dsh-relay-fast

Readme

# dsh-relay-fast

[中文](#中文) · [English](#english)

## 中文

面向 **DeepSeek Harness（DSH）Desktop/Web** 的中转站增强插件。

### 功能

- **思考等级自动补全**：为 `llm-pi-ai` 新增模型补上 `reasoningEfforts`；兼容 `openai-completions` 与 `anthropic-messages`，不覆盖已有声明。
- **模型同步**：在设置页从中转站 `/models` 同步缺失模型。
- **动态 Fast 开关**：仅在端点明确声明 Fast，或用户通过 `fastModels` 显式启用时显示；开启后发送 `service_tier: "priority"`。

### 安装

```sh
dsh plugin --profile desktop add github:Rock-ql/dsh-relay-fast
```

安装后重启 DSH Desktop。Web profile 可将 `desktop` 替换为对应 profile 名称。

### 可选配置

在 profile 的 `cordis.patch.yml` 中覆盖插件配置：

```yaml
- id: relay-fast
  config:
    autofill: true
    fastTier: priority
    fastModels:
      - gpt-5*          # 所有中转渠道的 GPT-5 模型
    defaultSpeed: fast  # 新会话默认 Fast；仍可手动切回 Standard
```

不配置 `fastModels` 时，插件只相信中转站 `/models` 返回的明确 Fast 能力元数据，不按模型名称猜测。

> 注意：开关能保证发送 `service_tier: priority`，但中转站可能忽略该字段。真实加速能力取决于上游。测试记录见 [docs](docs/)。

### 开发

```sh
pnpm install
pnpm check
```

---

## English

A relay enhancement plugin for **DeepSeek Harness (DSH) Desktop/Web**.

### Features

- **Reasoning-effort autofill**: adds missing `reasoningEfforts` to new `llm-pi-ai` models. It is protocol-aware for `openai-completions` and `anthropic-messages`, and never overwrites existing declarations.
- **Model sync**: imports missing models from a relay's `/models` endpoint in Settings.
- **Dynamic Fast toggle**: appears only when the endpoint explicitly advertises Fast, or when enabled through `fastModels`. Fast requests send `service_tier: "priority"`.

### Install

```sh
dsh plugin --profile desktop add github:Rock-ql/dsh-relay-fast
```

Restart DSH Desktop after installation. For Web, replace `desktop` with the target profile name.

### Optional configuration

Override the plugin in the profile's `cordis.patch.yml`:

```yaml
- id: relay-fast
  config:
    autofill: true
    fastTier: priority
    fastModels:
      - gpt-5*          # GPT-5 models on every relay route
    defaultSpeed: fast  # Fast by default; users can switch to Standard
```

Without `fastModels`, the plugin trusts only explicit Fast metadata returned by `/models`; it never guesses from model names.

> The toggle guarantees that `service_tier: priority` is sent, but a relay may ignore it. Actual acceleration depends on the upstream service. See [docs](docs/) for test reports.

### Development

```sh
pnpm install
pnpm check
```

## License

[MIT](LICENSE)