theme mode, and never hardcode a hex value. pnpm lint:css fails the build on a reference to a token the theme doesn't define — worth running, since an undefined token renders as nothing rather than erroring.

Logging goes through the logger from @wolffm/task-ui-components, not console.log:

import { logger } from '@wolffm/task-ui-components'
logger.info('Message', { key: 'value' })

License

MIT.

版本与发布记录

@hadoku/jobplatform 共发布 109 个版本,首次发布于 2026/10/31;过去 12 个月发布了 109 个版本。最新版本附带 npm provenance 构建来源证明,可追溯到公开的 CI 构建记录。

版本发布日期距今
3.42.100latest2026/11/17今天
3.42.992026/11/161 天前
3.42.982026/11/161 天前
3.42.972026/11/161 天前
3.42.962026/11/161 天前
3.42.952026/11/151 天前

关于 @hadoku/jobplatform 的常见问题

@hadoku/jobplatform 是什么?

Job listings aggregation and tracking dashboard

如何安装 @hadoku/jobplatform?

在项目目录运行 npm install @hadoku/jobplatform,或使用 pnpm add / yarn add / bun add。它还需要宿主项目提供 peer 依赖:react(^19.0.0)、react-dom(^19.0.0)、@wolffm/logger(>=1.3.2)、@wolffm/themes(>=5.7.3)、@wolffm/task-ui-components(>=4.7.3)。npm 7 及以上与 pnpm 8 及以上默认会自动安装必需的 peer 依赖,Yarn 需要手动安装。

@hadoku/jobplatform 支持 TypeScript 吗?

支持。包内自带 TypeScript 类型声明(package.json 声明了 types 入口),安装后即可获得类型提示,无需再安装 @types 包。

@hadoku/jobplatform 支持 ESM 还是 CommonJS?

以 ES Module 形式发布,推荐使用 import 引入。在 CommonJS 项目中,Node.js 20.19+ / 22.12+ 可以直接 require() 不含顶层 await 的 ES Module,更早的版本需要改用动态 import()。

@hadoku/jobplatform 可以商用吗?

@hadoku/jobplatform 使用 MIT 许可。宽松许可:允许商业使用、修改和再分发,通常只需在分发时保留原版权与许可声明。以上为概括,不构成法律意见,请以许可原文为准。

@hadoku/jobplatform 的下载量有多少?

根据 npm 官方下载统计,@hadoku/jobplatform 在 2026/10/17 – 2026/11/15 这 30 天内被下载 19,871 次,日均约 662 次。下载量会把 CI 构建、镜像同步和作为其他包依赖的安装一并计入,适合横向比较热度,不能直接等同于使用人数。

@hadoku/jobplatform 还在维护吗?

最新版本 3.42.100 发布于 2026/11/17(今天);至今共发布 109 个版本,其中过去 12 个月发布了 109 个;首次发布于 2026/10/31。发布频率只是参考信号:功能稳定的工具可能很少需要更新,建议结合源码仓库的 issue 和提交记录判断。