boss-zhipin-scraper
Guides users to provide resumes and job preferences, utilizing a dedicated BOSS-Agent Chrome Profile with macOS Computer Use to search for jobs, read full JDs, deduplicate, calibrate market requirements, score, and generate A/B/C outreach tiers on visible official BOSS Zhipin pages. After calibration, it prompts for the target greeting count (default recommended: 20) and continuously searches and reaches out to qualified positions only after user confirmation of specific batches, tiers, counts, and greeting modes. Designed for personal job-seeking tasks such as "finding jobs", "searching positions", "reading JDs", "analyzing BOSS positions", "filtering positions", and "continuous greeting outreach". CLI, CDP, Playwright, hidden APIs, Cookie extraction, and anti-detection workarounds are strictly prohibited.
1How It Works & Mechanism
Implements the job-seeking workflow as a state machine with confirmation gates across 10 sequential, unskippable states.
不得跳过当前状态的必需输入或确认。用户已提供的信息直接继承，不重复询问。
Sets mandatory boundaries: Operates exclusively within the dedicated Chrome Profile titled BOSS-Agent, explicitly prohibiting CLI/CDP/Playwright/Cookie extraction/anti-detection approaches.
- 禁止 BOSS CLI、CDP、远程调试、Bridge、Patchright、Playwright、无头浏览器、隐藏 API、`/wapi`、XHR/fetch 注入、Cookie 或 stoken 提取、代理轮换、指纹伪装和验证码绕过。
Prompts for resume upload during initial setup, extracting only job-matching information while suppressing irrelevant privacy data from echoes and reports.
只提取岗位匹配信息；姓名、手机号、邮箱等无关隐私不得主动回显或写入报告。
Completes job preferences across two rounds, with a maximum of 3 search-altering questions per round.
每轮最多问 3 个会改变搜索结果的问题：
Search confirmation gate authorizes read-only calibration only; operating BOSS pages without confirmation is prohibited.
只有用户确认搜索卡或明确确认同等内容后，才能操作 BOSS 页面。搜索确认仅授权只读校准。
Calibration phase defaults to reading 15–30 job cards and selecting approximately 10 full JDs, strictly requiring 0 external actions.
- 默认读取 15–30 张职位卡片，并从中选择约 10 条代表性完整 JD。
Dynamic A/B/C tiers must simultaneously pass hard criteria, total score threshold, core dimension threshold, and risk vetoes.
档位不能只看总分；必须同时通过：
Outreach confirmation gate requires verifying batch, tier, target greeting count, and greeting mode, defaulting to 20 when unspecified.
本批目标招呼数；未指定时主动询问，默认建议 `20` 个，不得静默代填；
Execution loop runs strictly sequentially per job: read list → deduplicate → read JD → filter → score → risk check → outreach → write checkpoint.
读列表 → 查历史台账 → 读完整 JD → 硬条件过滤 … → 触达或归档 → 写检查点
Hard stop on risk control: Immediately saves checkpoint and halts batch upon CAPTCHA, 429, session disconnect, or prompts to copy Cookies.
- 验证码、安全验证、`_security_check`、429、code 37、访问频繁、账号风险或环境异常；
2Core Capabilities
3External Dependencies
| Kind | Dependency / Endpoint |
|---|---|
| network | BOSS直聘（zhipin.com） |
| cli | macOS Computer Use + Chrome（专用 BOSS-Agent Profile） |
4Recommended Governance & Evidence
- 网络端点: External Network Access Declaration: Accesses public pages of BOSS Zhipin (zhipin.com) via host browser to read job listings and full JDs; submits in-platform greetings upon user authorization
- 本地工作目录: File I/O and Artifact Output: ~/.boss-zhipin-scraper/ledger/job_ledger.jsonl (job ledger)
5Second-Pass Verification
- [ok] 状态机状态数与顺序 — SKILL.md 的 ONBOARD→PROFILE→…→REVIEW 与正文 10 个小节一一对应。
- [ok] 校准只读约束 — SKILL.md 与 calibration-tiering.md 均写 external_actions=0 / 外部动作数必须为 0。
- [ok] 档位阈值一致性 — 主 SKILL.md 的 A≥85/B≥80/C≥72 与 references 表格一致，且要求核心维度最低值。
- [ok] 验收案例数量 — ledger-acceptance.md『离线验收案例』逐条编号 1–22，共 22 条。
- [ok] 禁止路线可定位 — §3 扫描『危險旗標/防護繞過』无命中，正文确有禁止 CDP/Playwright/隐藏 API 等条款。
- [ok] 凭证读取 — 全文无 Cookie/stoken 读取指令，反而明确『不得读取、代填或保存验证码和认证信息』。
- [ok] 档位复核 — 只经宿主浏览器操作公开站点 + 本地台账写入，无凭证读取、无防护绕过 → yellow。
6Verdict & Scenarios
e68954aed6706a48…