dsh-model-connect · 模型接入
中文 | English
给 DeepSeek Harness（下称 DSH）Web 界面左侧边栏加一个「模型接入」面板：
在图形界面里填写第三方 OpenAI / Anthropic 兼容 端点与 API Key，探测该端点提供的模型列表，
保存成 dsh-llm-pi-ai 的 provider 路由，随后就能在模型选择器里直接使用。
A sidebar panel for the DeepSeek Harness Web UI: fill in a third-party
OpenAI / Anthropic compatible endpoint and its API key, probe the endpoint's
model catalog, and save it as a dsh-llm-pi-ai provider route — no YAML editing,
no cordis.yml surgery, no restart.
中文
不需要改 cordis.yml、不需要手写 providers YAML、不需要重启 DSH。
功能
- 侧边栏入口：左侧
model-connect 图标 → 主区域整页面板（与内置 Plugins / Schedule / SSH 行同一套 shell 插槽）。
- 协议支持：
openai-completions（/chat/completions）、openai-responses（/responses）、anthropic-messages（/v1/messages）。
- 模型探测：填好 Base URL 与 Key，点「探测模型」即调用 LLM 服务的
discoverModels 拿回模型清单，自动回填模型 ID。
- 路由保存：写入
llm-pi-ai 命名空间的 providers.<route>，规则与官方 Models 设置页完全一致（providers 是易变字段）。
- Key 安全：API Key 只写进
credentials 服务，引用名形如 <ROUTE>_API_KEY，不会落进 cordis.patch.yml。
- 模型能力：保存时统一写入每个模型条目的上下文窗口、最大输出、输入模态与推理强度（默认
1M / 128K / text+image / high），不再吃适配器兜底值。
- 编辑 / 删除：已有路由可原地改写 Base URL、协议、模型集合；删除时可选择是否连带清除凭据。
- 原生观感：全部使用
--dsw-alias-* 设计令牌，跟随主题与深浅色；中英文按当前 locale 自动切换。
目录结构
model-connect/
├── index.js # Host 半：loopback 路由（state / probe / save / remove）+ settings/credentials 写入
├── client.js # 浏览器半：侧边栏行 + 主区域页面（纯 JS，无构建步骤，React 由模块表提供）
├── cordis.patch.yml # bundle 补丁：把自己 insert 进 Loader 组合
├── package.json # 清单：exports / files / dsh.bundle / dsh.client.inject
├── locale/
│ ├── zh.json # 面板元信息（中文）
│ └── en.json # 面板元信息（英文）
├── icon.svg # 侧边栏图标
└── tools/push-api.ps1 # 无 git 二进制时的 REST 发布脚本
安装
方式一：本地链接（开发 / 自用）
git clone https://github.com/ilmilm/dsh-model-connect.git C:\path\to\model-connect
dsh plugin --profile desktop add C:\path\to\model-connect
然后在 ~/.dsh/profiles/desktop/package.json 的 dsh.profile.bundles 数组里追加一项：
"@local/dsh-model-connect"
重启 DSH（或重新加载 profile）后，左侧边栏出现「模型接入」。
方式二：手动登记
- 把仓库放到任意目录；
- 编辑
~/.dsh/profiles/<profile>/package.json：
{
"dependencies": {
"@local/dsh-model-connect": "link:C:/path/to/model-connect"
},
"dsh": {
"profile": {
"bundles": ["...", "@local/dsh-model-connect"]
}
}
}
- 在 profile 目录执行
pnpm install，重启 DSH。
包名保持 @local/dsh-model-connect（不是发布到 npm 的包），@local/ 作用域配合 link: 依赖即可解析。
使用
- 左侧点「模型接入」；
- 填 提供方 ID（如
myrelay，只能字母数字点下划线连字符）、显示名、Base URL（如 https://api.example.com/v1）、协议；
- 填 API Key，点「探测模型」；
- 勾选要启用的模型 → 保存；
- 保存后该路由在原 Models 设置页与模型选择器里立即可见。
模型能力默认值
| 面板字段 | 默认 | 写入 providers.<route>.models[] |
|---|
| 上下文窗口 | 1M（1000000） | contextWindow |
| 最大输出 | 128K（128000） | maxTokens |
| 输入模态 | text + image 都勾选 | input: ['text', 'image'] |
| 推理强度 | high | reasoningEfforts: { high: 'high' } |
K / M 按 1000 进制，与官方 Models 页一致：1M = 1000000，128K = 128000。- 推理强度选「不启用」→ 不写
reasoningEfforts，模型不推理，选择器里也不出现强度档。
- 保存时这四项统一覆盖该路由下所有模型；编辑已有路由时用第一个模型的值回填表单。
- 字段留空按上面的默认值写入，不会退回适配器兜底。
Host HTTP 路由
所有特权操作都在 Host 半，浏览器半通过下面四条同源路由访问（仅回环地址可调）：
| 方法 | 路径 | 作用 |
|---|
GET | /model-connect/api/state | 当前路由、模型清单、Key 配置状态、命名空间诊断 |
POST | /model-connect/api/probe | 用草稿端点 + Key 探测模型列表 |
POST | /model-connect/api/save | 新增 / 更新一条 provider 路由（并写 Key） |
POST | /model-connect/api/remove | 删除一条路由（可选连带删除 Key） |
安全边界
- 仅回环：
socket.remoteAddress 必须是 127.0.0.0/8 或 ::1，Host/Origin 必须同源，Sec-Fetch-Site: cross-site 直接 403；不信任 X-Forwarded-For。
- 请求体上限 256 KiB，非 JSON 直接 400。
- Key 不落配置：
cordis.patch.yml 只出现 <ROUTE>_API_KEY 引用名；值存在 credentials 服务里。
- 参数校验：route ID 走
^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$；Base URL 必须是 http(s)；Key 只允许可见 ASCII。
兼容性
- DSH
0.2.x（开发验证于 0.2.1-rc.2）。
- 依赖 Host 服务：
webServer、settings、credentials、llm；对应插件需在组合中启用。
- 浏览器半无构建步骤（无 JSX / 无 bundler），React 从浏览器模块表读取。
English
A sidebar entry plus a full-page panel inside the DeepSeek Harness (DSH) Web UI
that registers third-party OpenAI / Anthropic compatible model endpoints as
first-class dsh-llm-pi-ai providers. Everything the official Models settings
page does, but reachable from the left rail and with a live model probe.
Features
- Sidebar entry — a
model-connect row on the left rail opens a page in the layout's keyed main slot, using the same shell-owned pattern as the built-in Plugins / Schedule / SSH rows.
- Three protocols —
openai-completions (/chat/completions), openai-responses (/responses), anthropic-messages (/v1/messages).
- Live model probe — enter a base URL and key, hit Probe, and the panel calls the LLM service's
discoverModels to populate the model list.
- Route persistence — writes
providers.<route> in the llm-pi-ai settings namespace, following the exact rules of the shipped Models page (providers is a volatile field).
- Key handling — the API key is written to the
credentials service only, under a <ROUTE>_API_KEY reference; it never lands in cordis.patch.yml.
- Model capabilities — every saved model entry gets an explicit context window, max output, input modalities and reasoning level (defaults
1M / 128K / text+image / high) instead of falling back to adapter-level defaults.
- Edit / remove — existing routes can be updated in place (base URL, protocol, model set); removal can optionally delete the stored credential.
- Native look — every colour comes from
--dsw-alias-* design tokens, so it follows the active theme; copy switches between Chinese and English by locale.
Layout
model-connect/
├── index.js # host half: loopback routes (state/probe/save/remove) + settings & credentials writes
├── client.js # browser half: sidebar row + main-slot page (plain JS, no build step, React from the module table)
├── cordis.patch.yml # bundle patch: inserts itself into the Loader composition
├── package.json # manifest: exports / files / dsh.bundle / dsh.client.inject
├── locale/
│ ├── zh.json # panel metadata (Chinese)
│ └── en.json # panel metadata (English)
├── icon.svg # sidebar icon
└── tools/push-api.ps1 # REST publishing helper for machines without a git binary
Install
Option 1 — local link (development / personal use)
git clone https://github.com/ilmilm/dsh-model-connect.git C:\path\to\model-connect
dsh plugin --profile desktop add C:\path\to\model-connect
Then append the package to the dsh.profile.bundles array in
~/.dsh/profiles/desktop/package.json:
"@local/dsh-model-connect"
Restart DSH (or reload the profile); the Model Connect row appears on the left rail.
Option 2 — manual registration
- Put the repository anywhere on disk.
- Edit
~/.dsh/profiles/<profile>/package.json:
{
"dependencies": {
"@local/dsh-model-connect": "link:C:/path/to/model-connect"
},
"dsh": {
"profile": {
"bundles": ["...", "@local/dsh-model-connect"]
}
}
}
- Run
pnpm install inside the profile directory and restart DSH.
The package keeps the name @local/dsh-model-connect (it is not published to npm);
the @local/ scope resolves through the link: dependency.
Usage
- Click Model Connect on the left rail.
- Fill in provider ID (e.g.
myrelay — letters, digits, dot, underscore, hyphen), display name, base URL (e.g. https://api.example.com/v1) and protocol.
- Enter the API key and click Probe models.
- Tick the models to enable, then save.
- The route shows up immediately in the shipped Models page and the model picker.
Capability defaults
| Panel field | Default | Written into providers.<route>.models[] |
|---|
| Context window | 1M (1000000) | contextWindow |
| Max output | 128K (128000) | maxTokens |
| Input modalities | text + image ticked | input: ['text', 'image'] |
| Reasoning level | high | reasoningEfforts: { high: 'high' } |
K / M are decimal, same as the shipped Models page: 1M = 1000000, 128K = 128000.- Reasoning "off" omits
reasoningEfforts entirely — the model does not reason and the picker offers no level.
- On save these four values overwrite every model of that route; editing an existing route prefills the form from its first stored model.
- A blank field falls back to the default above rather than to the adapter's own fallback.
Host routes
Every privileged operation lives in the host half. The browser half reaches it
through four same-origin routes (loopback only):
| Method | Path | Purpose |
|---|
GET | /model-connect/api/state | Current routes, model lists, key state, namespace diagnostics |
POST | /model-connect/api/probe | Probe a draft endpoint + key for its model catalog |
POST | /model-connect/api/save | Upsert one provider route (and its key) |
POST | /model-connect/api/remove | Remove one route (optionally its credential) |
Security boundaries
- Loopback only —
socket.remoteAddress must be 127.0.0.0/8 or ::1; Host and Origin must be same-origin; Sec-Fetch-Site: cross-site is rejected with 403. X-Forwarded-For is never trusted.
- Body cap — 256 KiB, non-JSON bodies are rejected with 400.
- Keys never enter config —
cordis.patch.yml only ever holds a <ROUTE>_API_KEY reference; the value lives in the credentials service.
- Validation — route ID must match
^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$; base URL must be http(s); keys are limited to visible ASCII.
Compatibility
- DSH
0.2.x (developed and verified against 0.2.1-rc.2).
- Requires the host services
webServer, settings, credentials, llm to be enabled in the composition.
- The browser half needs no build step (no JSX, no bundler); React is read from the browser module table.
License
MIT — see LICENSE.
