openhire 0.1.0__tar.gz

openhire-0.1.0/.claude/settings.local.json

@@ -0,0 +1,84 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(py --version)",
+      "Bash(python -m pip --version)",
+      "Bash(git --version)",
+      "Bash(psql --version)",
+      "Bash(curl -s -m 20 \"https://api.lever.co/v0/postings/ramp?mode=json\" -w \"\\\\nHTTP %{http_code}\\\\n\" -o /tmp/lever.json)",
+      "Read(//tmp/**)",
+      "Bash(python -c \"import json;d=json.load\\(open\\('/tmp/lever.json'\\)\\);print\\('jobs:',len\\(d\\)\\)\")",
+      "Bash(curl -s -m 20 \"https://api.ashbyhq.com/posting-api/job-board/openai\" -w \"\\\\nHTTP %{http_code}\\\\n\" -o /tmp/ashby.json)",
+      "Bash(python -c \"import json;d=json.load\\(open\\('/tmp/ashby.json'\\)\\);print\\('jobs:',len\\(d.get\\('jobs',[]\\)\\)\\)\")",
+      "Bash(python -m venv .venv)",
+      "Bash(./.venv/Scripts/python.exe -m pip install --quiet --upgrade pip)",
+      "Bash(./.venv/Scripts/python.exe -m pip install --quiet -e \".[dev]\")",
+      "Bash(./.venv/Scripts/python.exe -m pip install -e \".[dev]\")",
+      "Bash(export PYTHONIOENCODING=utf-8)",
+      "Bash(./.venv/Scripts/python.exe -c \"import openhire; from openhire.db import init_db; init_db\\(\\); print\\('DB init OK, v'+openhire.__version__\\)\")",
+      "Bash(rm -f ~/.openhire/openhire.db)",
+      "Bash(./.venv/Scripts/ohp.exe seed *)",
+      "Bash(./.venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/verify_m1.py\")",
+      "Bash(./.venv/Scripts/ohp.exe ingest *)",
+      "Bash(./.venv/Scripts/ohp.exe status *)",
+      "Bash(./.venv/Scripts/python.exe -m pytest -q)",
+      "Bash(./.venv/Scripts/python.exe -)",
+      "Bash(./.venv/Scripts/ohp.exe fix-apply-channels *)",
+      "Bash(./.venv/Scripts/python.exe -c \"import importlib.metadata as m; print\\('mcp', m.version\\('mcp'\\)\\)\")",
+      "Bash(./.venv/Scripts/python.exe -c \"import openhire.mcp_server, openhire.service, openhire.cli; print\\('imports OK'\\)\")",
+      "Bash(./.venv/Scripts/ohp.exe search *)",
+      "Bash(./.venv/Scripts/python.exe -m pytest tests/test_mcp_acceptance.py::test_all_five_acceptance_scripts -vv)",
+      "Bash(unset OPENHIRE_DATABASE_URL)",
+      "Bash(./.venv/Scripts/python.exe -m pytest --co -q)",
+      "Bash(export OPENHIRE_DATABASE_URL=\"sqlite+pysqlite:///C:/Users/gdche/.openhire/openhire.db\")",
+      "Bash(./.venv/Scripts/python.exe -c ' *)",
+      "Bash(xxd)",
+      "Bash(python -c ' *)",
+      "Bash(cd *)",
+      "Bash(./.venv/Scripts/ohp.exe extract-sample *)",
+      "Bash(./.venv/Scripts/python.exe -m pytest tests/test_rebuild.py tests/test_extract.py -q)",
+      "Bash(./.venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/verify_rebuild.py\")",
+      "Bash(./.venv/Scripts/python.exe -c \"import openhire.cli, openhire.client; print\\('import OK'\\)\")",
+      "Bash(./.venv/Scripts/ohp.exe --help)",
+      "Bash(./.venv/Scripts/python.exe -m pytest tests/test_client.py tests/test_cli_journey.py -q)",
+      "Bash(.venv/Scripts/python.exe -c 'from openhire.db import session_scope, Job; from sqlalchemy import select, func; import openhire.config as c; print\\('\\\\''DB URL:'\\\\'', c.DATABASE_URL\\); *)",
+      "Bash(.venv/Scripts/ohp.exe --help)",
+      "Bash(.venv/Scripts/ohp.exe status *)",
+      "Bash(.venv/Scripts/ohp.exe search *)",
+      "Bash(.venv/Scripts/python.exe -m pytest -q)",
+      "Bash(timeout 200 .venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/repro.py\")",
+      "Bash(.venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/repro2.py\")",
+      "Bash(.venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/repro3.py\")",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/test_concurrency.py -v)",
+      "Bash(timeout 120 .venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/repro_t3.py\")",
+      "Bash(.venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/diag.py\")",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/test_posting_dates.py -q)",
+      "Bash(.venv/Scripts/python.exe -c \"import sqlite3; c=sqlite3.connect\\('C:/Users/gdche/.openhire/openhire.db'\\); c.execute\\('PRAGMA wal_checkpoint\\(TRUNCATE\\)'\\); c.close\\(\\); print\\('checkpointed'\\)\")",
+      "Bash(cp \"C:/Users/gdche/.openhire/openhire.db\" \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/pre_backfill_backup.db\")",
+      "Bash(timeout 300 .venv/Scripts/ohp.exe backfill-dates)",
+      "Bash(.venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/regress.py\")",
+      "Bash(.venv/Scripts/python.exe -c \"import sqlite3; c=sqlite3.connect\\('C:/Users/gdche/.openhire/openhire.db'\\); c.execute\\('PRAGMA wal_checkpoint\\(TRUNCATE\\)'\\); c.close\\(\\); print\\('real DB checkpointed'\\)\")",
+      "Bash(.venv/Scripts/python.exe -c ' *)",
+      "Bash(timeout 120 .venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/rf_probe.py\")",
+      "Bash(.venv/Scripts/ohp.exe extract-role-family *)",
+      "Bash(echo \"EXIT $?\")",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/test_appendix_b.py -q)",
+      "Bash(.venv/Scripts/python.exe \"C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/rf_verify.py\")",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/test_snapshot.py -q)",
+      "Bash(.venv/Scripts/python.exe -c \"import sqlite3;c=sqlite3.connect\\('C:/Users/gdche/.openhire/openhire.db'\\);print\\('watches',c.execute\\('SELECT count\\(*\\) FROM watches'\\).fetchone\\(\\)[0],'applications',c.execute\\('SELECT count\\(*\\) FROM applications'\\).fetchone\\(\\)[0],'jobs',c.execute\\('SELECT count\\(*\\) FROM jobs'\\).fetchone\\(\\)[0]\\)\")",
+      "Bash(.venv/Scripts/ohp.exe snapshot-build *)",
+      "Bash(C:/Users/gdche/AppData/Local/Temp/claude/C--openhire/dc7ce18d-7256-41ad-8535-30df941b25d2/scratchpad/freshvenv/Scripts/ohp.exe --help)",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/test_privacy.py tests/test_ranking.py tests/test_snapshot.py -q)",
+      "Bash(.venv/Scripts/python.exe -c \"import tomllib; d=tomllib.load\\(open\\('pyproject.toml','rb'\\)\\); print\\('name:', d['project']['name']\\); print\\('urls:', d['project'].get\\('urls'\\)\\); print\\('scripts:', d['project']['scripts']\\)\")",
+      "Bash(rm -f __TRACKED_VAR__/dist/*.whl __TRACKED_VAR__/dist/*.tar.gz)",
+      "Bash(.venv/Scripts/python.exe -m build --outdir __TRACKED_VAR__/dist)",
+      "Bash(.venv/Scripts/python.exe -m pip install -q twine)",
+      "Bash(git init *)",
+      "Bash(git config *)",
+      "Bash(git add *)",
+      "Bash(grep -iE \"\\\\.env$|\\\\.pypirc|token|\\\\.db$|\\\\.db\\\\.gz$\")",
+      "Bash(winget install *)",
+      "Bash(echo \"EXIT=$?\")"
+    ]
+  }
+}

openhire-0.1.0/.env.example

@@ -0,0 +1,22 @@
+# OpenHire configuration — copy to .env (never commit real secrets).
+
+# --- Database ---
+# Default: local SQLite at ~/.openhire/openhire.db. For production use Postgres:
+# OPENHIRE_DATABASE_URL=postgresql+psycopg://openhire:secret@localhost:5432/openhire
+
+# --- LLM extraction ---
+# Extractor: auto (Anthropic if a key is present, else heuristic) | anthropic | heuristic
+OPENHIRE_EXTRACTOR=auto
+# Key is injected only at extraction time; nothing is ever sent about the user.
+# ANTHROPIC_API_KEY=sk-ant-...
+# OPENHIRE_EXTRACTION_MODEL=claude-haiku-4-5-20251001
+
+# --- Crawl politeness (defaults shown) ---
+OPENHIRE_MIN_TENANT_INTERVAL_MIN=30
+OPENHIRE_MAX_CONCURRENCY=5
+OPENHIRE_HTTP_TIMEOUT=30
+
+# --- Freshness tiers ---
+OPENHIRE_HOT_WINDOW_DAYS=7
+OPENHIRE_HOT_INTERVAL_HOURS=6
+OPENHIRE_COLD_INTERVAL_HOURS=24

openhire-0.1.0/.gitignore

@@ -0,0 +1,42 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+.eggs/
+build/
+dist/
+.venv/
+venv/
+
+# Env / secrets — NEVER commit API keys or PyPI tokens
+.env
+*.env
+!.env.example
+.pypirc
+*.pem
+*.key
+*_token
+*.token
+
+# Local state / data
+*.db
+*.db-wal
+*.db-shm
+*.db.gz
+*.sqlite
+*.sqlite3
+.openhire/
+~/.openhire/
+
+# Test / tooling
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+.coverage
+htmlcov/
+
+# OS / editor
+.DS_Store
+Thumbs.db
+.idea/
+.vscode/

openhire-0.1.0/CLAUDE.md

@@ -0,0 +1,41 @@
+# CLAUDE.md — OpenHire 工作约定（每个会话必读）
+
+## 新会话恢复指引（开工前先做这三步）
+
+1. **先读 `README.md`** — 了解产品定位、五个协议字段、三条隐私红线、安装与工具（公开版 README）。
+2. **再读 `PROGRESS.md`** — 了解已完成到哪一步、关键决策与理由、下一步、待用户确认事项。
+3. **禁止重做已完成的工作。** M1/M2/M3 已验收通过（见 PROGRESS.md 的验收证据）。除非用户明确要求返工，不要重建已完成的里程碑。
+
+## 常设工作制度（持续遵守）
+
+1. **进度记录：** 每完成一个里程碑、或每个工作日结束时，更新 `PROGRESS.md`：
+   - 日期
+   - 已完成（含验收证据：测试数、实测输出、花费等可核对的事实）
+   - 关键决策及理由
+   - 下一步
+   - 待用户确认事项
+   控制在一页以内；新进展追加在顶部（倒序），旧条目保留。
+2. **恢复指引常驻：** 本文件顶部的「新会话恢复指引」始终保留并保持最新。
+3. **每次会话结束前主动提醒用户：** 「今日请备份 `C:\openhire` 到 U 盘。」
+4. **里程碑节奏：** 每完成一个里程碑停下，对照验收标准向用户汇报后再继续。
+5. **汇报归档：** 每次完成任务后的完整汇报，除在终端显示外，**同时写入 `C:\openhire\reports\`**。
+   - 文件名 = 递增编号 + 主题，如 `001-真机验收三件套.md`、`002-M4打包.md`。
+   - Markdown 格式，**自足完整**：不依赖终端上下文，单独打开也能看懂（含背景、做了什么、证据、结论、下一步）。
+   - 每次干完活的**最后一行**告诉用户：「汇报已写入 C:\openhire\reports\xxx.md」。
+   - 编号取 `reports\` 里现有最大编号 +1（补零三位）。
+
+## 三条隐私红线（CI 强制，永不可破）
+
+1. 简历或任何 PII **绝不**经过服务端 —— 只有匿名指纹过网。
+2. 排序**绝不**是付费参数 —— 只是 f(匹配度, 新鲜度) 的纯函数，签名锁死。
+3. 雇主只为已授权、已交付的结果付费 —— 绝不为曝光付费（v0.1 无任何计费代码）。
+
+对应自动化测试见 `tests/test_privacy.py`、`tests/test_ranking.py`。改动排序/服务层/apply 后必须跑 `pytest` 确认全绿。
+
+## 关键路径与事实
+
+- 代码根：`C:\openhire\src\openhire`
+- 数据库（默认，绝对路径）：`C:\Users\gdche\.openhire\openhire.db`（11,825 职位 / 96 公司 / 全量 DeepSeek 抽取）
+- CLI 可执行文件：`C:\openhire\.venv\Scripts\ohp.exe`（**未在系统 PATH 上** —— 接入 Claude Desktop 时须写全路径）
+- 抽取后端：DeepSeek（`deepseek-chat`，key 从 `.env` 的 `DEEPSEEK_API_KEY` 读；serve 阶段不需要 key）
+- 设计交接文档：`design_handoff_openhire_v01\README.md`（唯一权威规格）；`design_refs\*.html` 仅供交互参考，**不复用其代码**。

openhire-0.1.0/PKG-INFO

@@ -0,0 +1,151 @@
+Metadata-Version: 2.4
+Name: openhire
+Version: 0.1.0
+Summary: OpenHire v0.1 「哨兵」— an MCP job protocol server. Résumés never leave your device.
+Project-URL: Homepage, https://github.com/gzchenhao/openhire
+Project-URL: Repository, https://github.com/gzchenhao/openhire
+Project-URL: Issues, https://github.com/gzchenhao/openhire/issues
+Author: OpenHire
+License: MIT
+Keywords: ashby,ats,greenhouse,jobs,lever,mcp,recruiting
+Requires-Python: >=3.11
+Requires-Dist: anthropic>=0.39
+Requires-Dist: apscheduler>=3.10
+Requires-Dist: httpx>=0.27
+Requires-Dist: mcp>=1.2.0
+Requires-Dist: python-dotenv>=1.0
+Requires-Dist: rich>=13.7
+Requires-Dist: sqlalchemy>=2.0
+Requires-Dist: typer>=0.12
+Provides-Extra: dev
+Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Provides-Extra: postgres
+Requires-Dist: psycopg[binary]>=3.1; extra == 'postgres'
+Description-Content-Type: text/markdown
+
+# openhire-mcp
+
+> **Jobs come to you. Your résumé never passes through OpenHire's servers, and we never store it.**
+> 岗位来找你，简历不经过我们的服务器，也不被我们存储。
+
+![MCP 1.0](https://img.shields.io/badge/MCP-1.0-58A6FF) ![privacy: local-first](https://img.shields.io/badge/privacy-local--first-3FB950) ![python ≥ 3.11](https://img.shields.io/badge/python-%E2%89%A5%203.11-C9D1D9) ![license: MIT](https://img.shields.io/badge/license-MIT-C9D1D9) ![v0.1 · sentinel](https://img.shields.io/badge/v0.1-sentinel-E3B341)
+
+An MCP server that turns your AI assistant (Claude, Cursor, Windsurf) into a private radar
+for **remote AI / Infra roles** — sourced directly from ~100 company career sites and their
+public ATS APIs (Greenhouse / Lever / Ashby). **No account. No signup. No résumé upload. Ever.**
+
+Matching runs on your machine; only an anonymous fingerprint and hard filters ever reach the
+server. This is the v0.1 「哨兵 / Sentinel」 reference implementation — see
+`design_handoff_openhire_v01/README.md` for the full protocol spec.
+
+---
+
+## Quickstart — under a minute
+
+```bash
+# 1. Install (pipx keeps it isolated and puts `ohp` on your PATH)
+pipx install openhire
+
+# 2. Get a job index. Default: download the public snapshot, then refresh it live.
+ohp bootstrap
+#   --fresh      crawl the public ATS from scratch (heuristic, free, no snapshot)
+#   --deepseek   higher-quality extraction using YOUR OWN DEEPSEEK_API_KEY
+
+# 3. Use it directly…
+ohp search --required-skills rust,k8s --remote --role-family engineering
+
+# …or connect it to an MCP client:
+ohp serve
+```
+
+For **Claude Desktop**, add to `claude_desktop_config.json`:
+
+```json
+{
+  "mcpServers": {
+    "openhire": {
+      "command": "ohp",
+      "args": ["serve"]
+    }
+  }
+}
+```
+
+> On Windows the config file lives at `%APPDATA%\Claude\claude_desktop_config.json` (for the
+> Microsoft Store build it is under `…\Packages\<Claude package>\LocalCache\Roaming\Claude\`).
+> Fully quit and reopen Claude Desktop after editing.
+
+---
+
+## What it does
+
+| Tool | What it gives you |
+|------|-------------------|
+| `search_jobs` | Hard-filter the live index; every result carries `verified_at`, `datePosted`, `days_open`, `ghost_score`, `remote_scope`, `eligible_regions`, `apply_channel`. Filter by `required_skills` (AND), `role_family`, `remote_scope`, `min_salary` + `currency`. |
+| `watch_intent` | Register a standing intent once — new matching jobs are waiting next time you check, even after you close the terminal. Accepts `required_skills` / `role_family` so sales / solutions roles stay out. |
+| `check_watches` | Pull the matches that are new since your last check (client-pull; stdio has no push). |
+| `authorize_application` | One explicit confirmation per job. It records your authorization and returns the employer's **own** application URL — you apply as yourself. It **cannot** accept a résumé. |
+| `get_company_info` | Aggregate, anonymous trust signals for one employer (`ghost_score_avg`, `active_jobs`, `index_built_at`). Never any candidate data. |
+
+Optional, entirely local: `ohp init --scan <dir>` derives a **skill fingerprint** from your
+own repos. You never write a résumé; the code never leaves your machine — only an anonymous
+vector does.
+
+## The five protocol fields
+
+Every listing is valid `schema.org/JobPosting`, plus:
+
+- `verified_at` — last moment confirmed live on the employer's own site
+- `source` — `employer_site | ats_public_api` (never a job board)
+- `ghost_score` — 0–1 likelihood the listing is not a real, active hire (ages off the **real**
+  posting date; lower is better)
+- `response_sla_days` — employer's committed response window (v0.1: always null)
+- `apply_channel` — always the employer's own application URL, deep-linked to the specific job
+
+## Privacy model
+
+| | |
+|---|---|
+| **Résumé / PII upload** | **never** — matching runs locally; a résumé never transits the server, and we never store one |
+| **What the server sees** | one anonymous, client-generated fingerprint + hard filters |
+| **Repo scan** | local-only · personal projects · explicit consent · opt-out anytime |
+| **Job sources** | first-party only: employer career pages + public ATS APIs (Greenhouse / Lever / Ashby) |
+
+## First-run data — the snapshot vs. fresh
+
+`ohp bootstrap` (default) downloads a small **public** index snapshot (a GitHub Release
+asset — `companies` + `jobs` only, **zero** user data) and then runs one incremental crawl to
+refresh `verified_at` / delisting. `--fresh` skips the snapshot and crawls the public ATS from
+scratch with the free offline heuristic extractor. Either way: no account, no PII.
+
+## Three rules this project will never break
+
+1. Your résumé stays on your machine — it never transits the server, and we never store it.
+2. Ranking is not for sale — it is only `f(match_quality, freshness)`, a locked pure function.
+3. Employers pay only for authorized, delivered outcomes — never for exposure. (v0.1 has no
+   billing at all.)
+
+These are enforced by CI (`tests/test_privacy.py`, `tests/test_ranking.py`,
+`tests/test_snapshot.py`).
+
+## Development
+
+```bash
+python -m venv .venv && . .venv/Scripts/activate   # Windows
+pip install -e ".[dev]"
+pytest        # privacy red lines + ranking + snapshot must be green
+```
+
+Set `OPENHIRE_DATABASE_URL=postgresql+psycopg://…` to run against Postgres instead of the
+default local SQLite file (`~/.openhire/openhire.db`).
+
+## Roadmap
+
+- **v0.2** — CN ATS adapters (Beisen / Moka) · `ghost_score` public beta
+- **v0.3** — Employer claim + verified badges · response-SLA enforcement (7-day auto-delist)
+- **v1.0** — Open, vendor-neutral schema extension for AI-readable job postings
+
+## License
+
+MIT © OpenHire Protocol · PRs welcome.

openhire-0.1.0/PROGRESS.md

@@ -0,0 +1,136 @@
+# PROGRESS · OpenHire `openhire-mcp` v0.1「哨兵」
+
+> 一页进度台账。新条目置顶。恢复会话请先读 `README.md` 再读本文件，勿重做已完成工作。
+
+---
+
+## 2026-07-12 — 真机验收修复 campaign（报告 003 → 逐项修，详见 reports/004）
+
+**P0-1 写路径挂死 — ✅ 已修复**
+- 根因：真库为 SQLite rollback-journal 模式，长驻 stdio server 的流式读（search/check 遍历 11,825 行）阻塞 watch_intent 写；async 事件循环里表现为 4 分钟挂死。apply 秒回是因当时无重叠读，证伪「写路径整体挂死」。
+- 复现对照：rollback-journal 读活跃时写 5.46s 后 `database is locked`；WAL 下 0.004s 提交。
+- 修复：`db/session.py` SQLite 引擎启用 `journal_mode=WAL` + `busy_timeout=5000` + `synchronous=NORMAL` + `check_same_thread=False`。真库已持久转 WAL。
+- 验证：新增 `tests/test_concurrency.py`（2 项）；全量 **100 passed**；T3 真库重跑（fp `88ba1102edb9205d`，4 线程读压下）watch_intent **0.046s** 返回、只落匿名数据。
+
+**P0-2 datePosted 造假 + ghost_score 恒 0 — ✅ 已修复**
+- 根因：入库丢弃 ATS 真实发布日，`first_seen_at`/`verified_at` 全写抓取时刻（全库 1 个 distinct）；ghost 用 first_seen（岗龄≈0）+ relist=0 → 恒 0。DeepSeek 未污染日期。
+- 修复：`Job` 加 `posted_at`/`updated_at` 列 + migrate；ingest 落库真实发布日、ghost 以真实岗龄为锚；`job_posting` 的 `datePosted` 用 posted_at、新增 `days_open`；新增 **免费** `ohp backfill-dates` 重抓 ATS 回填（不重跑抽取）。
+- 回填真库：96/0，真实发布日 11446，ghost 重算 11825。回归：`mongodb:7727896` days_open=110（≈3.6 月）、ghost 0→0.3565；全库 ghost min0/median0.081/max1.0，**57.7% >0**。新增 `tests/test_posting_dates.py`。
+
+**API 定型 5 项免费项 — ✅ 完成（107 passed）**
+- required_skills(AND)；min_salary 加 currency + 拆 require_stated_salary（未标薪资默认保留）；remote_scope 枚举 + eligible_regions；MCP `apply`→`authorize_application`；get_company_info 去 verified、last_crawled_at→index_built_at；watch_intent 加 fingerprint_notice。
+- role_family：参数+空列已冻结，**数据待跑**——报价 ≈¥29（¥50 硬上限），**等用户批准**。
+
+**role_family 抽取 — ✅ 已批准跑完**：新增 `classify_role_family_with_usage` + 断点续跑成本封顶的 `rebuild_role_family` + CLI `ohp extract-role-family`。全量 11,825/11,825 标注、0 失败、**¥17.53**（低于 ¥29 估价）。分布 eng4207/sales3373/ops2621/…。回归达标：`--role-family engineering` 搜索 0 条 sales；两个附录 B 销售岗均 role_family=sales。
+
+**附录 B 回归 — ✅**：`tests/test_appendix_b.py`（5 job_id 各锁一个契约：datePosted/销售岗排除/Ashby 写路径/荷兰无薪资岗保留-排除）。
+
+**文案红线 — ✅**：README 标语 + CLI 帮助/横幅「简历永不离开设备」→「简历不经过我们服务器，也不被我们存储」。
+
+**测试基线：112 passed。** 报告 003 的 P0-1/P0-2/6 项 API 定型/文案/附录 B 回归全部完成。
+
+**CLI 真机体验 7 步单 — ✅ 交付**（`reports/005`，新手向：每步一命令 + 预期输出 + 出错时截图哪里；预期输出为真库副本+临时目录实跑截取）。
+**顺手修一处小摩擦：** `ohp apply` 原来无论是否 `--no-open` 都打印「申请页（已打开）」；改为按实际结果打印「已在浏览器打开」/「请手动打开」。112 passed 不变。
+
+**下一步：** 用户按 `reports/005` 亲跑最后一轮验收 → 通过即开 **M4**。
+
+---
+
+## 2026-07-12（续）— M4 就绪修复 + M4 启动
+
+**两处 M4 就绪修复 — ✅（114 passed）**
+- ① `ohp search` 回显行漏掉 `--role-family`/`--limit` → 补上。
+- ② `ohp watch` + `watch_intent` 加可选 `required_skills`(AND) + `role_family`（非破坏；`_clean_filters` 白名单 + `check_watches` 透传）。实测 watch 加这两参数后命中不再混入销售/SA 岗。新增 2 项服务测试。
+
+**M4 代码就绪 — ✅（118 passed，详见 reports/006）**
+- 首跑数据（用户拍板混合）：快照=GitHub Release 资产、仅 jobs/companies、**零用户态**（构建时红线校验，违规即失败）。`ohp bootstrap` 默认 拉快照→增量刷新→打印快照龄；`--fresh` 现抓（启发式免费）；`--deepseek` 自带 key。开跑前先声明将做什么。新增 `pipeline/snapshot.py`、`ohp bootstrap`、`ohp snapshot-build`、`db.session.dispose_engine`、`config.SNAPSHOT_URL`、`tests/test_snapshot.py`(4)。
+- 实测：真快照 96/11825·13.2MB·零用户态校验通过；bootstrap 全流程（安装 96/11825 龄3天→刷新 新323/更340/下线452）✓；护栏（已有索引提示 --force）✓。
+- pipx：wheel 73KB 无数据大文件 → 全新隔离 venv 安装 → `ohp version`/命令齐全 ✓。终端用户 pipx 装后 config 可直接 `command: "ohp"`。
+- README：design_refs→真 Markdown，隐私口径收窄。
+- **待用户拍板**：GitHub 仓库地址、是否发 PyPI、是否落地 `smithery.yaml`（草案在 006）、快照刷新节奏。MCP 目录提交材料清单已备好、**未提交**。
+
+**M4 发布准备 — 四项拍板已落地（详见 reports/007）**
+- ① GitHub：个人账号下 `openhire` 仓库（URL 待用户名，`pyproject`/README/`smithery`/`config` 里占位 `OWNER`，拿到用户名我一次性 wire）。
+- ② PyPI：查得 `openhire` **未被占用 → 沿用**（无需改名）；CLI 保持 `ohp`。
+- ③ `smithery.yaml` **已落地**（stdio + 可选 OPENHIRE_DATABASE_URL）。
+- ④ 快照每周手动刷新：维护者操作卡 `docs/maintainer-snapshot-refresh.md`。
+- 另：`.gitignore` 补 `.pypirc`/token/WAL/`*.db.gz` 防护；`pyproject` 加 `[project.urls]`；sdist+wheel 构建通过。
+- **发布日操作单 `reports/007`**：GitHub/PyPI 双新手逐步指引，标注【可代劳】步骤；红线 token 只进本机 env/.pypirc、不进 git。
+
+**下一步：** 用户给 GitHub 用户名 → 我 wire URL；建仓库/推代码/建 Release 传快照；PyPI 注册+2FA+token → 我 twine 上传。
+
+---
+
+## 2026-07-11 — M4 前真机验收准备
+
+**已完成**
+- 建立常设工作制度：新增 `CLAUDE.md`（新会话恢复指引 + 工作约定 + 隐私红线 + 关键路径），补记本 `PROGRESS.md`（回填 M1–M3）。
+- 核对真机事实：默认 DB URL = `sqlite:///C:/Users/gdche/.openhire/openhire.db`（绝对路径，11,825 职位 / 96 公司 / 全量 deepseek 抽取）；`ohp.exe` 在 `.venv\Scripts\`，**不在系统 PATH**。
+
+**关键决策**
+- Claude Desktop 接入 `command` 用 `ohp.exe` 全路径（因未上 PATH，且 M4 才做 pipx/uvx）。这是**预判的头号发布阻塞项**：陌生开发者裸装无法直接 `command: "ohp"`。
+- serve 阶段 DB 走绝对路径默认值 → Claude Desktop 从任意 cwd 启动都命中同一份数据，无需在 config 里配 env。
+
+**下一步**
+- 交付真机验收三件套（Claude Desktop 分步接入指引 + 5 条验收对话由用户亲跑 + CLI `init→watch→apply` 体验清单）。
+- 用户跑完反馈摩擦点 → 按 v0.1 发布阻塞项处理 → 再开 M4（pipx/uvx 打包、正式 README、提交 MCP 目录）。
+
+**已发现并修复的摩擦（发布阻塞项 #1）**
+- `ohp search` / `check` 只打印 apply URL，**不打印 `job_id`**，而 `ohp apply` 恰需 `job_id` 参数 → 新手搜完不知道 apply 该敲什么。已修 `_print_job`：每条结果新增 `id=<job_id>` 及 `# 投递：ohp apply <job_id>` 提示行。98 测试仍全绿。
+
+**待用户确认**
+- 真机 5 条对话与 CLI 清单跑下来的摩擦点（报错/看不懂/步骤多），逐条反馈。
+
+---
+
+## 2026-07-10 — M3 验收通过（求职者旅程 CLI）
+
+**已完成（验收证据）**
+- 求职者旅程 CLI：`ohp init --scan <dir>` / `watch` / `check` / `apply` / `status`，客户端状态在 `~/.openhire`。
+- `init` 本地扫描仓库导出技能指纹（按扩展名 + 依赖清单 → 技能标签，规范化），先确认后落盘；代码/内容不出机器，匿名 id 形如 `#a3f9`。
+- `apply <job_id>` 打开链接**前**先在终端打印 JD 摘要（公司/职位/薪资/技能/JD 要点），`[Y/n]` 授权后再跳 —— 修复 Greenhouse embed 表单无 JD 上下文的问题。落 receipt 到 `~/.openhire/receipts.jsonl` + 服务端 `Application`。
+- 实测链路：watch → 新 JD 入库 → `check` 增量浮现；apply 全流程 + receipt。**98 项测试全绿**。
+
+**关键决策**
+- `check_watches` 以 `last_notified_at` 为游标（None=基线返回当前全部匹配），修复首拉返回 0 的问题。
+- stdio 无服务端推送 → `check` 为客户端主动拉取。
+
+---
+
+## 2026-07-09 — DeepSeek 全量抽取重建
+
+**已完成（验收证据）**
+- 新增 DeepSeek OpenAI 兼容后端（`base_url=https://api.deepseek.com`, `model=deepseek-chat`, key 从 `DEEPSEEK_API_KEY`）。
+- 先抽 100 条对比启发式确认质量提升 → 分批断点续跑全量：**11,825/11,825 转换，0 失败，实际花费 ¥27.03**（硬上限 ¥50，超则停）。
+- 抽取结果保留原启发式值作回退列（`*_fallback`），可对比/回滚。
+- 合并策略：skills 永远用 LLM；remote/salary 保留权威 ATS 值（除非 ATS 缺失）—— 修复朴素全量替换导致的 remote(48 改)/salary(43 失) 回退。
+- 复测「远程 Rust/K8s 基础设施岗」搜索：Top 8 中销售岗归零，回归确认。
+
+**关键决策**
+- 用最便宜 `deepseek-chat`；先 100 条抽样报差异、确认再全量；超 ¥50 先停问用户 —— 均按用户要求落地。
+- 共享连接池 `httpx.Client` + 24 workers（原每调用新建连接过慢）。
+
+---
+
+## 2026-07（M2）— MCP 服务端（四工具 + apply）
+
+**已完成（验收证据）**
+- FastMCP 服务 `openhire`，5 个工具：`search_jobs` / `get_company_info` / `watch_intent` / `check_watches` / `apply`。
+- 通过 5 条 Claude Desktop 验收对话脚本（搜索 / 公司信任 / 常驻意向 / 跨会话拉取 / 隐私红线对抗）+ 2 条内部验收。
+- 红线②落地：排序函数签名锁死 `f(match_quality, freshness)`，无任何付费参数（`test_ranking.py`）。
+- `search_jobs` 服务端只做硬过滤 + 固定排序，精排留客户端。
+- `apply` 结构上无 resume 参数 —— 无法接收简历；`assert_no_resume` 拒绝 resume/cv/file/email 等键 → `ERR_RESUME_NEVER_TRANSMITTED`。
+
+---
+
+## 2026-07（M1）— 数据管线
+
+**已完成（验收证据）**
+- 从公开 ATS API 入库（Greenhouse / Lever / Ashby），零爬虫；96 家已验证公司（含 anthropic/openai/cohere/mistral 等）。
+- 五协议字段：`verified_at` / `source` / `ghost_score` / `response_sla_days`(v0.1 恒 NULL) / `apply_channel`。
+- 变更/下线/重挂检测；`ghost_score = min(1.0, 0.15*relist_count + max(0, days_since_first_seen-45)/90*0.5)`（纯函数、可注入时钟）。
+- **apply_channel 修复**：21 家嵌入式公司（4,188 职位）全部重生成，DB 不变量「0 个非 ATS 宿主」，全部 HTTP 200 直达具体职位；补自动化测试「apply_channel 必须直达具体职位」。
+- Postgres 优先 schema，SQLite 为开发默认；方言感知 TypeDecorator（StringArray / JSONDict / TZDateTime 修复 SQLite naive datetime）。
+
+**关键决策**
+- 嵌入页无法可靠直达时，回退 Greenhouse `/embed/job_app` 官方申请表单（CoreWeave 等案例）。

openhire-0.1.0/README.md

@@ -0,0 +1,125 @@
+# openhire-mcp
+
+> **Jobs come to you. Your résumé never passes through OpenHire's servers, and we never store it.**
+> 岗位来找你，简历不经过我们的服务器，也不被我们存储。
+
+![MCP 1.0](https://img.shields.io/badge/MCP-1.0-58A6FF) ![privacy: local-first](https://img.shields.io/badge/privacy-local--first-3FB950) ![python ≥ 3.11](https://img.shields.io/badge/python-%E2%89%A5%203.11-C9D1D9) ![license: MIT](https://img.shields.io/badge/license-MIT-C9D1D9) ![v0.1 · sentinel](https://img.shields.io/badge/v0.1-sentinel-E3B341)
+
+An MCP server that turns your AI assistant (Claude, Cursor, Windsurf) into a private radar
+for **remote AI / Infra roles** — sourced directly from ~100 company career sites and their
+public ATS APIs (Greenhouse / Lever / Ashby). **No account. No signup. No résumé upload. Ever.**
+
+Matching runs on your machine; only an anonymous fingerprint and hard filters ever reach the
+server. This is the v0.1 「哨兵 / Sentinel」 reference implementation — see
+`design_handoff_openhire_v01/README.md` for the full protocol spec.
+
+---
+
+## Quickstart — under a minute
+
+```bash
+# 1. Install (pipx keeps it isolated and puts `ohp` on your PATH)
+pipx install openhire
+
+# 2. Get a job index. Default: download the public snapshot, then refresh it live.
+ohp bootstrap
+#   --fresh      crawl the public ATS from scratch (heuristic, free, no snapshot)
+#   --deepseek   higher-quality extraction using YOUR OWN DEEPSEEK_API_KEY
+
+# 3. Use it directly…
+ohp search --required-skills rust,k8s --remote --role-family engineering
+
+# …or connect it to an MCP client:
+ohp serve
+```
+
+For **Claude Desktop**, add to `claude_desktop_config.json`:
+
+```json
+{
+  "mcpServers": {
+    "openhire": {
+      "command": "ohp",
+      "args": ["serve"]
+    }
+  }
+}
+```
+
+> On Windows the config file lives at `%APPDATA%\Claude\claude_desktop_config.json` (for the
+> Microsoft Store build it is under `…\Packages\<Claude package>\LocalCache\Roaming\Claude\`).
+> Fully quit and reopen Claude Desktop after editing.
+
+---
+
+## What it does
+
+| Tool | What it gives you |
+|------|-------------------|
+| `search_jobs` | Hard-filter the live index; every result carries `verified_at`, `datePosted`, `days_open`, `ghost_score`, `remote_scope`, `eligible_regions`, `apply_channel`. Filter by `required_skills` (AND), `role_family`, `remote_scope`, `min_salary` + `currency`. |
+| `watch_intent` | Register a standing intent once — new matching jobs are waiting next time you check, even after you close the terminal. Accepts `required_skills` / `role_family` so sales / solutions roles stay out. |
+| `check_watches` | Pull the matches that are new since your last check (client-pull; stdio has no push). |
+| `authorize_application` | One explicit confirmation per job. It records your authorization and returns the employer's **own** application URL — you apply as yourself. It **cannot** accept a résumé. |
+| `get_company_info` | Aggregate, anonymous trust signals for one employer (`ghost_score_avg`, `active_jobs`, `index_built_at`). Never any candidate data. |
+
+Optional, entirely local: `ohp init --scan <dir>` derives a **skill fingerprint** from your
+own repos. You never write a résumé; the code never leaves your machine — only an anonymous
+vector does.
+
+## The five protocol fields
+
+Every listing is valid `schema.org/JobPosting`, plus:
+
+- `verified_at` — last moment confirmed live on the employer's own site
+- `source` — `employer_site | ats_public_api` (never a job board)
+- `ghost_score` — 0–1 likelihood the listing is not a real, active hire (ages off the **real**
+  posting date; lower is better)
+- `response_sla_days` — employer's committed response window (v0.1: always null)
+- `apply_channel` — always the employer's own application URL, deep-linked to the specific job
+
+## Privacy model
+
+| | |
+|---|---|
+| **Résumé / PII upload** | **never** — matching runs locally; a résumé never transits the server, and we never store one |
+| **What the server sees** | one anonymous, client-generated fingerprint + hard filters |
+| **Repo scan** | local-only · personal projects · explicit consent · opt-out anytime |
+| **Job sources** | first-party only: employer career pages + public ATS APIs (Greenhouse / Lever / Ashby) |
+
+## First-run data — the snapshot vs. fresh
+
+`ohp bootstrap` (default) downloads a small **public** index snapshot (a GitHub Release
+asset — `companies` + `jobs` only, **zero** user data) and then runs one incremental crawl to
+refresh `verified_at` / delisting. `--fresh` skips the snapshot and crawls the public ATS from
+scratch with the free offline heuristic extractor. Either way: no account, no PII.
+
+## Three rules this project will never break
+
+1. Your résumé stays on your machine — it never transits the server, and we never store it.
+2. Ranking is not for sale — it is only `f(match_quality, freshness)`, a locked pure function.
+3. Employers pay only for authorized, delivered outcomes — never for exposure. (v0.1 has no
+   billing at all.)
+
+These are enforced by CI (`tests/test_privacy.py`, `tests/test_ranking.py`,
+`tests/test_snapshot.py`).
+
+## Development
+
+```bash
+python -m venv .venv && . .venv/Scripts/activate   # Windows
+pip install -e ".[dev]"
+pytest        # privacy red lines + ranking + snapshot must be green
+```
+
+Set `OPENHIRE_DATABASE_URL=postgresql+psycopg://…` to run against Postgres instead of the
+default local SQLite file (`~/.openhire/openhire.db`).
+
+## Roadmap
+
+- **v0.2** — CN ATS adapters (Beisen / Moka) · `ghost_score` public beta
+- **v0.3** — Employer claim + verified badges · response-SLA enforcement (7-day auto-delist)
+- **v1.0** — Open, vendor-neutral schema extension for AI-readable job postings
+
+## License
+
+MIT © OpenHire Protocol · PRs welcome.