@mazjq/zhiro 0.1.2 → 0.2.1

package/README.md

@@ -2,7 +2,8 @@
 
 # Zhiro · 职路
 
-**Local-first AI career research and job hunting — with the agent you already use.**
+A job-search assistant that runs on your own computer: find jobs, look into companies, keep track of applications.
+Works with AI assistants such as Claude Code and Codex.
 
 English | [简体中文](README.zh-CN.md)
 
@@ -12,90 +13,144 @@ English | [简体中文](README.zh-CN.md)
 
 </div>
 
-Zhiro keeps your whole job hunt in one place on your own computer: which companies are hiring, what fits you, what
-closes this week, where each application stands — and lets your own AI agent (Claude Code, Codex, WorkBuddy, or any
-MCP client) research industries and companies, track things for you and fill in application forms, **without ever
-pressing Submit for you**. It is built first for campus recruiting in China (校招), and works for any job search.
+Every day Zhiro collects jobs from companies' recruiting sites, picks the ones that fit your degree, major and
+experience, and reminds you which ones close soon. It can look up listed companies' financial reports and
+announcements to help you decide whether a company is worth joining. After you apply, it can also keep track of your
+progress, read interview emails and fill in application forms.
+
+It is built mainly for campus recruiting in China (校招), and works for other job searches too.
 
 ![Zhiro: jobs ranked against your profile](docs/assets/jobs.png)
 
-| Today: what closes soon, what waits for you | Applications: every stage, with the job's own requirements |
+<sub>Every company, job and number in the screenshots is made up.</sub>
+
+## Why Zhiro
+
+### 1. Know what to learn next
+
+Zhiro gathers the requirements of the jobs related to your field, lists the skills your profile doesn't mention yet,
+and counts how many jobs and companies ask for each one. In the screenshot below, C++ is required by 4 jobs and ROS by 2.
+
+If you already have a skill, click *I can do this* and it is added to your profile. If you don't, click *haven't done
+this*. Job recommendations are recalculated from your answers.
+
+![Opportunity gaps: how many jobs ask for each skill](docs/assets/gaps.png)
+
+### 2. See how a company is doing
+
+For listed companies (A-shares, Hong Kong, US, NEEQ), Zhiro can look up:
+
+- three years of revenue, profit, growth, gross margin and R&D spending;
+- headcount, and the average cash the company pays per employee (salary, social insurance and so on);
+- the past year's announcements, such as employee share plans or equity incentives, new production lines, government
+  grants, being named a "专精特新" company (usually a niche firm that is hard to replace in its industry), and share
+  sales by major shareholders.
+
+You can also ask your AI assistant to research a whole industry. It writes a report from this data, broker research
+and what it finds online. Every point names its source, and the numbers it quotes are checked against the original
+data. If a company in the report is hiring, you can click through to its jobs.
+
+![A research report: key points, sources, the companies it names](docs/assets/research.png)
+
+<img src="docs/assets/company-intel.png" alt="A company's financials: revenue, R&D, pay per employee" width="520">
+
+### 3. Don't miss a job or a deadline
+
+Every day Zhiro reads the recruiting sites you subscribe to: companies' own hiring systems (Moka, Beisen, Feishu Jobs,
+51job, Zhaopin, iGuopin and more) and 高校人才网 (gaoxiaojob.com, where research institutes, labs, university teams and
+public institutions post). It can also read spreadsheets you keep yourself (Feishu Base, CSV files).
+
+The same job sometimes shows different deadlines on different sites. Zhiro compares them and notes which site each
+date came from and whether it has been verified.
+
+Your education, skills, projects and preferred cities are kept in your profile. Job recommendations, choosing a
+resume and filling in forms all take their information from there.
+
+| Today: jobs closing soon, things waiting for you | Profile: your education, skills and experience |
 |---|---|
-| ![Today](docs/assets/today.png) | ![Applications](docs/assets/pipeline.png) |
-
-<sub>Screenshots use made-up data.</sub>
-
-## What it does
-
-- **One job pool, always fresh.** Reads companies' own recruiting sites (Moka, Beisen, Feishu Jobs, Dayee, 51job,
-  Zhaopin, iGuopin, and more) and your own spreadsheets (Feishu Base, CSV links or files), every day. Deadlines are
-  reconciled across sources, and each one says where it came from.
-- **What fits you.** Your profile is a set of small, checkable cards (education, skills, projects, preferences); jobs
-  are matched against them, and gaps become directions to grow in, not silent filters.
-- **Today.** What closes in the next 14 days, what changed, and the few things waiting for your decision.
-- **Applications.** Watch → prepare → applied → tests and interviews → result, with reminders and an inbox that
-  turns recruiting mail into suggestions.
-- **Form filling up to Submit.** A separate Chrome profile fills in application forms from your profile and the
-  right resume; it stops before *Submit*, and you press it.
-- **Research with your own agent.** Ask about an industry or a company; your agent searches the web and gets
-  financial and industry data through Zhiro's tools, and the report links back to the jobs in your pool.
-- **Your agent, your rules.** Reversible actions happen and can be undone; anything with consequences becomes a card
-  you confirm; logins and credentials are never an agent's to touch.
-- **Your data stays yours.** Everything is in one folder on your machine. Upgrades run on a copy and are checked
-  before they replace anything; one command verifies, one rolls back.
+| ![Today](docs/assets/today.png) | ![Profile](docs/assets/profile.png) |
 
-## Install
+### 4. After you apply
 
-**Have Node 22.13 or newer?**
+Zhiro records where each application stands (preparing, applied, written test, interview, result) and reminds you as
+dates come up.
 
-```bash
-npx @mazjq/zhiro                 # try it
-npm i -g @mazjq/zhiro && zhiro   # keep it (the command is `zhiro`)
-```
+Once you connect your mailbox, Zhiro reads recruiting emails, recognises test and interview invitations, finds the
+application each one belongs to, and asks whether to update it.
 
-**No Node?** One line installs everything it needs (its own Node, checked against nodejs.org's checksums; your system
-is not touched):
+It can fill in application forms for you. It opens a separate Chrome window, fills the form from your profile and the
+right resume, and then stops so you can check it and press *Submit* yourself.
+
+| Applications: the progress and requirements of each one | Inbox: test and interview invitations found in your mail |
+|---|---|
+| ![Applications](docs/assets/pipeline.png) | ![Inbox](docs/assets/inbox.png) |
+
+### 5. Your data stays safe
+
+An AI assistant can only directly do things that can be undone. Anything else waits until you click to confirm. It
+never gets your logins for any site.
+
+All your data is kept in one folder on your computer. Zhiro backs it up before every upgrade, and if an upgrade goes
+wrong, one command restores it.
+
+## Install
+
+Windows: search for "PowerShell" in the Start menu, open it, paste this line and press Enter:
 
 ```powershell
-# Windows (PowerShell)
 irm https://github.com/mazjq/zhiro/releases/latest/download/install.ps1 | iex
 ```
 
+macOS / Linux: open Terminal, paste this line and press Enter:
+
 ```bash
-# macOS / Linux
 curl -fsSL https://github.com/mazjq/zhiro/releases/latest/download/install.sh | sh
 ```
 
-<details><summary>Prefer to read the script first?</summary>
+The first start asks where to keep your data (press Enter for the default), then opens Zhiro in your browser. Running
+the same line again later upgrades it.
 
-Download [`install.ps1`](https://github.com/mazjq/zhiro/releases/latest/download/install.ps1) or
-[`install.sh`](https://github.com/mazjq/zhiro/releases/latest/download/install.sh), look at it, then run it. Running it
-again upgrades; the previous version is kept.
+<details><summary>Already have Node.js 22.13 or newer</summary>
+
+```bash
+npx @mazjq/zhiro                 # try it
+npm i -g @mazjq/zhiro && zhiro   # keep it; start it later by typing zhiro
+```
 </details>
 
-**Want to change the code?** See [CONTRIBUTING.md](CONTRIBUTING.md): `git clone`, `pnpm install`, `pnpm dev`.
+<details><summary>Want to see what the install script does first</summary>
+
+Download [`install.ps1`](https://github.com/mazjq/zhiro/releases/latest/download/install.ps1) or
+[`install.sh`](https://github.com/mazjq/zhiro/releases/latest/download/install.sh) and read it before running it. The
+script downloads an official copy of Node.js into Zhiro's own folder (checked against the checksums nodejs.org
+publishes) and leaves the software already on your computer alone. When you upgrade, the previous version is kept.
+</details>
 
-The first start asks where to keep your data (press Enter for the usual place), then opens the app in your browser.
+To work on the code, see [CONTRIBUTING.md](CONTRIBUTING.md).
 
-### Let your agent install it
+### Let an AI assistant install it
 
-Paste this into Claude Code, Codex or any coding agent:
+If you use an AI coding assistant such as Claude Code or Codex, send it this:
 
 ```text
 Install Zhiro for me by following https://raw.githubusercontent.com/mazjq/zhiro/main/llms-install.md
 (fetch it with curl, not a summarizer).
 ```
 
-## Use it from your own agent
+## Connect your AI assistant
+
+You can use Zhiro in your browser without any AI assistant. Once you connect one, you can ask it in chat to find jobs,
+research companies or fill in applications for you.
 
-Zhiro is a set of MCP tools. Start Zhiro, then connect your agent once:
+AI assistants use Zhiro through MCP (a common way for AI to call outside tools). Start Zhiro, then set it up once for
+the assistant you use:
 
 <details><summary>Claude Code</summary>
 
 ```bash
 claude mcp add zhiro -- zhiro mcp            # Windows: claude mcp add zhiro -- cmd /c zhiro mcp
 ```
-Or open Claude Code inside Zhiro's agent folder (shown on the settings page) — everything is set up there.
+Or open Claude Code in the folder Zhiro prepares for it (shown on Zhiro's settings page); it is already set up there.
 [More](docs/harness/claude-code.md)
 </details>
 
@@ -104,7 +159,7 @@ Or open Claude Code inside Zhiro's agent folder (shown on the settings page) —
 ```bash
 codex mcp add zhiro --env ZHIRO_RESULT_MODE=text -- zhiro mcp
 # Windows: codex mcp add zhiro --env ZHIRO_RESULT_MODE=text -- cmd /c zhiro mcp
-# (ZHIRO_RESULT_MODE: Codex then gets each result once instead of twice — about half the tokens)
+# (ZHIRO_RESULT_MODE=text stops Codex receiving each result twice, which saves about half the usage)
 ```
 [More](docs/harness/codex.md)
 </details>
@@ -114,22 +169,38 @@ codex mcp add zhiro --env ZHIRO_RESULT_MODE=text -- zhiro mcp
 [WorkBuddy](docs/harness/workbuddy.md) · [any MCP client](docs/harness/others.md)
 </details>
 
-Then just ask: *"What closes this week?"*, *"Watch the control-algorithm jobs in Wuhan"*, *"Research the humanoid
-robot industry"*, *"Fill in the application for the first one."*
+Then just ask, for example: *"What closes this week?"*, *"Watch the control-algorithm jobs in Wuhan"*, *"Research the
+humanoid robot industry"*, *"Fill in the application for the first job."*
+
+### Use AI inside Zhiro
+
+You can also use AI without leaving Zhiro. Pick who answers on the settings page (Claude Code, Codex, DeepSeek, or
+Zhiro's built-in assistant), then click *Assistant* in the bottom-left corner to chat.
+
+With Claude Code or Codex you pick the model from a list. Zhiro reads the list from the Claude Code and Codex installed
+on your computer, newest first; if you don't pick one, it uses whatever you have set in them. The choice only applies
+inside Zhiro and doesn't change their own settings.
+
+The folder shown on the settings page is set up for your own AI assistant: open Claude Code or Codex there and there is
+nothing else to configure. [More](docs/harness/embedded.md)
+
+![Settings: choose the AI assistant and its model](docs/assets/settings.png)
+
+With the built-in assistant, choose a model provider (DeepSeek, OpenAI, Anthropic and others) in the same card and sign
+in or enter an API key.
 
-You can also pick which agent answers in Zhiro's own chat panel (Claude Code, Codex, or an optional built-in
-assistant): [docs/harness/embedded.md](docs/harness/embedded.md).
+![Settings: the built-in assistant and its model](docs/assets/builtin.png)
 
 ## Your data
 
-One folder, on your machine: `%APPDATA%\Zhiro\data` (Windows), `~/Library/Application Support/Zhiro` (macOS),
-`~/.local/share/zhiro` (Linux), or wherever you chose. Nothing is uploaded; the only network traffic is the
-recruiting sites you subscribe to and the agent or model you pick.
+Everything is kept in one folder on your computer. By default that is `%APPDATA%\Zhiro\data` (Windows),
+`~/Library/Application Support/Zhiro` (macOS) or `~/.local/share/zhiro` (Linux), or you can choose a folder at the first
+start. Zhiro only connects to the recruiting sites you subscribe to and the AI assistant you choose.
 
 ```bash
-zhiro vault verify      # is everything intact?
-zhiro vault backup      # a checked backup, any time
-zhiro vault rollback    # back to before the last upgrade
+zhiro vault verify      # check that the data is intact
+zhiro vault backup      # make a backup now
+zhiro vault rollback    # go back to before the last upgrade
 ```
 
 More: [docs/data.md](docs/data.md) · Bring your own spreadsheet: [docs/data-sources.md](docs/data-sources.md)
@@ -147,11 +218,12 @@ Most user docs are in Chinese for now; PRs translating them are welcome.
 
 ## Contributing
 
-Bug reports, recruiting sites to support, and pull requests are welcome — see [CONTRIBUTING.md](CONTRIBUTING.md).
-Adding a recruiting site is one file plus a test ([how](apps/daemon/src/data/readers/README.md)).
-Security issues: please report privately ([SECURITY.md](SECURITY.md)).
+Bug reports, suggestions for recruiting sites to support, and pull requests are welcome. See
+[CONTRIBUTING.md](CONTRIBUTING.md). Adding a recruiting site takes one file plus a test
+([how](apps/daemon/src/data/readers/README.md)). Please report security issues privately ([SECURITY.md](SECURITY.md)).
 
 ## License
 
 [MIT](LICENSE). The Moka reader includes three files from [career-ops](https://github.com/career-ops-hq/career-ops)
-(MIT).
+(MIT). The financial-report and announcement sources behind company profiles follow the public endpoints collected in
+Simon Lin's [a-stock-data](https://github.com/simonlin1212/a-stock-data) (Apache-2.0).

package/README.zh-CN.md

@@ -2,7 +2,7 @@
 
 # Zhiro · 职路
 
-**本地优先的 AI 职业调研与求职工具——用你已经在用的 agent。**
+在你自己电脑上运行的求职助手：找岗位、了解公司、管理投递。可以配合 Claude Code、Codex 等 AI 助手使用。
 
 [English](README.md) | 简体中文
 
@@ -12,75 +12,119 @@
 
 </div>
 
-职路把整个求职过程放在你自己的电脑上：哪些公司在招、哪些适合你、这周什么截止、每份申请走到哪一步。你自己的 AI agent（Claude Code、Codex、WorkBuddy，或任何支持 MCP 的客户端）可以帮你调研行业和公司、盯进度、填网申——**但提交按钮永远由你来点**。先为国内校招做，社招也能用。
+职路每天从各公司的招聘网站收集岗位，按你的学历、专业和经历挑出合适的，并提醒你哪些快截止了。它能查上市公司的财报和公告，帮你判断一家公司值不值得去。投递以后的进度、面试邮件和网申表格，也可以交给它。
 
-![职路：按你的画像排好的岗位](docs/assets/jobs.png)
+目前主要面向国内校招，社招也能用。
 
-| 今日：快截止的、等你拍板的 | 投递：每一步进度，岗位要求原文随手可查 |
+![职路：按你的资料排好的岗位](docs/assets/jobs.png)
+
+<sub>截图里的公司、岗位和数字都是虚构的。</sub>
+
+## 为什么用职路
+
+### 1. 知道自己该补什么
+
+职路会把和你方向相关的岗位要求汇总起来，列出你的资料里还没写到的技能，并算出每一项出现在多少个岗位、多少家公司的要求里。比如下图中，C++ 出现在 4 个岗位的要求里，ROS 出现在 2 个。
+
+其实会的，点「我会」，它会记进你的资料；没学过的点「没做过」。岗位推荐会按你的回答重新算一遍。
+
+![机会缺口：每项技能出现在几个岗位的要求里](docs/assets/gaps.png)
+
+### 2. 了解公司的经营情况
+
+对上市公司（A 股、港股、美股、新三板），职路能查到：
+
+- 近三年的营收、利润、增长率、毛利率和研发投入；
+- 员工人数，以及公司平均为每位员工支付的现金（工资、社保等）；
+- 近一年的公告，比如员工持股或股权激励、扩建产线、获得政府补助、被认定为「专精特新」企业（一般是某个细分领域里很难被替代的公司）、大股东减持。
+
+你也可以让 AI 助手调研一个行业。它会结合这些数据、券商研报和网上的资料写一份报告，每条结论注明出处，引用的数字会自动和原始数据核对一遍。报告里提到的公司如果正在招聘，可以直接点进去看岗位。
+
+![调研报告：要点、出处、提到的公司](docs/assets/research.png)
+
+<img src="docs/assets/company-intel.png" alt="一家公司的经营情况：营收、研发、人均薪酬" width="520">
+
+### 3. 不错过岗位和截止日期
+
+你订阅的招聘网站，职路每天自动去读一遍：公司自己的招聘系统（Moka、北森、飞书招聘、前程无忧、智联校园、国聘等），还有高校人才网（研究院、实验室、高校团队、事业单位常在这里发公告）。也可以读你自己整理的表格（飞书多维表格、CSV 文件）。
+
+同一个岗位在不同网站上写的截止日期有时对不上。职路会把它们放在一起比对，注明每个日期来自哪个网站、有没有核实过。
+
+你的学历、技能、项目经历和意向城市都存在「求职资料」里。推荐岗位、挑简历、填网申，都从这里取信息。
+
+| 今日：快截止的岗位、等你确认的事 | 求职资料：你的学历、技能、经历 |
 |---|---|
-| ![今日](docs/assets/today.png) | ![投递](docs/assets/pipeline.png) |
+| ![今日](docs/assets/today.png) | ![求职资料](docs/assets/profile.png) |
 
-<sub>截图里都是虚构数据。</sub>
+### 4. 投递之后
 
-## 能做什么
+每份申请到了哪一步（准备中、已投递、笔试、面试、结果）都会记下来，日子临近时提醒你。
 
-- **一个岗位库，每天自动更新。** 读公司自己的招聘站（Moka、北森、飞书招聘、大易、前程无忧、智联校园、国聘等）和你自己的表格（飞书多维表格、CSV 链接或文件）。多个来源的截止日期会对照合并，每条都标明出处。
-- **哪些适合你。** 画像是一张张能核对的小卡片（教育、技能、项目、偏好）；岗位按它匹配，差距会变成「往哪补」，而不是悄悄筛掉。
-- **今日。** 14 天内截止的、有变化的、等你拍板的几件事。
-- **投递进度。** 关注 → 准备申请 → 已投递 → 笔试面试 → 结果，带提醒；收件箱把招聘邮件变成建议，由你确认。
-- **网申填到提交前。** 用一个单独的 Chrome 档案，按你的画像和合适的那份简历填表；停在「提交」前，由你自己点。
-- **用你自己的 agent 做调研。** 问一个行业或一家公司：你的 agent 上网查，再通过职路的工具拿财报、行业研报等数据，报告和你岗位库里的岗位对上号。
-- **你的 agent，你定规矩。** 能撤销的事它直接做、你能撤销；有影响的事变成确认卡等你点；登录和凭据 agent 碰不到。
-- **数据在你手里。** 全部在你电脑上的一个文件夹里。升级在副本上做、核对全对才换上；一条命令检查，一条命令回退。
+接入邮箱后，职路会读招聘邮件，认出笔试和面试通知，找到对应的申请，问你要不要更新进度。
 
-## 安装
+网申表格也可以让它填。它会打开一个单独的 Chrome 窗口，用你的资料和合适的简历把表格填好，然后停下来，由你检查后自己点提交。
+
+| 投递：每份申请的进度和岗位要求 | 收件箱：从招聘邮件里认出的笔试、面试通知 |
+|---|---|
+| ![投递](docs/assets/pipeline.png) | ![收件箱](docs/assets/inbox.png) |
 
-**装了 Node 22.13 或更新？**
+### 5. 数据安全
 
-```bash
-npx @mazjq/zhiro                 # 试一下
-npm i -g @mazjq/zhiro && zhiro   # 长期用（装好后命令叫 zhiro）
-```
+AI 助手只能直接做可以撤销的操作，其他操作要你点确认才会执行。你在各个网站的登录信息它拿不到。
+
+所有数据都存在你电脑上的一个文件夹里。每次升级前会先备份，升级出了问题，一条命令就能恢复。
+
+## 安装
 
-**没有 Node？** 一行命令，自带它需要的一切（自带一份官方 Node，按 nodejs.org 的校验和核对过；不动你系统里的东西）：
+Windows：在开始菜单搜「PowerShell」并打开，粘贴下面这行，回车：
 
 ```powershell
-# Windows（PowerShell）
 irm https://github.com/mazjq/zhiro/releases/latest/download/install.ps1 | iex
 ```
 
+macOS / Linux：打开「终端」，粘贴下面这行，回车：
+
 ```bash
-# macOS / Linux
 curl -fsSL https://github.com/mazjq/zhiro/releases/latest/download/install.sh | sh
 ```
 
-<details><summary>想先看看脚本再运行？</summary>
+装好后，第一次启动会问数据存在哪（直接回车用默认位置），然后在浏览器里打开职路。以后再运行同一行命令就是升级。
+
+<details><summary>已经装了 Node.js 22.13 或更新版本</summary>
 
-下载 [`install.ps1`](https://github.com/mazjq/zhiro/releases/latest/download/install.ps1) 或 [`install.sh`](https://github.com/mazjq/zhiro/releases/latest/download/install.sh)，看过再运行。再运行一次就是升级，上一个版本会保留。
+```bash
+npx @mazjq/zhiro                 # 先试一下
+npm i -g @mazjq/zhiro && zhiro   # 长期用，之后输入 zhiro 就能启动
+```
 </details>
 
-**想改代码？** 见 [CONTRIBUTING.md](CONTRIBUTING.md)：`git clone`、`pnpm install`、`pnpm dev`。
+<details><summary>想先看看安装脚本做了什么</summary>
 
-第一次启动会问数据放哪（直接回车用默认位置），然后在浏览器里打开。
+下载 [`install.ps1`](https://github.com/mazjq/zhiro/releases/latest/download/install.ps1) 或 [`install.sh`](https://github.com/mazjq/zhiro/releases/latest/download/install.sh) 看过再运行。脚本会下载一份官方的 Node.js 放在职路自己的文件夹里（按 nodejs.org 公布的校验值核对），不改动你电脑上原有的软件。升级时上一个版本会保留。
+</details>
+
+想参与开发，见 [CONTRIBUTING.md](CONTRIBUTING.md)。
 
-### 让你的 agent 帮你装
+### 让 AI 助手帮你装
 
-把这句话贴给 Claude Code、Codex 或任何编程 agent：
+如果你在用 Claude Code、Codex 这类 AI 编程助手，把这句话发给它：
 
 ```text
 按 https://raw.githubusercontent.com/mazjq/zhiro/main/llms-install.md 帮我安装 Zhiro（用 curl 取原文，不要用网页摘要）。
 ```
 
-## 在你自己的 agent 里用
+## 接到你的 AI 助手
+
+不接 AI 助手，也能直接在浏览器里用职路。接上以后，你可以用聊天的方式让 AI 助手替你查岗位、做调研、填网申。
 
-职路是一组 MCP 工具。启动职路后，给你的 agent 接一次：
+AI 助手通过 MCP（一种让 AI 调用外部工具的通用接口）使用职路。先启动职路，再按你用的助手设置一次：
 
 <details><summary>Claude Code</summary>
 
 ```bash
 claude mcp add zhiro -- zhiro mcp            # Windows：claude mcp add zhiro -- cmd /c zhiro mcp
 ```
-或者在职路的 agent 工作目录（设置页里有位置）里打开 Claude Code，那里都配好了。[详细](docs/harness/claude-code.md)
+也可以在职路准备好的文件夹里打开 Claude Code（位置在职路的设置页），那里已经设置好了。[详细](docs/harness/claude-code.md)
 </details>
 
 <details><summary>Codex</summary>
@@ -88,7 +132,7 @@ claude mcp add zhiro -- zhiro mcp            # Windows：claude mcp add zhiro --
 ```bash
 codex mcp add zhiro --env ZHIRO_RESULT_MODE=text -- zhiro mcp
 # Windows：codex mcp add zhiro --env ZHIRO_RESULT_MODE=text -- cmd /c zhiro mcp
-#（ZHIRO_RESULT_MODE：让 Codex 每个结果只收一份，省约一半 token）
+#（ZHIRO_RESULT_MODE=text 让 Codex 少收一份重复内容，大约省一半用量）
 ```
 [详细](docs/harness/codex.md)
 </details>
@@ -98,18 +142,30 @@ codex mcp add zhiro --env ZHIRO_RESULT_MODE=text -- zhiro mcp
 [WorkBuddy](docs/harness/workbuddy.md) · [任何 MCP 客户端](docs/harness/others.md)
 </details>
 
-然后直接说：「这周有什么截止的」「关注武汉的控制算法岗」「调研一下人形机器人行业」「给第一个岗位填网申」。
+设置好后直接跟它说，比如：「这周有什么快截止的」「关注武汉的控制算法岗」「调研一下人形机器人行业」「帮我填第一个岗位的网申」。
+
+### 在职路里直接用 AI
+
+不想另开 AI 助手，也可以在职路里直接用。在设置页选好由谁来回答（Claude Code、Codex、DeepSeek，或职路自带的助手），之后点左下角的「助手」就能聊天。
+
+选 Claude Code 或 Codex 时，模型可以直接在下拉框里挑。列表是职路从你电脑上装的 Claude Code、Codex 里读出来的，最新的排在前面；不挑就和你平时在它们里面设的一样。这里的选择只影响职路，不会改动它们自己的设置。
+
+设置页里那行文件夹地址，是给你自己的 AI 助手准备好的：在那里打开 Claude Code 或 Codex，不用再做任何设置。[详细](docs/harness/embedded.md)
+
+![设置页：选 AI 助手和模型](docs/assets/settings.png)
+
+选「内置助手」时，在同一张卡片里选模型服务（DeepSeek、OpenAI、Anthropic 等）并登录或填 API key。
 
-职路窗口里的聊天面板也可以选由谁来回答（Claude Code、Codex，或可选的内置助手）：[docs/harness/embedded.md](docs/harness/embedded.md)。
+![设置页：内置助手和它用的模型](docs/assets/builtin.png)
 
 ## 你的数据
 
-一个文件夹，在你的电脑上：`%APPDATA%\Zhiro\data`（Windows）、`~/Library/Application Support/Zhiro`（macOS）、`~/.local/share/zhiro`（Linux），或者你选的位置。不上传任何东西；联网的只有你订阅的招聘站、你选的 agent 或模型。
+都存在你电脑上的一个文件夹里，默认位置是 `%APPDATA%\Zhiro\data`（Windows）、`~/Library/Application Support/Zhiro`（macOS）、`~/.local/share/zhiro`（Linux），也可以第一次启动时自己选。职路只会连接你订阅的招聘网站，以及你选用的 AI 助手。
 
 ```bash
-zhiro vault verify      # 数据完整吗
-zhiro vault backup      # 随时备份一次（带检查）
-zhiro vault rollback    # 回到上次升级前
+zhiro vault verify      # 检查数据是否完整
+zhiro vault backup      # 手动备份一次
+zhiro vault rollback    # 恢复到上次升级前
 ```
 
 详细：[docs/data.md](docs/data.md) · 接你自己的表格：[docs/data-sources.md](docs/data-sources.md)
@@ -129,4 +185,4 @@ zhiro vault rollback    # 回到上次升级前
 
 ## 许可
 
-[MIT](LICENSE)。Moka 读取器包含来自 [career-ops](https://github.com/career-ops-hq/career-ops) 的三个文件（MIT）。
+[MIT](LICENSE)。Moka 读取器包含来自 [career-ops](https://github.com/career-ops-hq/career-ops) 的三个文件（MIT）。企业情况的财报、公告接口参考了 Simon Lin 的 [a-stock-data](https://github.com/simonlin1212/a-stock-data)（Apache-2.0）整理的公开数据源。

package/agent-skills/zhiro/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: zhiro
-description: 用 Zhiro 管校招：找岗位、关注、准备申请、网申填写（到提交前）、申请进展、今日截止、主题调研、接岗位表。当用户提到 岗位 / 校招 / 秋招 / 投递 / 截止 / 申请 / 简历投到哪 / 调研某行业 / zhiro / jobos 时使用。前提是 Zhiro 在跑且 MCP `zhiro` 可用（老配置里可能叫 `jobos`，是同一个）。
+description: 用 Zhiro 管校招：找岗位、关注、准备申请、网申填写（到提交前）、申请进展、今日截止、主题调研、接岗位表。当用户提到 岗位 / 校招 / 秋招 / 投递 / 截止 / 申请 / 简历投到哪 / 调研某行业 / BOSS 直聘 / 实习 / zhiro / jobos 时使用。前提是 Zhiro 在跑且 MCP `zhiro` 可用（老配置里可能叫 `jobos`，是同一个）。
 ---
 
 # Zhiro · 职路
@@ -59,9 +59,15 @@ Zhiro 是本机的校招系统（岗位库、截止、申请、画像、调研
 
 交回的报告只是存档，不会改动岗位库。
 
+## BOSS 直聘：用户自己浏览，用捕获书签收
+
+BOSS 直聘的用户协议不允许程序（爬虫、「拟人程序」）代为读取、复制平台内容。不要替用户在 BOSS 上自动搜索、翻页或读取岗位，也不要用浏览器工具代开 BOSS 页面。用户说「去 BOSS 看看」时，请用户自己在浏览器里看，遇到合适的岗位用 Zhiro 的捕获书签收进库；agent 用 `listing.capture` 送 BOSS 链接会被拒绝。
+
 ## 公司经营状况
 
-`company.intel` 先看卡；`stale` 或没数据再 `company.intelRefresh`；非上市公司用 `research.company`（先问档，联网调研）。
+`company.intel` 先看卡：过期的类会在后台自动更新（`refreshing` 为真时稍后再读）；从没生成过（`listing` 为空）才调 `company.intelRefresh`；非上市公司用 `research.company`（先问档，联网调研）。
+
+用户问「这几家哪家更值得去」「帮我排一下」时，先 `employer.compare` `{ "companies": ["汇川技术", "埃斯顿", …] }`（最多 15 家 A 股、新三板、港股、美股公司，不花模型额度；新三板按同一证监会行业的 A 股公司来量，港美股按 peerIndustry 或一起比的 A 股公司所在行业、没有待遇信号；每家的 yardstick 要告诉用户）。它不给总分：每家在本行业同行里的百分位、在三种偏好下的名次和名次范围、谁在四项上都不如谁、同期一起发生的变化。转述时：说「在均衡偏好下第 2 名，通常第 2–3 名」，不要说「综合得分」；「进前三的比例」是偏好情景的比例，不是概率；`movedTogether` 只是同时发生，原因要读年报 / 公告再说；未上市和港美股单独讲，不要硬排进去。用户说更看重哪方面，就把 `weights` 带上重算。
 
 ## 接岗位表、运维
 

package/dist/acp-bridge.js

@@ -1,11 +1,11 @@
 import {
   jobosTool
-} from "./chunk-ZL2RKXOF.js";
-import "./chunk-UMCKC7XP.js";
-import "./chunk-NNH4ZG6F.js";
-import "./chunk-LH25QWMP.js";
-import "./chunk-YYNUPNBH.js";
-import "./chunk-5KCN32HQ.js";
+} from "./chunk-ZUUYRVX6.js";
+import "./chunk-VOUILWMI.js";
+import "./chunk-H3EFVYGD.js";
+import "./chunk-ZKD2BKN3.js";
+import "./chunk-4UWMDKMH.js";
+import "./chunk-4EPNDXRS.js";
 import "./chunk-4JJEOOW2.js";
 import {
   __name