HK Job Radar · 香港岗位雷达

一个本地运行、AI 辅助的香港求职自动化工具。完全免费，没有付费版本。 自动监控三个招聘网站 + 各大企业官网直招的新岗位 → 关键词初筛 → 大模型打分 → 每日推送（企业微信 / WxPusher / 邮件 / Telegram）→ 半自动填表。 数据不出本机，合规克制，配置驱动。

Your personal, AI-assisted job search automation for Hong Kong. Free, with no paid tier. Watch new jobs across three job boards and company career sites → filter → score → daily push (WeCom / WxPusher / Email / Telegram) → semi-auto apply. Local-first, privacy-first, compliance-first.

🍎 On a Mac? One-click install via .command scripts — see Mac 安装教程 → (Chinese, click-by-click).

🧰 Never used a terminal? / 完全新手？ Follow the step-by-step beginner's guide (Chinese, from installing Python to receiving your first digest): 保姆级新手教程 → No coding experience needed — the one-click installer does the work for you.

🌐 English version: jump to English ↓

香港岗位雷达是一个本地运行的求职助手：定时监控 JobsDB / LinkedIn / Indeed 三个招聘网站，外加企业官网直招（走 Workday 招聘系统的机构，内置宏利、永明、 中信银行国际等五家，也能自己加），按你的关键词初筛， 用大模型（DeepSeek）对照你的个人画像打分， 每天推送到你自选的一个通道——企业微信群机器人 / WxPusher / 邮件（HTML 岗位卡片）/ Telegram（带 ✅/❌ 反馈按钮）；还附带一个半自动投递副驾， 帮你填申请表——但最后的「提交」永远由你亲手点击。

它不是求职网站、不是猎头、不是代写简历，它只自动处理求职里最重复的 那部分：找岗位 + 筛选岗位。

• 任何在香港求职的人——学生、应届生、转行者都能用；用 career_stage 选身份， 岗位类型过滤会自动跟着变；
• 尤其适合持学生签证的港硕/非本地学生（默认 career_stage: student，内置签证合规过滤，见下）；
• 岗位方向完全由你自己定——双击 setup.bat（Mac 上是 setup.command），在打开的网页里填； 向导会先预填几个常见方向当示例，改成你自己的就行，覆盖任意岗位。

flowchart LR
    A[JobsDB] --> D[抓取]
    B[LinkedIn] --> D
    C[Indeed] --> D
    W[企业官网 Workday] --> D
    D --> E[关键词初筛]
    E --> F[大模型打分]
    F --> G[SQLite 去重]
    G --> H[日报推送]
    H --> I[半自动填表]

1. 抓取 — 每个开启的站都按随机的人为延迟抓，并限制单站每日页数（防封号优先）。 企业官网走的是各家招聘系统自己的公开接口（Workday），比翻网页稳，也不用过人机验证。
2. 初筛 — 先用关键词和同义词做便宜的匹配，把噪音丢掉，这样只有值得看的岗位才会 送进大模型，省的是你自己的 API 钱。
3. 打分 — DeepSeek 对着你的个人画像给每个岗位打 0–100 分，顺带读一遍岗位写明的 准入条件（须为香港永久居民、要求粤语之类），作为提示标在卡片上——只提示、不扣分， 能不能投由你自己判断；调用失败会降级处理，不会把整条流水线卡死。
4. 去重入库 — SQLite 用「岗位 ID + 归一化 URL」双键去重，保证同一个岗位不会被 推第二遍。
5. 推送 — 按分数排序的日报（电脑开着就 09:00 发，当天第一次开机也会补发）； 新出现的 ≥80 分岗位会立刻单独推一条，不用等到第二天。
6. 投递（可选） — 帮你填表、传简历、起草筛选问题的答案，然后请你逐项核对、 自己点「提交」。

• 🔎 四个来源，各自可开关——JobsDB / LinkedIn / Indeed 三个招聘站，加上企业官网（Workday）
• 🧠 LLM 打分（DeepSeek，OpenAI 兼容，想换别家也行）：按你的画像逐条打分，附一句理由
• 📬 四个推送通道，爱开几个开几个——企业微信 / WxPusher / 邮件（HTML 卡片）/ Telegram， 同时开也行（Telegram 看推送、邮件留个档）。其中 Telegram 多一层反馈：日报里每个岗位都带 ✅ / ❌ 按钮，点一下就把「感兴趣 / 不感兴趣」记进本机数据库，岗位状态跟着变成「感兴趣」 或「已排除」——一条条攒下来，就是你自己看得懂的求职偏好档案
• ⚡ 高分岗位抓到就推；日报想发几份也由你定，早晚各一份， 或者只在真有新岗位落地时才响。香港的岗位经常上午挂出来、下午就约面， 这个差别直接体现在你投得早不早
• ♻️ 同一个岗位在几个站各挂一份，只当一份看——公司名与职位名归一化后比对， 日报里两条都留但标出「同一岗位另见」，实时推送只响一次，不会同一个职位把你叫起来两回
• 🗂️ 积压的岗位摊在一张页面上挑，而不是在聊天记录里往上翻——工作台按分数排开， 可以按来源 / 分数 / 新鲜度筛，看过的、标过的、不想看的各自归位
• 📈 市场在怎么变、你投到了哪一步，两件事放在一起看——趋势看板盯着你关注的方向： 这周放出多少岗位、集中在哪些薪资区间、哪几家在连续招。投递那边不用来回翻记录： 一张本地看板把岗位摊成卡片，点标题直接跳回原站去投；投完在卡片上点「已投」，约到了点 「约面」，告一段落点「结束」——公司、岗位、日期、走的哪个站自动留档， 投出去多少、有多少回音，图上直接看
• ✍️ 投递副驾——申请页上那些「你有几年经验」「期望月薪多少」的选择题，按你的口径替你 选好，停在最后一页等你核对；提交那一下永远是你点
• 🔁 已投记录同步——你已经投过的岗位，从招聘网站读回来一次性标好，不用回卡片一张张点
• 🎯 投之前知道自己差在哪，投之后知道会被问什么——简历和 JD 逐项对照，指出它要、 而你没写出来的那句该补在哪；面试预测按这条 JD 把可能的问题先过一遍
• 📄 最后一份能直接发出去的东西——精美 HTML 报告，也能导出 Excel / PDF， 给导师、给朋友，或者自己留档
• 🎓 按求职阶段自动过滤岗位类型（career_stage：学生 / 已毕业 / 两者）
• 🛡️ 防封号：随机延迟、每日页数上限、指数退避，不破解验证码
• 🧩 配置驱动，可复用到任何求职目标
• 🔓 不需要服务器，数据不出本机——整套东西都在你自己电脑上跑

真实录屏（本地库里 394 条岗位）：筛出 80 分以上 → 展开岗位描述 → 标记已投 → 撤销。

• 有一定技术门槛——不过「一键安装脚本」已省去大部分；仍需自备 DeepSeek key 和配好一个推送通道 （Windows 双击 install.bat，macOS 双击 install.command，流程一样）
• 依赖第三方站点，抓取可能被限流或临时封锁。LinkedIn 需要真实账号并登录一次； JobsDB、Indeed 和企业官网不需要账号，只要过一次人机验证（企业官网连这一步都没有）
• 不保证 100% 覆盖——它是「雷达」，不是「全量抓取」
• 关键词要靠你自己想——它只会按你写的词去找，写得太窄就会漏。 好消息是没有数量限制，把同义词、缩写、中英文写法都铺上去就行。

程序只连下面这几个地方，除此之外不往任何地方发数据：

地区由网址决定，不由 IP 决定。 三个招聘站抓的都是 hk. 子域、或用 location: Hong Kong 参数定位；企业官网那边按城市筛，只要 Hong Kong 的。 所以不论你人在香港、英国、加拿大还是内地，抓回来的都是香港的岗位。

LinkedIn 的异地登录：LinkedIn 对「会话从一个地区突然跳到另一个地区」很敏感， 可能要求重新验证甚至临时限制账号。所以在哪个地区用，就在那个地区登一次 （双击 login_linkedin.bat / .command）。 JobsDB 和 Indeed 不需要账号，换地区完全没影响——省事的话，关掉 LinkedIn 只跑这两个。

方式 A —— 一键安装，全程不用敲命令（Windows，推荐）

1. 安装 Python——务必勾选 "Add python.exe to PATH"
2. 在本页面点 Code → Download ZIP，下载后解压
3. 双击 install.bat，然后跟着窗口走就行——它会一口气做完这些事：
   • 装好依赖和浏览器内核，并引导你填 DeepSeek key；
   • 自动打开配置页面：你在这里填要找的岗位、选一个推送通道、 填你自己的背景，点「保存配置」→「关闭向导」；
   • 自动做 15 项自检，全绿才往下走，有 ❌ 就直接写明该改哪一行；
   • 自动打开浏览器，让你过一次 JobsDB 的安全验证。
4. 看到 全部就绪。 之后，双击 start.bat（保持窗口开着）； 只想跑一轮就双击 run_once.bat。 当天第一次运行就会收到一份日报，之后新出现的达标岗位会推送。 start.bat 关掉之后就不会再自动跑了——Windows 上「每天 09:00」 靠的是这个窗口一直在（Mac 上不一样，见方式 A2）。

以后再双击一次 install.bat 只会检查环境，不会再弹配置页， 收尾显示「环境已就绪，你的配置保持原样」。这是正常的，不是出错。 想改配置就双击 setup.bat。

整个流程里只有一件事机器替不了你：告诉它你想找什么岗位、你是谁。 这两处就在第 3 步自动打开的配置页面里。跳过它们，程序照样会跑，但会抓错岗位、 再拿一份空画像去打分。 每一步点哪里，见保姆级新手教程。

拿不准自己配对了没有？双击 doctor.bat。 它会把 15 项逐个查一遍， 缺什么就直说该改哪一行——不用猜，也不用等我回复。 装完之后想改关键词或推送通道，双击 setup.bat 再走一遍那个页面即可。

方式 A2 —— Mac 一键安装，同样不用敲命令

流程完全一样，只是文件后缀从 .bat 换成 .command：

1. 安装 Mac 版 Python 3.10+ （Mac 自带的那个太旧，不够用）

   系统要求：macOS 11（Big Sur）或更新——python.org 从 Python 3.10 起只提供 macOS 11 及以上的安装包；10.15 及更早的系统装不上官方 Python，不建议硬来。 查版本：屏幕左上角 Apple 菜单 →「关于本机」。

2. Code → Download ZIP，解压后把文件夹挪到 /Users/你的名字/ 下 （别放在桌面/文稿/下载里运行，那三个位置被系统保护，浏览器可能报 Operation not permitted）
3. 按住 Control 点 install.command →「打开」→ 再点「打开」 （只有第一次需要这样，因为文件是从网上下载的；找不到「打开」按钮就看 项目里的 MAC-打不开就读我.txt） —— 之后它和 Windows 一样，把配置页面、15 项自检、 JobsDB 安全验证 一口气带完，中途不用你再判断该点哪个文件
4. 看到 全部就绪。 之后，双击 run_once.command（只跑一轮）或 start.command（一直挂着）

install.command 会顺手注册每天 09:00 自动运行（电脑当时在睡觉也会在 唤醒后补跑），不用一直开着窗口；不想要就双击 关闭自动运行.command。

Windows 上没有这层定时任务，那边靠 start.bat 一直开着。 不想整天开着电脑的话，用 run_once.bat 手动跑一轮，或者开「任务计划程序」 自己加一条定时任务指向 run_once.bat。

每一步点哪里、每个窗口长什么样，见 Mac 安装教程。

方式 B —— 手动安装（终端，只需一次）

git clone https://github.com/prui0474/hk-job-radar.git
cd hk-job-radar
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
playwright install chromium

Copy-Item config.example.yaml config.yaml
Copy-Item data/profiles/profile.example.yaml data/profiles/profile.yaml
$env:DEEPSEEK_API_KEY = "sk-..."   # 填自己的 key
# 编辑 config.yaml：设 notify.channels（wecom / wxpusher / email / telegram 任选其一）并填好凭据

python run.py login jobsdb        # 首次过盾/登录（JobsDB / LinkedIn / Indeed 各一次，企业官网不用）
python run.py login linkedin
python run.py login indeed

Windows 中文编码提示（别删那一行）。 requirements.txt 的第一行 # -*- coding: utf-8 -*- 是刻意保留的：简体中文 Windows 下 pip 默认会用 GBK 解码该文件，从而崩在 UnicodeDecodeError: 'gbk' codec can't decode byte ...。 项目里的 11 个 .bat 也统一设了 PYTHONUTF8=1，保证窗口输出被重定向到文件时 中文与 emoji 不会抛 UnicodeEncodeError。

日常使用 —— 无论哪种方式安装，常用的就下面这几条：

python run.py once                # 只抓一轮就退出并推日报（或直接双击 run_once.bat）
python run.py run                 # 启动监控（或直接双击 start.bat）
python run.py report now          # 立刻补发一次日报
python run.py notify test         # 给推送通道发一条测试消息（配完通道后先点这个，10 秒验证）
python run.py doctor              # 一键自检：15 项体检，逐条告诉你哪里没配好、怎么修
python run.py stats               # 查看今日抓取/推送统计
python run.py apply <job_uid>     # 投递副驾（半自动填表）

不想整天开着电脑？ 用 once / run_once.bat：抓一轮 → 打分 → 推日报 → 自动退出， 开机十几分钟就够。日报若在开机时还没发过，也会自动补发。 详见 docs/uptime-strategy.md。

所有设置在 config.yaml 里（config.example.yaml 是带逐行注释的模板）。 不想碰文本文件就双击 setup.bat（Mac 上是 setup.command），在网页里点。

data/profiles/profile.yaml 的 background: 段是 AI 打分的依据，必须换成你自己的 技能、经历与目标岗位。若该段缺失、或仍是一堆 <占位符>，程序会退回一段通用描述来打分， 分数参考价值很低（但不会报错——这是最容易被忽略的一步）。

• 不自动提交投递——代码里没有任何自动提交路径，最后一下永远由你点；
• 不破解验证码、不绕过 Cloudflare；
• 按求职阶段过滤——默认 career_stage: student 只抓实习/兼职（学生签证合规）； 拿到 IANG 后改成 graduate 或 both 即可。

如果你对这个项目感兴趣，点个 star 或者随缘打赏 就是最直接的鼓励。

如果你用它拿到了面试、或者只是觉得省了时间，可以请我喝杯咖啡。万分感谢你的支持，祝你早日找到高薪的心仪offer！

有问题、想反馈，或者想一起改这个项目：邮箱 P_RUI@outlook.com （标题里带一句「岗位雷达」，我一眼就能认出来）。

Questions, feedback, or want to help improve it: email P_RUI@outlook.com (put "HK Job Radar" in the subject line).

MIT

HK Job Radar is a self-hosted job search assistant for the Hong Kong market. It periodically watches JobsDB / LinkedIn / Indeed as well as company career sites running on Workday (Manulife, Sun Life, CNCBI and more built in), filters them by your keywords, scores each one with an LLM against your personal profile, and delivers a daily digest to whichever channel you pick — WeCom group bot, WxPusher, email (HTML job cards), or Telegram (with ✅/❌ buttons). It also ships a semi-automatic application assistant that fills in forms for you — the final "Submit" is always yours to click.

It is not a job board, not a recruiter, and not a resume writer. It automates the repetitive part of job hunting: finding and shortlisting.

• Anyone job-hunting in Hong Kong — students, fresh graduates, career switchers. Set career_stage: student / graduate / both and the employment-type filter follows.
• Especially non-local students on a Hong Kong student visa — career_stage: student ships as the default and applies visa-compliance filtering (see Compliance).
• Works for any role — list your own target job titles on the setup page (a local browser form). Nothing is hard-coded to one field; the wizard just prefills a few common examples for you to overwrite.

flowchart LR
    A[JobsDB] --> D[Scraper]
    B[LinkedIn] --> D
    C[Indeed] --> D
    W[Company career sites] --> D
    D --> E[Keyword Filter]
    E --> F[LLM Scoring]
    F --> G[SQLite Dedup]
    G --> H[Daily Digest Push]
    H --> I[Semi-auto Apply]

1. Scrape — each enabled source is fetched with random human-like delays and a per-site daily page cap (anti-ban first). Company career sites are read through their own public job API (Workday): steadier than page scraping, and no human verification involved.
2. Filter — cheap keyword + synonym matching drops noise so the LLM only sees worthwhile roles.
3. Score — DeepSeek scores each role 0–100 against your profile and also flags any eligibility conditions the ad states outright (permanent-resident only, Cantonese required, …) as a note on the card — a note only, never a deduction; whether to apply stays your call. Failures degrade gracefully instead of blocking the pipeline.
4. Dedup + store — SQLite double-key (job ID + normalized URL) guarantees no duplicate pushes.
5. Notify — a digest sorted by score (09:00 if the PC is on, otherwise the first time you power on that day). Roles at ≥ 80 are also pushed the moment they are found.
6. Apply (optional) — fills forms + uploads resume + drafts screening answers, then asks you to review every field and click submit yourself.

• 🔎 4 sources, each toggleable — JobsDB / LinkedIn / Indeed plus company career sites (Workday).
• 🧠 LLM scoring — DeepSeek (OpenAI-compatible; swap base_url for others), with a one-line reason attached to every score.
• 📬 4 push channels — turn on as many as you like — WeCom group bot / WxPusher / Email (HTML job cards) / Telegram. All at once is fine (Telegram for the ping, email for the archive). Telegram adds a feedback layer: every job in the digest carries ✅ / ❌ buttons, one tap records interested / not interested in your local database and flips that job's status, so your own taste accumulates as a record you can read back.
• ⚡ High-scoring roles are pushed the moment they are found. Hong Kong roles often go up in the morning and are interviewing by the afternoon — this is the difference between applying early and applying late. Digest cadence is yours too: twice a day, or only when something new actually lands.
• ♻️ One posting listed on several boards counts once — company and title are normalised and compared; the digest keeps both lines but marks "same job, see #N", and instant push fires once instead of waking you up twice for the same role.
• 🗂️ The backlog laid out on one page instead of scrolled up in a chat — a graphical workspace sorted by score, filterable by source / score / freshness.
• 📈 How the market moves and how far you got, side by side — the trend board watches your directions: how many roles came out this week, which salary bands they cluster in, which employers keep posting. The application side is just clicks: a local board lays the roles out as cards, and applied / interview / closed is one click on the card. One glance shows what went out and what came back.
• ✍️ Application co-pilot — picks the answers to those screening questions (years of experience, expected salary, right to work), stops on the last page for you to check. The final click is always yours.
• 🔁 Applied-jobs sync — reads back the applications you already sent on JobsDB and marks them in one go, instead of clicking each card.
• 🎯 Know the gap before you apply, know the questions after — resume-vs-JD comparison points at the line the JD wants and your resume never shows; interview prep rehearses the questions that JD invites.
• 📄 And one thing you can actually send — a polished HTML report, exportable to Excel / PDF, for a mentor, a friend, or your own records.
• 🎓 Career-stage aware — career_stage: student / graduate / both sets the employment types automatically (student visa → internship/part-time only).
• 🛡️ Anti-ban — random delays, daily page caps, exponential backoff, no CAPTCHA cracking, no Cloudflare bypass.
• 🧩 Config-driven — keywords, location, career stage, profile, thresholds all in YAML.
• 🔓 No server involved, nothing leaves your machine.

Recorded from the real board (394 jobs in the local database): filter to 80+ → expand a job description → mark it applied → undo.

• Some technical comfort helps — though the one-click installer (install.bat on Windows, install.command on macOS) removes most of it; you still need a DeepSeek key and one configured push channel.
• Depends on third-party sites — scraping may be rate-limited or blocked at any time. LinkedIn needs a real account and a one-time login; JobsDB, Indeed and company career sites need no account at all — the first two just need one human-verification pass.
• Not 100% coverage — it's a radar, not a guarantee of every posting.
• Keywords are on you — it only looks for what you write, so a narrow list means missed roles. There is no cap, so pile on synonyms, abbreviations and the Chinese / English variants of the same title.

The tool talks to these hosts and nothing else:

hk.jobsdb.com · hk.indeed.com · www.linkedin.com · api.deepseek.com · your employers' own hiring domains (*.myworkdayjobs.com when that company runs on Workday) — plus one push channel (api.telegram.org / qyapi.weixin.qq.com / wxpusher.zjiecode.com / your SMTP server).

Region is decided by URL, not by your IP. All three boards are queried through their Hong Kong subdomain or a location: Hong Kong parameter, and company career sites are filtered down to Hong Kong — so you get Hong Kong postings wherever you are.

LinkedIn and changing regions: LinkedIn is sensitive to a session jumping between countries and may demand re-verification or temporarily restrict the account. Log in again from wherever you actually are (double-click login_linkedin.bat / .command). JobsDB and Indeed need no account, so they are unaffected — disabling LinkedIn and running those two is the friction-free option.

Option A — One-click install, no terminal needed (Windows, recommended)

1. Install Python — tick "Add python.exe to PATH".
2. On this page click Code → Download ZIP, then unzip it.
3. Double-click install.bat and follow the window. It chains the rest of the setup for you:
   • installs dependencies and the browser engine, and asks for your DeepSeek key;
   • opens the setup page automatically — fill in what you're looking for (target job titles), who you are (the background the AI scores against), and where to send messages (pick one of the four channels and paste its credentials). Click Save, then Close wizard. No YAML editing, no indentation to get wrong;
   • runs the 15-item self-check automatically, and only continues when it is all green — anything marked with a cross tells you exactly what to change;
   • opens a browser for the one-time JobsDB verification.
4. Once it prints 全部就绪。 ("all set"), double-click start.bat and leave it running — or run_once.bat for a single sweep. Either way you get a digest the first time you run it that day, and new high-scoring jobs are pushed as they appear. On Windows the 09:00 schedule lives inside that window: close it and the daily digests stop (macOS is different — see Option A2).

Running install.bat a second time only re-checks the environment: it does not reopen the setup page, and it ends with "环境已就绪，你的配置保持原样" ("environment ready, your settings are untouched"). That is expected, not a failure.

The one thing the installer cannot do for you is describing what you're looking for and who you are — that is the setup page it opens for you in step 3. Skip it and the tool still runs, but it will fetch the wrong roles and score them against an empty profile. Exact click-by-click steps: beginner's guide (Chinese). Want to change keywords or the push channel later? Double-click setup.bat.

Option A2 — One-click install on macOS (no terminal needed)

Same flow, but the files end in .command instead of .bat:

1. Install Python 3.10+ for macOS (the Python that ships with macOS is 3.9 and is too old).
2. Code → Download ZIP, unzip it, and move the folder to /Users/<you>/ (avoid Desktop / Documents / Downloads — macOS guards those and the browser may fail with Operation not permitted).
3. Control-click install.command → Open → Open (first launch only; macOS blocks downloaded scripts — see MAC-打不开就读我.txt if the button is missing). It then walks the same chain as Windows: setup page → self-check → JobsDB verification, without asking you to pick the next file.
4. Once it prints 全部就绪。, double-click run_once.command (single sweep) or start.command (keep running).

install.command also registers a daily 09:00 launchd job; if the Mac is asleep at that moment, it runs on wake. No window needs to stay open. Remove it any time with 关闭自动运行.command.

Windows has no equivalent scheduled task — there the 09:00 digest depends on start.bat staying open. To avoid leaving the PC on all day, use run_once.bat in a single sweep, or add a Task Scheduler entry pointing at it.

Full version, with every screen explained: Mac 安装教程 →

Option B — Manual (terminal, one-time setup)

# 1. Clone & enter
git clone https://github.com/prui0474/hk-job-radar.git
cd hk-job-radar

# 2. Virtual env + deps + Chromium
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
playwright install chromium

# 3. Config + profile
Copy-Item config.example.yaml config.yaml
Copy-Item data/profiles/profile.example.yaml data/profiles/profile.yaml

# 4. LLM key — either an env var, or a one-line api_key.txt next to config.yaml
$env:DEEPSEEK_API_KEY = "sk-..."      # env var (or: create api_key.txt containing just sk-...)

# 5. Fill in the config — this opens a setup page in your browser
python run.py setup
#    (prefer editing YAML by hand? config.example.yaml is fully commented)

# 6. One-time login (Cloudflare / LinkedIn)
python run.py login jobsdb
python run.py login linkedin
python run.py login indeed

Windows encoding note. The first line of requirements.txt (# -*- coding: utf-8 -*-) is there on purpose. On Simplified-Chinese Windows, pip would otherwise decode that file with GBK and abort with UnicodeDecodeError: 'gbk' codec can't decode byte .... Keep the line. The bundled .bat files also set PYTHONUTF8=1, so Chinese/emoji console output stays intact even when the window output is redirected to a file.

Daily use — after either option, just this:

python run.py once             # run ONE sweep, push the digest, then exit (or double-click run_once.bat)
python run.py run              # keep monitoring in the foreground (or double-click start.bat)
python run.py report now       # send today's digest immediately
python run.py notify test      # send a test message to your push channel(s) — verify setup in 10s
python run.py doctor           # self-check: 15 items, says exactly what is still misconfigured
python run.py setup            # open the in-browser setup page (edit the config without YAML)
python run.py stats            # today's fetch/new/push counts
python run.py apply <job_uid>  # semi-auto application assistant

Don't leave your PC on all day? Use once / run_once.bat. It does one full sweep, scores the jobs, pushes the digest, and exits — so a 15-minute session is enough. The digest is also auto-sent when you power on if it hasn't gone out yet. See docs/uptime-strategy.md.

Everything is in config.yaml (see config.example.yaml for commented defaults):

The background section of data/profiles/profile.yaml drives LLM scoring — you must replace the placeholders with your own skills, experience, and target roles. If that section is missing or still holds only <placeholders>, the tool falls back to a generic description and scores become far less meaningful.

中文：data/profiles/profile.yaml 的 background: 段是 AI 打分的依据，必须换成 你自己的背景与目标岗位。若该段缺失、或仍是一堆 <占位符>，程序会退回一段通用 描述来打分，分数参考价值很低（但不会报错——这是最容易被忽略的一步）。

• No auto-submit — the code has no path that submits an application; the final click is always yours.
• No CAPTCHA cracking / Cloudflare bypass — challenges are waited on or handed to you, never solved.
• Student-visa aware — career_stage: student (the default) keeps internship/part-time only; switch to graduate / both only after you hold IANG.

If this project is useful to you, a star — or a coffee if you feel like it — is the most direct way to say so.

If it got you an interview, or just saved you an evening, you can buy me a coffee. Thank you so much for the support — and may you land the well-paid offer you are after!

hk_job_radar/
├── install.bat   / install.command   # one-click install (Windows / macOS)
├── setup.bat     / setup.command     # change settings: opens the setup page
├── doctor.bat    / doctor.command    # 15-item self-check
├── start.bat     / start.command     # keep running (this is what drives 09:00 on Windows)
├── run_once.bat  / run_once.command  # single sweep, then exit
├── login_jobsdb.bat / login_linkedin.bat / login_indeed.bat   # one-time verification / login
├── report_now.bat / notify_test.bat  # resend digest / test your push channel
├── board.bat     / board.command     # local job board: see the backlog, track progress
├── MAC-打不开就读我.txt              # macOS: what to do if they won't open
├── scripts/
│   ├── common.sh                  # shared shell helpers for the .command scripts
│   └── check_encoding.py          # guard against the line-ending / encoding class of bugs
├── .github/workflows/check.yml    # free macOS + Windows CI on every push
├── config.example.yaml            # config template (copy → config.yaml)
├── data/profiles/profile.example.yaml  # profile template (copy → profile.yaml)
├── src/hk_job_radar/
│   ├── scrapers/                  # jobsdb / linkedin / indeed / workday adapters
│   ├── scoring/                   # keyword filter + LLM scorer
│   ├── copilot/                   # application co-pilot (JobsDB wizard + form filler)
│   ├── notify/                    # 4 push channels: base / dispatch / wecom / wxpusher / email / telegram
│   ├── guard.py                   # strips hidden text and prompt-injection from job ads
│   ├── liveness.py                # re-checks whether a posting is still open
│   ├── doctor.py                  # 15-item self-check (double-click doctor.bat / .command)
│   ├── setup_wizard.py            # in-browser setup (double-click setup.bat / .command)
│   ├── setup_ui.html              # the setup page itself (single file, no external assets)
│   ├── channels.py                # the four push channels, one place
│   ├── workbench.py / workbench_ui.html   # local job board (server + single-file UI)
│   ├── dedup.py                   # cross-site duplicate detection
│   ├── applied_sync.py            # read applied jobs back from JobsDB
│   ├── export.py                  # HTML / Excel / PDF export
│   ├── resume.py / matcher.py     # resume parsing + resume-vs-JD comparison
│   ├── pipeline.py / scheduler.py
│   └── config.py / db.py / models.py
├── docs/                          # beginner guides (Windows + macOS)
└── tests/                         # unit tests (pure, no network)

Why the CI runs on macOS: a real macOS runner is the only honest way to verify the Mac path — Windows can be checked locally, macOS cannot. Every push installs the project, downloads Chromium, runs stats, and checks the launchd job — see .github/workflows/check.yml.

MIT