dsh-trae-connect

1. 安装并启动 DeepSeek Harness：npx @deepseek-ai/dsh web
2. 在终端执行上面的安装命令（CLI 会解析插件并核验来源）
3. 用 dsh plugins list 确认已安装，必要时重启 Harness 生效

插件以当前 dsh 进程的权限运行，安装时可能执行代码。请先通读仓库源码与许可证，确认无破坏性命令与越权访问；本站只做索引，不对第三方插件安全性作担保。

事实信息来自公开插件目录快照（2026-10-01），介绍文案由本站再加工。

以下为插件仓库 README 全文（原始内容，由公开目录抓取整理）。

# DSH Trae Connect

English | [简体中文](./README.zh.md)

Plug the models built into the [Trae](https://www.trae.cn/) desktop app (ByteDance's AI coding assistant) — DeepSeek V3/R1, GPT-4o, GPT-4.1, Claude 3.5/3.7 Sonnet, Gemini 2.5 and more — into DeepSeek Harness (DSH) as a zero-configuration model provider. It follows the bridging architecture of [dsh-workbuddy-connect](https://github.com/corrinehu/dsh-workbuddy-connect): a loopback OpenAI-compatible endpoint + a pi-ai provider + a browser settings card.

The upstream protocol was reverse-engineered from [trae2api](https://github.com/linqiu919/trae2api) (MIT), which translates OpenAI chat-completions into Trae's private IDE API.

## Features

- **Zero-configuration credentials.** The plugin reads Trae's desktop sign-in state directly; there is nothing to paste into DSH.
- **Automatic model catalog.** The available Trae models are fetched from the upstream at startup and refreshed periodically; they show up in DSH's model picker under the `trae` provider (for example `deepseek-v3`, `deepseek-r1`, `gpt-4o`, `claude-3-7-sonnet`).
- **OpenAI-compatible loopback shim.** A local endpoint on `127.0.0.1` translates Trae's private SSE stream into OpenAI chat-completions SSE in real time, including reasoning content. The shim is protected by four independent loopback checks and is never exposed beyond localhost.
- **Two credential sources with automatic preference.** Desktop session token first, self-renewing OAuth configuration second (details below).
- **Stable device identity.** Device headers (`x-device-id`, `x-machine-id`, ...) are generated once and persisted, so the upstream sees a consistent client instead of a new device per request.
- **Settings card in the browser.** A card under Settings -> Plugins (`settings.trae`) shows the current sign-in state and the number of available models, and tolerates registration failures of either half.
- **Web status routes.** The host half exposes status/capture/checkin routes under its settings namespace, which the companion plugin [dsh-points-checkin](https://github.com/lament-z/dsh-points-checkin) reuses for TRAE points and daily check-in.
- **CLI diagnostics.** `dsh-trae-connect` ships subcommands `doctor`, `status`, `logout`, and `setup` for checking credential health without opening the browser.

## Requirements

- DSH `>=0.1.1-rc.1` (built and tested against `0.1.5-rc.2`).
- The Trae desktop app signed in at least once (for the desktop-session credential source), or a Trae OAuth refresh token (for the auto-renewing source).

## Install

```sh
# Preferred: install straight from GitHub (no npm involved; prebuilt lib is committed)
dsh plugin --profile web add github:lament-z/dsh-trae-connect

# Alternative: from npm
dsh plugin --profile web add @lament_z/dsh-trae-connect

# From a local clone / working copy
dsh plugin --profile web add link:
```

Then restart `dsh web` and reload the page. The first start also compiles nothing and asks for nothing — the plugin activates with the host.

## Credentials (two sources, automatic preference)

1. **Desktop session (zero configuration).** Trae stores its `x-ide-token` in the Electron `Local Storage/leveldb`. Place an exported `/trae-auth.json` in the Trae app directory and the plugin picks it up automatically (fields: `ideToken` / `appId` / `clientId` / `userId` / `device` / `expiresAt`). This source takes priority over the OAuth copy below.
   - macOS app directory candidates: `~/Library/Application Support/TRAE SOLO CN` and `Trae`.
   - Override the directory with the `TRAE_APP_DIR` environment variable.
2. **OAuth configuration (auto-renewing).** Set the environment variables `TRAE_APP_ID` / `TRAE_CLIENT_ID` / `TRAE_REFRESH_TOKEN` / `TRAE_USER_ID`. The plugin exchanges the refresh token at Trae's OAuth endpoint and rotates `x-ide-token` on its own, without reopening the desktop app.

> The desktop session token cannot be renewed by itself — when it expires, reopen Trae once. The OAuth configuration source self-heals and is recommended for long-running hosts.

## Usage

1. Install and restart `dsh web` (see above).
2. Open the DSH web page and check Settings -> Plugins: the `trae` card should show a signed-in state and a model count.
3. Open the model picker: Trae models appear under the `trae` provider and can be used in any conversation like built-in models.
4. Optional diagnostics from a terminal:

```sh
dsh-trae-connect doctor   # check credential health and upstream reachability
dsh-trae-connect status   # current credential source, expiry, model count
dsh-trae-connect logout   # drop the stored desktop-session copy
dsh-trae-connect setup    # guided setup of the OAuth environment variables
```

## Upstream protocol (implemented)

| Purpose | Endpoint |
|---|---|
| Chat | `https://a0ai-api-sg.byteintlapi.com/api/ide/v1/chat` (POST) |
| Model catalog | `https://a0ai-api-sg.byteintlapi.com/api/ide/v1/model_list?type=chat` (GET) |
| Token exchange | `https://api-sg-central.trae.ai/cloudide/api/v3/trae/oauth/ExchangeToken` (POST) |

Every upstream request carries the Trae client headers: `x-app-id`, `x-ide-version` (1.2.10), `x-ide-version-code` (20250325), `x-ide-version-type` (stable), `x-device-cpu`, `x-device-id`, `x-machine-id`, `x-device-brand`, `x-device-type`, `x-ide-token` (auth), `Host`, and an empty `User-Agent`.

The upstream returns Trae's private SSE events (`output` carrying `response`/`reasoning_content`, `request_wait_in_queue` for queueing); the shim in `src/upstream.ts` rewrites them into OpenAI chat-completions SSE on the fly for the pi-ai provider. Model-name mapping (OpenAI id <-> Trae upstream name) lives in the same file, for example `deepseek-v3` <-> `deepseek-V3`, `claude-3-7-sonnet` <-> `aws_sdk_claude37_sonnet`.

## Billing

Trae has no public balance endpoint (same as workbuddy-connect), so no points are shown — only the number of available models. For TRAE points and daily check-in, use [dsh-points-checkin](https://github.com/lament-z/dsh-points-checkin).

## Architecture

- **Host** (`src/index.ts`): starts the loopback shim, registers the `trae` provider, installs the settings section, and refreshes the model catalog.
- **Shim** (`src/shim.ts`): the OpenAI-compatible endpoint bound to `127.0.0.1` with four-way loopback authentication.
- **Client** (`src/client/index.tsx`): the settings card under Settings -> Plugins showing sign-in state and model count.
- **CLI** (`src/bin.ts`): `doctor` / `status` / `logout` / `setup`.
- **Web status** (`src/web-status.ts`): status/capture/checkin routes reused by dsh-points-checkin.

## Troubleshooting

- **No models appear.** Run `dsh-trae-connect doctor`. If the desktop session expired, reopen the Trae desktop app and restart `dsh web`; or configure the OAuth environment variables for self-renewal.
- **The settings card is missing.** Both halves register independently; check the DSH boot log for `dsh-trae-connect` errors. A failed card never blocks the provider.
- **Credential errors right after Trae re-login.** The exported `trae-auth.json` is stale — re-export it or switch to the OAuth source.

## License

MIT