dsh-model 0.0.1 → 0.15.0

package/README.en.md

@@ -0,0 +1,127 @@
+<div align="center">
+
+# dsh-model
+
+**Many model sources behind one OpenAI-compatible endpoint — for dsh and every other tool you use.**
+
+<p>
+  <a href="LICENSE"><img src="https://img.shields.io/github/license/AIcivilization/dsh-model" alt="MIT license"></a>
+  <a href="https://www.npmjs.com/package/dsh-model"><img src="https://img.shields.io/npm/v/dsh-model?color=cb3837&logo=npm" alt="npm version"></a>
+  <img src="https://img.shields.io/badge/DeepSeek%20Harness-%E2%89%A5%200.2.0--rc.2-4176E6" alt="DeepSeek Harness ≥ 0.2.0-rc.2">
+  <img src="https://img.shields.io/badge/platform-macOS%20%7C%20Linux-blue" alt="Platform: macOS / Linux">
+  <img src="https://img.shields.io/badge/node-%E2%89%A5%2020-339933?logo=node.js&logoColor=white" alt="Node ≥ 20">
+  <a href="https://github.com/router-for-me/CLIProxyAPI"><img src="https://img.shields.io/badge/engine-CLIProxyAPI%208.0.13-555" alt="Engine: CLIProxyAPI 8.0.13"></a>
+</p>
+
+English · [中文](README.md)
+
+</div>
+
+## What it is
+
+You may have several model sources at once: free WorkBuddy credits, an OpenCode Zen key, paid Codex or Kimi subscriptions. Each has its own API and sign-in, and every tool needs its own setup.
+
+dsh-model turns them into **one address and one kind of key**:
+
+- **In dsh**: the models show up in dsh's model list automatically, grouped by source. Nothing to configure.
+- **Everywhere else**: any editor or script that speaks the OpenAI API uses the same address and key to reach every model.
+- **Low maintenance**: one command to install; a key is always required; uninstall restores dsh's config exactly.
+- **On a server**: runs next to [dsh-vps](https://github.com/AIcivilization/dsh-vps) on the same VPS and opens a public address on dsh's own domain, so your laptop and phone can use it too.
+
+> A personal tool: one person, one account per provider, not a public service.
+
+## Quick start
+
+```bash
+npm install -g dsh-model
+dsh-model setup
+```
+
+On a dsh-vps server, use `sudo dsh-model setup`.
+
+`setup` is safe to re-run. It:
+
+1. downloads and verifies the engine and runs it on `127.0.0.1:8317/v1`, with key auth enforced;
+2. connects OpenCode Zen (enter your key; it is tested before being saved) and WorkBuddy (dsh-model's own bridge);
+3. installs the management page into dsh and syncs the models into dsh;
+4. on a VPS, also opens a public address `https://<dsh's domain>:9443/v1` (or `<IP>.sslip.io` when there is no domain).
+
+## Manage it in dsh: Settings → dsh-model
+
+| Section | What it holds |
+|---|---|
+| **Endpoint** | Local and public addresses, each with a Copy button |
+| **Sources** | One switch per provider: on connects (and signs in if needed), off pauses and keeps the sign-in. Each row is tagged Free / Free · limited / Paid and shows plan usage or remaining credits |
+| **Models** | Pick which models of each source appear in dsh (5 suggested per source by default); the rest stay available through the endpoint |
+| **API keys** | OpenAI-style, one per device. Requests, success rate, latency and speed over 24 hours; copy, rotate, revoke |
+| **Model stats** | Success rate, latency and speed per model |
+| **About** | Current and latest npm version; one-click Update and Uninstall (on a VPS these need root, so the commands are shown instead) |
+
+If the service is not set up on this machine yet, the page says what is missing and offers a "Set up now" button, so you can also start by installing dsh-model from the dsh plugin marketplace and finish from this page.
+
+dsh-model does every sign-in itself. No provider CLI is installed and the server needs no browser:
+
+- **WorkBuddy, Codex, Kimi, Grok, Muse**: open the link and approve; if a code is needed, the dialog shows it;
+- **Devin, Claude, Antigravity**: these send the browser to a fixed local address that will not load. That is expected: bring the address back and press "Paste from clipboard".
+
+If a source is signed in but has no usable plan (say, a Kimi account without Kimi Code), dsh-model detects it, hides its models in dsh and shows where to subscribe.
+
+> After updating dsh-model, restart dsh once so the management page picks up the new code. Model and switch changes need no restart.
+
+## Sources and cost
+
+| Source | Cost | Notes |
+|---|---|---|
+| WorkBuddy / WorkBuddy AI | Free | Credits come with the account |
+| Codex | Paid | ChatGPT Plus / Pro |
+| Kimi | Paid | Kimi Code plan |
+| Grok | Paid | SuperGrok / X Premium+ |
+| Muse | Paid | Muse Code plan |
+| Devin | Paid | The free tier is refused in testing |
+| OpenCode Zen | Paid | Pay as you go after a top-up; its free models only work inside the OpenCode app |
+| Claude, Antigravity | High risk, hidden by default | The providers have banned third-party use before |
+
+dsh-model only connects; it provides no quota of its own.
+
+## Commands
+
+| Command | What it does |
+|---|---|
+| `setup` | Install / repair everything (idempotent) |
+| `sources` | Sources: switch, cost, sign-in state, usage |
+| `source enable\|disable\|logout <source>` | Turn on (signing in if needed) / off / sign out |
+| `key list\|add\|revoke\|rotate <name>` | Access keys, one per device |
+| `stats` | Success rate, latency and speed by key, source and model |
+| `status` / `doctor [--e2e]` | Overview / health checks |
+| `models [sync]` | List models / sync them to dsh |
+| `remote enable --via caddy\|ssh\|tailscale` / `remote disable` | Remote access (on a VPS, setup already enables caddy) |
+| `engine version\|upgrade\|rollback` | Engine version |
+| `logs [--bridge]` / `repair` | Logs / repair |
+| `update [--to ver]` | Update to the latest npm version: program, dsh plugin and services together |
+| `uninstall` | Clean uninstall |
+
+Every command accepts `--lang zh|en` and `--json`.
+
+## Architecture
+
+```
+dsh · editors · scripts ──Bearer key──▶ engine 127.0.0.1:8317/v1 (the only entry)
+                                          │      ▲ VPS: Caddy https://<dsh domain>:9443
+             ┌────────────────────────────┼─────────────────────────┐
+             ▼                            ▼                         ▼
+   subscription OAuth (Codex, Kimi…)   opencode/* (your key)   workbuddy/* → dsh-model's bridge
+```
+
+- The engine is [CLIProxyAPI](https://github.com/router-for-me/CLIProxyAPI) (MIT), version-pinned and sha256-verified. The WorkBuddy protocol layer is ported from [dsh-workbuddy-connect](https://github.com/corrinehu/dsh-workbuddy-connect) (MIT); the usage endpoints follow [CodexBar](https://github.com/steipete/CodexBar) (MIT).
+- dsh-model writes only two things into dsh: the model providers and `refs.DSH_MODEL_API_KEY`. On uninstall they are restored byte for byte if nobody changed them.
+
+See the [design doc](docs/DESIGN.md) (Chinese).
+
+## Notes
+
+- Using subscription and WorkBuddy credentials outside the official clients may break the providers' terms and put the account at risk. Decide for yourself.
+- macOS and Linux, Node ≥ 20.
+
+## License
+
+MIT

package/README.md

@@ -1,73 +1,126 @@
+<div align="center">
+
 # dsh-model
 
-> **无需 API key，装上就能用前沿模型。** No login, no sign-up, no API key — the models are just there.
+**把多家模型汇成一个 OpenAI 兼容地址，dsh 和你的其他软件都能直接用。**
+
+<p>
+  <a href="LICENSE"><img src="https://img.shields.io/github/license/AIcivilization/dsh-model" alt="MIT license"></a>
+  <a href="https://www.npmjs.com/package/dsh-model"><img src="https://img.shields.io/npm/v/dsh-model?color=cb3837&logo=npm" alt="npm version"></a>
+  <img src="https://img.shields.io/badge/DeepSeek%20Harness-%E2%89%A5%200.2.0--rc.2-4176E6" alt="DeepSeek Harness ≥ 0.2.0-rc.2">
+  <img src="https://img.shields.io/badge/platform-macOS%20%7C%20Linux-blue" alt="Platform: macOS / Linux">
+  <img src="https://img.shields.io/badge/node-%E2%89%A5%2020-339933?logo=node.js&logoColor=white" alt="Node ≥ 20">
+  <a href="https://github.com/router-for-me/CLIProxyAPI"><img src="https://img.shields.io/badge/engine-CLIProxyAPI%208.0.13-555" alt="Engine: CLIProxyAPI 8.0.13"></a>
+</p>
+
+[English](README.en.md) · 中文
+
+</div>
+
+## 它是什么
+
+你可能同时有好几个模型来源：WorkBuddy 的免费积分、OpenCode Zen 的 key，还有已经付费的 Codex、Kimi 订阅。它们的接口和登录方式各不相同，每个软件都要分别配置。
 
-一条命令，把你本机（或海外 VPS）的 Codex、Claude Code、Grok Build、Muse 等 CLI 订阅，变成 dsh 和所有本地软件都能用的模型——OpenAI 兼容，填个地址就行。
+dsh-model 把它们合成**一个地址、一种 key**：
 
-## 它解决什么
+- **dsh 里**：装好后自动出现在 dsh 的模型列表里，按来源分组，不用手动配置。
+- **其他软件**：编辑器、脚本，凡是支持 OpenAI 接口的，填上同一个地址和 key 就能用全部模型。
+- **省心**：一条命令装好；始终要求 key，不会出现谁都能用的情况；卸载时把 dsh 的配置原样还原。
+- **服务器上**：可以和 [dsh-vps](https://github.com/AIcivilization/dsh-vps) 装在同一台 VPS 上，自动用 dsh 的域名开一个对外地址，你的电脑、手机都能访问。
 
-- 装了 Claude Code / Codex 订阅，却没法在 dsh 里直接用；
-- 想在编辑器插件、脚本、其他 Agent 框架里用订阅模型，到处买 API key；
-- 国内连不上国际模型：海外 VPS 部署 dsh-model 作中继，国内直连 VPS 即可；
-- 一个 key 都没有：内置免费模型池（DeepSeek V4 Flash、MiMo、Hy3、Muse Spark 等），开箱即聊。
+> 个人自用工具：一个人、每家一个账号，不对外提供服务。
 
 ## 快速开始
 
 ```bash
 npm install -g dsh-model
-dsh-model setup        # 安装引擎 + 内置免费模型，接入 dsh
+dsh-model setup
 ```
 
-打开 dsh → 获取可用模型 → 直接对话。想用 GPT/Claude 再增量登录：
+服务器上（装了 dsh-vps 的 VPS）用 `sudo dsh-model setup`。
 
-```bash
-dsh-model login codex  # device-code 登录，按提示在浏览器完成
-```
+`setup` 可以重复执行，它会依次：
 
-海外中继（国内用户推荐）：
+1. 下载并校验引擎，在本机 `127.0.0.1:8317/v1` 上运行，强制 key 鉴权；
+2. 接入 OpenCode Zen（输入你的 key，先实测再保存）和 WorkBuddy（dsh-model 自己的 bridge）；
+3. 把管理页装进 dsh，把模型同步进 dsh；
+4. 在 VPS 上，额外开通对外地址 `https://<dsh 的域名>:9443/v1`（没有域名时用 `<IP>.sslip.io`）。
 
-```bash
-# 在海外 VPS 上
-dsh-model setup && dsh-model relay enable
-# 国内任何软件填 https://<vps>/v1 + token
-```
+## 在 dsh 里管理：设置 → dsh-model
 
-## 架构
+| 区块 | 内容 |
+|---|---|
+| **访问地址** | 本机地址和外网地址，各带「复制」按钮 |
+| **来源** | 每家一个开关：打开就接入（没登录会弹出登录框），关闭只停用、保留登录。每行标出「免费 / 免费·额度少 / 需付费」，下面显示套餐用量或剩余积分 |
+| **选模型** | 每个来源选哪些模型显示在 dsh 里，默认每家推荐 5 个，其余的经统一地址照样可用 |
+| **API key** | OpenAI 格式，每台设备一把。显示 24 小时内的请求数、成功率、延迟和速度，可复制、轮换、吊销 |
+| **模型统计** | 按模型显示成功率、延迟和速度 |
+| **关于** | 当前版本和 npm 上的最新版本；「更新」「卸载」一键完成（VPS 上要 root，这里给出命令） |
 
-```
-dsh · 编辑器 · 脚本 · 任意 OpenAI 兼容软件
-        │
-        ▼
-  dsh-model（薄集成层：setup / connect-dsh / relay / status）
-        │
-        ▼
-  CLIProxyAPI 引擎（成熟核心：各家 OAuth 直连 + 全协议出口）
-```
+还没在这台机器上装好服务时，这一页会说明缺什么，并提供「一键安装」按钮。所以也可以先从 dsh 的插件市场装 dsh-model，再在这一页里点安装。
+
+登录全由 dsh-model 自己完成，服务器上不装各家的 CLI，也不需要浏览器：
+
+- **WorkBuddy、Codex、Kimi、Grok、Muse**：打开链接授权即可，要验证码的弹窗里会给出；
+- **Devin、Claude、Antigravity**：这几家会把浏览器跳到一个固定的本机地址，页面打不开是正常的。把地址栏里的地址带回来，点「从剪贴板粘贴」即可。
+
+已登录但没有可用套餐的来源（比如 Kimi 账号没有开通 Kimi Code），dsh-model 会实测出来，把它的模型从 dsh 里隐藏，并给出开通链接。
+
+> 更新 dsh-model 后，要重启一次 dsh，管理页的新代码才会生效。模型和开关的变化不需要重启。
 
-- 引擎：[CLIProxyAPI](https://github.com/router-for-me/CLIProxyAPI)（Go，MIT），上游 OAuth 直连 Codex / Claude Code / Grok Build / Muse / Gemini / Kimi 等；
-- dsh-model 自身只做安装编排、dsh 接线（`connect-dsh` / 干净卸载）、已装探测、海外中继产品化、免费模型池；
-- 不自建界面——模型选择与使用全部发生在 dsh 的 Web UI（VPS 场景配合 [dsh-vps](https://github.com/AIcivilization/dsh-vps)）。
+## 来源与费用
 
-完整设计见 [docs/DESIGN.md](docs/DESIGN.md)。
+| 来源 | 费用 | 说明 |
+|---|---|---|
+| WorkBuddy / WorkBuddy AI | 免费 | 账号自带积分 |
+| Codex | 需付费 | ChatGPT Plus / Pro |
+| Kimi | 需付费 | Kimi Code 套餐 |
+| Grok | 需付费 | SuperGrok / X Premium+ |
+| Muse | 需付费 | Muse Code 套餐 |
+| Devin | 需付费 | 免费档实测调用会被拒 |
+| OpenCode Zen | 需付费 | 充值后按量计费；它的免费模型只能在 OpenCode 软件里用 |
+| Claude、Antigravity | 高风险，默认隐藏 | 服务商有封禁第三方使用的先例 |
 
-## 命令
+dsh-model 只负责转接，不提供额度。
+
+## 常用命令
 
 | 命令 | 作用 |
 |---|---|
-| `dsh-model setup` | 安装引擎、内置免费模型、接入 dsh |
-| `dsh-model connect-dsh` / `disconnect-dsh` | 接线 / 干净还原 dsh 配置 |
-| `dsh-model login <cli>` | device-code 登录各家订阅 |
-| `dsh-model relay enable` | 海外中继模式（公网 + token） |
-| `dsh-model status` / `doctor` | 已装/登录状态 / 自检 |
-| `dsh-model uninstall` | 干净卸载，不留痕 |
+| `setup` | 安装 / 修复全部（可重复执行） |
+| `sources` | 来源列表：开关、费用、登录状态、用量 |
+| `source enable\|disable\|logout <来源>` | 打开（需要时登录）/ 关闭 / 退出登录 |
+| `key list\|add\|revoke\|rotate <名称>` | 访问 key，每台设备一把 |
+| `stats` | 按 key、来源、模型的成功率、延迟和速度 |
+| `status` / `doctor [--e2e]` | 总览 / 逐项自检 |
+| `models [sync]` | 列出模型 / 同步到 dsh |
+| `remote enable --via caddy\|ssh\|tailscale` / `remote disable` | 远程访问（VPS 上 setup 已自动开通 caddy） |
+| `engine version\|upgrade\|rollback` | 引擎版本 |
+| `logs [--bridge]` / `repair` | 日志 / 修复 |
+| `update [--to 版本]` | 更新到 npm 上的最新版：程序、dsh 插件和服务一起更新 |
+| `uninstall` | 干净卸载 |
+
+所有命令都支持 `--lang zh|en` 和 `--json`。
+
+## 架构
+
+```
+dsh · 编辑器 · 脚本 ──Bearer key──▶ 引擎 127.0.0.1:8317/v1（唯一入口）
+                                      │      ▲ VPS：Caddy https://<dsh 域名>:9443
+             ┌────────────────────────┼────────────────────────┐
+             ▼                        ▼                        ▼
+   订阅 OAuth（Codex、Kimi…）   opencode/*（你的 key）   workbuddy/* → dsh-model 的 bridge
+```
 
-## 状态
+- 引擎是 [CLIProxyAPI](https://github.com/router-for-me/CLIProxyAPI)（MIT），版本锁定并校验 sha256。WorkBuddy 协议层移植自 [dsh-workbuddy-connect](https://github.com/corrinehu/dsh-workbuddy-connect)（MIT），订阅用量的查询接口参考了 [CodexBar](https://github.com/steipete/CodexBar)（MIT）。
+- dsh 里只写两样东西：模型 provider 和 `refs.DSH_MODEL_API_KEY`。卸载时，如果这两处没被改过，就逐字节还原。
 
-开发中（M0 验证阶段）。平台：macOS / Linux（VPS）优先，Windows 延后。
+详见 [设计文档](docs/DESIGN.md)。
 
-## 免责声明
+## 注意
 
-个人学习研究用途。dsh-model 是将你本人已授权的 CLI 订阅封装为本地 API 的工具，请遵守各家服务商的服务条款，自行承担账号风险。不支持也不鼓励多账号、配额规避等行为。
+- 订阅和 WorkBuddy 的凭据在官方客户端之外使用，可能不符合服务商的条款，账号有风险，请自行判断。
+- 支持 macOS 和 Linux，需要 Node ≥ 20。
 
 ## License
 

package/bin/dsh-model.js

@@ -0,0 +1,8 @@
+#!/usr/bin/env node
+import('../dist/cli.js').then((m) => m.main(process.argv.slice(2))).then(
+  (code) => process.exit(code ?? 0),
+  (error) => {
+    console.error(error?.stack ?? String(error))
+    process.exit(1)
+  },
+)

package/dist/bridge/aggregate.js

@@ -0,0 +1,91 @@
+// bridge/aggregate.ts — 把 OpenAI chat.completion.chunk 的 SSE 流拼成一个非流式 chat.completion
+//
+// WorkBuddy 上游只给流式（prepareChatBody 强制 stream:true），而引擎对非流式客户端会发 stream:false，
+// 期待一个完整的 JSON。这里按 OpenAI 的增量规则累积：content / reasoning_content 拼接，
+// tool_calls 按 index 合并 id、name 与 arguments 片段，finish_reason 与 usage 取最后出现的值。
+export function aggregateSse(text) {
+    let id = '';
+    let model = '';
+    let created = Math.floor(Date.now() / 1000);
+    let usage;
+    const choices = new Map();
+    for (const rawLine of text.split('\n')) {
+        const line = rawLine.trim();
+        if (!line.startsWith('data:'))
+            continue;
+        const data = line.slice(5).trim();
+        if (!data || data === '[DONE]')
+            continue;
+        let chunk;
+        try {
+            chunk = JSON.parse(data);
+        }
+        catch {
+            continue;
+        }
+        if (typeof chunk.id === 'string' && chunk.id)
+            id = chunk.id;
+        if (typeof chunk.model === 'string' && chunk.model)
+            model = chunk.model;
+        if (typeof chunk.created === 'number')
+            created = chunk.created;
+        if (chunk.usage && typeof chunk.usage === 'object')
+            usage = chunk.usage;
+        for (const c of (Array.isArray(chunk.choices) ? chunk.choices : [])) {
+            const index = typeof c.index === 'number' ? c.index : 0;
+            let acc = choices.get(index);
+            if (!acc) {
+                acc = { index, content: '', reasoning: '', toolCalls: new Map(), finishReason: null, role: 'assistant' };
+                choices.set(index, acc);
+            }
+            const delta = (c.delta ?? c.message ?? {});
+            if (typeof delta.role === 'string')
+                acc.role = delta.role;
+            if (typeof delta.content === 'string')
+                acc.content += delta.content;
+            const reasoning = delta.reasoning_content ?? delta.reasoning;
+            if (typeof reasoning === 'string')
+                acc.reasoning += reasoning;
+            for (const tc of (Array.isArray(delta.tool_calls) ? delta.tool_calls : [])) {
+                const ti = typeof tc.index === 'number' ? tc.index : acc.toolCalls.size;
+                let t = acc.toolCalls.get(ti);
+                if (!t) {
+                    t = { type: 'function', function: { name: '', arguments: '' } };
+                    acc.toolCalls.set(ti, t);
+                }
+                if (typeof tc.id === 'string' && tc.id)
+                    t.id = tc.id;
+                const fn = (tc.function ?? {});
+                if (typeof fn.name === 'string')
+                    t.function.name += fn.name;
+                if (typeof fn.arguments === 'string')
+                    t.function.arguments += fn.arguments;
+            }
+            if (typeof c.finish_reason === 'string' && c.finish_reason)
+                acc.finishReason = c.finish_reason;
+        }
+    }
+    const ordered = [...choices.values()].sort((a, b) => a.index - b.index);
+    if (!ordered.length)
+        ordered.push({ index: 0, content: '', reasoning: '', toolCalls: new Map(), finishReason: 'stop', role: 'assistant' });
+    return {
+        id: id || `chatcmpl-bridge-${Date.now()}`,
+        object: 'chat.completion',
+        created,
+        model,
+        choices: ordered.map((c) => {
+            const tools = [...c.toolCalls.entries()].sort(([a], [b]) => a - b).map(([i, t]) => ({ ...t, id: t.id ?? `call_${i}` }));
+            return {
+                index: c.index,
+                message: {
+                    role: c.role,
+                    content: c.content || (tools.length ? null : ''),
+                    ...(c.reasoning ? { reasoning_content: c.reasoning } : {}),
+                    ...(tools.length ? { tool_calls: tools } : {}),
+                },
+                finish_reason: c.finishReason ?? (tools.length ? 'tool_calls' : 'stop'),
+            };
+        }),
+        ...(usage ? { usage } : {}),
+    };
+}

package/dist/bridge/login.js

@@ -0,0 +1,127 @@
+// bridge/login.ts — dsh-model 自己完成 WorkBuddy 登录（不需要桌面 App，也不需要 CodeBuddy CLI）
+//
+// 流程与 WorkBuddy 内置 CLI 的 cli-external-link 相同（读其 product.json 与登录源码确认，2026-10-08）：
+//   POST {base}/v2/plugin/auth/state?platform=<workbuddy|workbuddy-ai>  → { state, authUrl }
+//   用户在任意设备的浏览器打开 authUrl 并授权（服务器上不需要浏览器，也没有本地回调）
+//   轮询 GET {base}/v2/plugin/auth/token?state=…                         → 令牌
+//   GET {base}/v2/plugin/login/account?state=…（Bearer）                 → 账号（uid / 企业 / 昵称）
+// 令牌按移植凭据库的"自有副本"格式写进 $DSH_MODEL_HOME/workbuddy/<ownFilename>（0600），bridge 直接用。
+// 网络请求走 curl：它认 -x 代理（Node 的 fetch 不走代理）；国内版域名强制直连。
+import { join } from 'node:path';
+import { DshModelError } from '../errors.js';
+import { L } from '../i18n.js';
+import { run } from '../util/exec.js';
+import { atomicWrite, ensureDir } from '../util/fs.js';
+import { bridgeDir } from './runtime.js';
+export const LOGIN_ENDPOINTS = {
+    workbuddy: { base: 'https://www.workbuddy.cn', platform: 'workbuddy' },
+    'workbuddy-ai': { base: 'https://www.workbuddy.ai', platform: 'workbuddy-ai' },
+};
+const NO_AUTH_HEADERS = ['X-No-Authorization: true', 'X-No-User-Id: true', 'X-No-Enterprise-Id: true', 'X-No-Department-Info: true'];
+const POLL_INTERVAL_MS = 2000;
+/** 一次 JSON 请求。headers 经 stdin 的 curl 配置传入，令牌不出现在进程参数里 */
+async function curlJson(method, url, headers, proxy) {
+    const direct = /\.workbuddy\.cn$|\.codebuddy\.cn$|\.tencent\.com$|^127\.0\.0\.1$|^localhost$/.test(new URL(url).hostname);
+    const cfg = [...headers, 'Content-Type: application/json', 'Accept: application/json'].map((h) => `header = "${h.replace(/"/g, '')}"`).join('\n') + '\n';
+    const args = ['-s', '-m', '20', '-w', '\n%{http_code}', '-X', method, '-K', '-', ...(proxy && !direct ? ['-x', proxy] : ['--noproxy', '*']), ...(method === 'POST' ? ['-d', '{}'] : []), url];
+    const r = await run('curl', args, { input: cfg, timeoutMs: 30_000 });
+    if (r.code !== 0 && !r.stdout) {
+        throw new DshModelError('workbuddy_network', L(`连不上 ${new URL(url).host}（curl 退出码 ${r.code}）${proxy ? '' : '；海外 / 国际版可能需要代理：dsh-model setup --proxy <地址>'}`, `Cannot reach ${new URL(url).host} (curl exit ${r.code})${proxy ? '' : '; the international site may need a proxy: dsh-model setup --proxy <url>'}`));
+    }
+    const lines = r.stdout.trimEnd().split('\n');
+    const status = Number(lines.pop()) || 0;
+    let body = {};
+    try {
+        body = JSON.parse(lines.join('\n'));
+    }
+    catch {
+        // 非 JSON（网关错误页等）
+    }
+    return { status, body };
+}
+const str = (v) => (typeof v === 'string' && v ? v : undefined);
+const num = (v) => (typeof v === 'number' && Number.isFinite(v) ? v : undefined);
+/** 秒或毫秒的时间戳统一成毫秒 */
+const toMs = (v) => (v < 1e12 ? v * 1000 : v);
+export async function workbuddyLogin(variant, home, opts) {
+    const known = LOGIN_ENDPOINTS[variant.id];
+    if (!known)
+        throw new DshModelError('workbuddy_unknown_variant', `unknown WorkBuddy product: ${variant.id}`);
+    // 测试用：DSH_MODEL_WORKBUDDY_BASE 指向本机假服务
+    const ep = process.env.DSH_MODEL_WORKBUDDY_BASE ? { ...known, base: process.env.DSH_MODEL_WORKBUDDY_BASE } : known;
+    const st = await curlJson('POST', `${ep.base}/v2/plugin/auth/state?platform=${ep.platform}`, NO_AUTH_HEADERS, opts.proxy);
+    const state = str(st.body.data?.state);
+    const authUrl = str(st.body.data?.authUrl);
+    if (!state || !authUrl) {
+        throw new DshModelError('workbuddy_auth_state_failed', L(`${variant.displayName} 没有返回授权链接（HTTP ${st.status}${st.body.msg ? `：${st.body.msg}` : ''}）`, `${variant.displayName} returned no authorization link (HTTP ${st.status}${st.body.msg ? `: ${st.body.msg}` : ''})`));
+    }
+    opts.onUrl(authUrl);
+    // 轮询令牌：授权前服务端返回"等待中"，授权后返回令牌
+    const deadline = Date.now() + (opts.timeoutMs ?? 10 * 60_000);
+    let token;
+    while (Date.now() < deadline) {
+        if (opts.signal?.aborted)
+            throw new DshModelError('workbuddy_login_cancelled', L('已取消登录', 'Login cancelled'));
+        await new Promise((r) => setTimeout(r, POLL_INTERVAL_MS));
+        let res;
+        try {
+            res = await curlJson('GET', `${ep.base}/v2/plugin/auth/token?state=${encodeURIComponent(state)}`, NO_AUTH_HEADERS, opts.proxy);
+        }
+        catch {
+            continue; // 偶发网络抖动：继续等
+        }
+        if (str(res.body.data?.accessToken)) {
+            token = res.body.data;
+            break;
+        }
+    }
+    if (!token)
+        throw new DshModelError('workbuddy_login_timeout', L('等待授权超时（10 分钟）', 'Timed out waiting for authorization (10 minutes)'), L('重新执行 dsh-model workbuddy login', 'Run dsh-model workbuddy login again'));
+    const accessToken = str(token.accessToken);
+    // 账号信息：uid / 企业 / 昵称决定请求头（X-User-Id 等），拿不到就用 X-No-* 头
+    let account = {};
+    for (let i = 0; i < 5; i++) {
+        try {
+            const a = await curlJson('GET', `${ep.base}/v2/plugin/login/account?state=${encodeURIComponent(state)}`, [`Authorization: Bearer ${accessToken}`, 'X-No-User-Id: true', 'X-No-Enterprise-Id: true', 'X-No-Department-Info: true'], opts.proxy);
+            if (a.body.data && typeof a.body.data === 'object') {
+                account = a.body.data;
+                break;
+            }
+        }
+        catch {
+            // 重试
+        }
+        await new Promise((r) => setTimeout(r, POLL_INTERVAL_MS));
+    }
+    const expiresIn = num(token.expiresIn);
+    const expiresAt = num(token.expiresAt);
+    const refreshExpiresAt = num(token.refreshExpiresAt);
+    const enterpriseId = str(account.enterpriseId) ?? str(token.enterpriseId);
+    const nickname = str(account.nickname) ?? str(account.userNickname) ?? str(account.name);
+    const credential = {
+        accessToken,
+        refreshToken: str(token.refreshToken) ?? '',
+        expiresAtMs: expiresAt ? toMs(expiresAt) : expiresIn ? Date.now() + expiresIn * 1000 : 0,
+        ...(refreshExpiresAt ? { refreshExpiresAtMs: toMs(refreshExpiresAt) } : {}),
+        // 凭据库按 domain 判断国内 / 国际：令牌没给就用登录站点的域名
+        domain: str(token.domain) ?? str(account.domain) ?? new URL(ep.base).hostname,
+        uid: str(account.uid) ?? str(account.userId) ?? str(token.uid) ?? '',
+        ...(enterpriseId ? { enterpriseId } : {}),
+        ...(nickname ? { nickname } : {}),
+        source: 'dsh',
+    };
+    const dir = bridgeDir(home);
+    await ensureDir(dir, { owner: opts.owner });
+    await atomicWrite(join(dir, variant.ownFilename), JSON.stringify({ version: 1, credential }, null, 2) + '\n', { mode: 0o600, owner: opts.owner });
+    return credential;
+}
+/** 退出登录：删掉 dsh-model 保存的那份（App 的文件不动） */
+export async function workbuddyLogout(variant, home) {
+    const { rm } = await import('node:fs/promises');
+    const { exists } = await import('../util/fs.js');
+    const file = join(bridgeDir(home), variant.ownFilename);
+    if (!(await exists(file)))
+        return false;
+    await rm(file, { force: true });
+    return true;
+}