English | 中文

English

The Problem

DeepSeek Harness (DSH) is a powerful LLM agent framework, but it has no built-in image generation capability. You can't ask the agent to draw a picture, create a diagram, or generate a visual — it simply doesn't know how.

dsh-draw-router solves this by acting as a universal image generation router. It registers draw_image and draw_list_sources as agent tools, so the DSH agent can generate images just like it runs bash commands — naturally, in conversation.

✨ Features

🧠 How it works

User: "帮我画一张流程图，展示API调用过程"
         │
         ▼
Agent calls draw_image(source="sensenova", prompt="...")
         │
         ▼
Plugin calls GET {baseURL}/v1/models → auto-detect image models
         │
         ▼
Plugin calls POST {baseURL}/v1/images/generations
         │
         ▼
Returns image URL → agent displays it in conversation

Model detection works by name pattern matching — no hardcoded model lists. The plugin recognizes 30+ patterns including image, draw, generation, flux, sd-3, hunyuan, seedream, agnes-image, step-image, qwen-image, wan, imagen, gpt-image, dall, t2i, i2i and more.

Each provider's API quirks are handled automatically (per official docs):

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│                    DeepSeek Harness (DSH)                   │
│                                                             │
│  ┌─────────────────────────────────────────────────────┐    │
│  │              dsh-draw-router (plugin)               │    │
│  │                                                     │    │
│  │  ┌──────────────┐   ┌────────────────────────────┐  │    │
│  │  │ Agent Tools  │   │     REST API (webServer)   │  │    │
│  │  │              │   │                            │  │    │
│  │  │ draw_image   │   │ GET  /dsh-draw-router/     │  │    │
│  │  │ draw_list_   │   │      sources              │  │    │
│  │  │   sources    │   │ POST /dsh-draw-router/     │  │    │
│  │  └──────┬───────┘   │      sources              │  │    │
│  │         │           │ POST /dsh-draw-router/     │  │    │
│  │         │           │      draw                 │  │    │
│  │  ┌──────▼───────────────────────────────┐        │    │
│  │  │        Draw Router Core              │        │    │
│  │  │  - Source registry (baseURL+apiKey)  │        │    │
│  │  │  - Model discovery (GET /v1/models)  │        │    │
│  │  │  - Pattern matching (30+ patterns)   │        │    │
│  │  │  - Per-provider API adaptation       │        │    │
│  │  │  - draw-config.json persistence      │        │    │
│  │  └──────┬───────────────────────────────┘        │    │
│  │         │                                          │    │
│  └─────────┼──────────────────────────────────────────┘    │
│            ▼                                               │
│  ┌────────────────────────────────────────────────┐        │
│  │          Provider Image APIs                    │        │
│  │  SenseNova · StepFun · Agnes · Qwen · Flux ...  │        │
│  └────────────────────────────────────────────────┘        │
└─────────────────────────────────────────────────────────────┘

📦 Installation

Two installation methods are supported.

Method A: File-based installation (推荐 / Recommended)

Clone the repo and copy files:

# 1. Clone
git clone https://github.com/xiaozhe7772222/dsh-draw-router.git
cd dsh-draw-router

# 2. Copy plugin files
mkdir -p ~/.dsh/profiles/web/plugins/dsh-draw-router
cp -r lib package.json cordis.patch.yml ~/.dsh/profiles/web/plugins/dsh-draw-router/

# 3. Link to node_modules (so DSH bundle loader can find it)
ln -sfn ../plugins/dsh-draw-router ~/.dsh/profiles/web/node_modules/dsh-draw-router

# 4. Add to bundles in ~/.dsh/profiles/web/package.json
#    (add "dsh-draw-router" to the bundles array)

Method B: Terminal command installation (终端命令安装)

Install directly via the DSH CLI:

# If you have published the plugin to npm registry:
cd ~/.dsh/profiles/web
dsh plugin add dsh-draw-router --profile web

# Or install from local package:
dsh plugin add /path/to/dsh-draw-router --profile web

After either method, restart DSH:

npx @deepseek-ai/dsh web

Verify it's working:

# Check the HTTP server is up
curl http://127.0.0.1:3080/dsh-draw-router/sources

# Should return: {"sources":{}}

⚙️ Configuration

Option 1: Static config via cordis.patch.yml

- insert:
    - id: draw-router
      name: dsh-draw-router
      inject: [tools, webServer]
      config:
        sources:
          stepfun:
            baseUrl: https://api.stepfun.com/v1
            apiKey: your-stepfun-key
          sensenova:
            baseUrl: https://token.sensenova.cn/v1
            apiKey: your-sensenova-key
          agnes:
            baseUrl: https://apihub.agnes-ai.com/v1
            apiKey: your-agnes-key

baseURL hint: give the API root (e.g. https://api.stepfun.com/v1). The plugin appends /images/generations and /models automatically.

Option 2: REST API (runtime)

# Add a source (auto-probes models)
curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"add","name":"stepfun","baseUrl":"https://api.stepfun.com/v1","apiKey":"sk-xxx"}'

# List sources + detected models
curl http://127.0.0.1:3080/dsh-draw-router/sources

# Add a manual model (for providers not exposing image models via /v1/models)
curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"addModel","name":"tokenrhythm","model":"qwen-image-2.0"}'

# Generate an image
curl -X POST http://127.0.0.1:3080/dsh-draw-router/draw \
  -H 'Content-Type: application/json' \
  -d '{"source":"sensenova","prompt":"a cute cat","size":"2048x2048"}'

🔌 REST API Reference

🛠️ Troubleshooting (故障排查)

Installation issues (安装问题)

1. cannot resolve profile bundle "dsh-draw-router"

running dsh web fails with this during boot.

The bundle loader can't find the plugin. Fix:

# Ensure the symlink exists
ln -sfn ../plugins/dsh-draw-router ~/.dsh/profiles/web/node_modules/dsh-draw-router

# And confirm "dsh-draw-router" is in the bundles array of package.json
# (as a string, NOT as an object)

2. failed to apply loader entry draw-router: unsupported JSON schema

Plugin loads but the agent tool schema fails validation.

This happens if you modified the source. The output.schema must include additionalProperties: false (or true), and optional parameters must not have required: false. Use the shipped lib/index.js as-is.

3. Plugin loads but draw_image tool is not in chat

Agent can't see the draw tools.

The tools service must be available. Check inject: [tools, webServer] in cordis.patch.yml is intact, then restart DSH. Also verify with:

cd ~ && DSH_SKINS_DIR=... npx -y @deepseek-ai/dsh --profile web --dump-config 2>/dev/null | grep -A3 draw-router

Runtime issues (运行问题)

4. Adding a source returns imageModels: []

Auto-detection found no draw models.

Some providers (e.g. TokenRhythm) don't expose image models via /v1/models — only chat models. Fix: add models manually.

curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"addModel","name":"tokenrhythm","model":"qwen-image-2.0"}'

5. Draw fails with HTTP 402 / quota_exceeded

The provider account has no balance.

This is an account billing issue, not a plugin bug. Top up the provider account or switch to another source:

curl -X POST http://127.0.0.1:3080/dsh-draw-router/draw \
  -H 'Content-Type: application/json' \
  -d '{"source":"sensenova","prompt":"a cat","size":"2048x2048"}'

6. Draw fails with field Size invalid

The size value isn't supported by the provider's model.

Different providers support different sizes:

• SenseNova sensenova-u1-fast: 2048x2048, 1664x2496, 2752x1536, etc.
• StepFun: 1024x1024, 768x768, 1280x800
• Agnes: 1K/2K/3K/4K + ratio

If in doubt, omit size and let the provider use its default.

7. Draw fails with HTTP 401 / 403

The API key is invalid or missing.

Check the source's key via the masked value, and re-add with the correct key:

curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"remove","name":"bad-source"}'
# then add again with correct apiKey

8. Requests time out

Large images or slow providers can take 30-120s.

The plugin uses a 120s request timeout. If your provider is slow, retry once — image generation naturally takes time.

9. Settings card not visible in Web UI

The plugin works (REST API responds) but no settings card shows.

Client bundles load from lib/client.js. Make sure:

• package.json has "client": { "platform": "web" }
• Restart DSH with a clean cache (hard refresh browser: Ctrl+Shift+R)
• If it still fails, the plugin's server-side tools still work — configure via REST API instead.

10. draw-config.json corrupted

Persistence file is broken after a crash.

Delete it and restart — sources configured via cordis.patch.yml will reload, and you can re-add runtime sources.

🗂️ Project structure

dsh-draw-router/
├── lib/
│   ├── index.js      # Server: model discovery, draw logic, agent tools, REST API
│   └── client.js     # Client: settings panel card (optional)
├── package.json      # Package metadata + DSH bundle declaration
├── cordis.patch.yml  # Static config example
├── draw-config.json  # Runtime persistence (gitignored)
├── README.md
├── CHANGELOG.md
└── LICENSE

🏷️ Topics

deepseek-harness · dsh · dsh-plugin · draw · image-generation · text-to-image · sensenova · stepfun · agnes · qwen · wanxiang · tongyi · flux · imagen · stable-diffusion · hunyuan · seedream · jimeng · doubao · gpt-image · dall-e · openai · midjourney · kling · cogview

📄 License

MIT License — free to use, modify, distribute. Never commit your real API keys to a public repo.

中文文档

问题

DeepSeek Harness (DSH) 是一个强大的 LLM Agent 框架，但它没有内置的生图能力。你无法让 agent 画图、创建图表或生成视觉内容——它根本不会。

dsh-draw-router 作为一个统一的图像生成路由器解决了这个问题。它向 DSH agent 注册了 draw_image 和 draw_list_sources 工具，让 agent 像执行 bash 命令一样自然地生图——在对话中一句话就够了。

✨ 功能特性

🧠 工作原理

用户: "帮我画一张流程图，展示API调用过程"
         │
         ▼
Agent 调用 draw_image(source="sensenova", prompt="...")
         │
         ▼
插件调用 GET {baseURL}/v1/models → 自动识别绘图模型
         │
         ▼
插件调用 POST {baseURL}/v1/images/generations
         │
         ▼
返回图片 URL → agent 在对话中展示图片

模型识别通过名称关键词匹配——没有硬编码模型列表。插件识别 30+ 种模式，包括 image、draw、generation、flux、sd-3、hunyuan、seedream、agnes-image、step-image、qwen-image、wan、imagen、gpt-image、dall、t2i、i2i 等。

每个厂商的 API 差异按官方文档自动处理：

🏗️ 系统架构

┌─────────────────────────────────────────────────────────────┐
│                   DeepSeek Harness (DSH)                    │
│                                                             │
│  ┌─────────────────────────────────────────────────────┐    │
│  │              dsh-draw-router（插件）                 │    │
│  │                                                     │    │
│  │  ┌──────────────┐   ┌────────────────────────────┐  │    │
│  │  │  Agent 工具   │   │    REST API (webServer)    │  │    │
│  │  │              │   │                            │  │    │
│  │  │ draw_image   │   │ GET  /dsh-draw-router/     │  │    │
│  │  │ draw_list_   │   │      sources              │  │    │
│  │  │   sources    │   │ POST /dsh-draw-router/     │  │    │
│  │  └──────┬───────┘   │      sources              │  │    │
│  │         │           │ POST /dsh-draw-router/     │  │    │
│  │         │           │      draw                 │  │    │
│  │  ┌──────▼───────────────────────────────┐        │    │
│  │  │         绘图路由器核心                 │        │    │
│  │  │  - 绘图源注册表 (baseURL+apiKey)      │        │    │
│  │  │  - 模型发现 (GET /v1/models)          │        │    │
│  │  │  - 模式匹配 (30+ 种模式)              │        │    │
│  │  │  - 厂商专属 API 适配                  │        │    │
│  │  │  - draw-config.json 持久化            │        │    │
│  │  └──────┬───────────────────────────────┘        │    │
│  │         │                                          │    │
│  └─────────┼──────────────────────────────────────────┘    │
│            ▼                                               │
│  ┌────────────────────────────────────────────────┐        │
│  │            厂商图像 API                        │        │
│  │  商汤 · 阶跃 · Agnes · 千问 · Flux ...           │        │
│  └────────────────────────────────────────────────┘        │
└─────────────────────────────────────────────────────────────┘

📦 安装

支持两种安装方式。

方式一：文件安装（推荐）

克隆仓库并复制文件：

# 1. 克隆
git clone https://github.com/xiaozhe7772222/dsh-draw-router.git
cd dsh-draw-router

# 2. 复制插件文件
mkdir -p ~/.dsh/profiles/web/plugins/dsh-draw-router
cp -r lib package.json cordis.patch.yml ~/.dsh/profiles/web/plugins/dsh-draw-router/

# 3. 软链到 node_modules（让 DSH bundle 加载器能找到）
ln -sfn ../plugins/dsh-draw-router ~/.dsh/profiles/web/node_modules/dsh-draw-router

# 4. 在 ~/.dsh/profiles/web/package.json 的 bundles 数组中添加 "dsh-draw-router"

方式二：终端命令安装

通过 DSH CLI 直接安装：

# 如果插件已发布到 npm registry：
cd ~/.dsh/profiles/web
dsh plugin add dsh-draw-router --profile web

# 或从本地包安装：
dsh plugin add /path/to/dsh-draw-router --profile web

两种方式任选其一后，重启 DSH：

npx @deepseek-ai/dsh web

验证是否安装成功：

# 检查 HTTP 服务
curl http://127.0.0.1:3080/dsh-draw-router/sources

# 应返回：{"sources":{}}

⚙️ 配置说明

方式一：静态配置（cordis.patch.yml）

- insert:
    - id: draw-router
      name: dsh-draw-router
      inject: [tools, webServer]
      config:
        sources:
          stepfun:
            baseUrl: https://api.stepfun.com/v1
            apiKey: your-stepfun-key
          sensenova:
            baseUrl: https://token.sensenova.cn/v1
            apiKey: your-sensenova-key
          agnes:
            baseUrl: https://apihub.agnes-ai.com/v1
            apiKey: your-agnes-key

baseURL 提示：填 API 根地址（如 https://api.stepfun.com/v1）。插件会自动补全 /images/generations 和 /models。

方式二：REST API（运行时）

# 添加绘图源（自动探测模型）
curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"add","name":"stepfun","baseUrl":"https://api.stepfun.com/v1","apiKey":"sk-xxx"}'

# 列出绘图源 + 已识别的模型
curl http://127.0.0.1:3080/dsh-draw-router/sources

# 手动添加模型（用于不通过 /v1/models 暴露绘图模型的厂商）
curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"addModel","name":"tokenrhythm","model":"qwen-image-2.0"}'

# 直接生图
curl -X POST http://127.0.0.1:3080/dsh-draw-router/draw \
  -H 'Content-Type: application/json' \
  -d '{"source":"sensenova","prompt":"一只可爱的猫","size":"2048x2048"}'

🔌 REST API 参考

🛠️ 故障排查

安装问题

1. 启动报错 cannot resolve profile bundle "dsh-draw-router"

启动 DSH 时 bundle 加载器找不到插件。

解决：

# 确保软链存在
ln -sfn ../plugins/dsh-draw-router ~/.dsh/profiles/web/node_modules/dsh-draw-router

# 确认 package.json 的 bundles 数组中有 "dsh-draw-router"（字符串格式，不是对象）

2. 启动报错 unsupported JSON schema

插件加载时工具 schema 验证失败。

说明你修改了源码。output.schema 必须包含 additionalProperties: false（或 true），可选参数不能有 required: false。请使用原版 lib/index.js。

3. 插件加载成功但 draw_image 工具不可用

Agent 看不到绘图工具。

检查 cordis.patch.yml 中的 inject: [tools, webServer] 是否完整，然后重启 DSH。也可通过以下命令验证：

cd ~ && DSH_SKINS_DIR=... npx -y @deepseek-ai/dsh --profile web --dump-config 2>/dev/null | grep -A3 draw-router

运行问题

4. 添加绘图源后返回 imageModels: []

自动探测没有找到绘图模型。

有些厂商（如 TokenRhythm）不通过 /v1/models 暴露绘图模型。解决：手动添加模型。

curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"addModel","name":"tokenrhythm","model":"qwen-image-2.0"}'

5. 生图失败 HTTP 402 / quota_exceeded

厂商账户余额不足。

这是账户计费问题，不是插件 bug。给厂商充值或切换到其他绘图源：

curl -X POST http://127.0.0.1:3080/dsh-draw-router/draw \
  -H 'Content-Type: application/json' \
  -d '{"source":"sensenova","prompt":"一只猫","size":"2048x2048"}'

6. 生图失败 field Size invalid

尺寸参数不被该厂商的模型支持。

不同厂商支持的尺寸不同：

• 商汤 sensenova-u1-fast：2048x2048、1664x2496、2752x1536 等
• 阶跃：1024x1024、768x768、1280x800
• Agnes：1K/2K/3K/4K + ratio

如果不确定，省略 size 让厂商使用默认值。

7. 生图失败 HTTP 401 / 403

API Key 无效或缺失。

通过脱敏值检查源 Key，重新添加正确的 Key：

curl -X POST http://127.0.0.1:3080/dsh-draw-router/sources \
  -H 'Content-Type: application/json' \
  -d '{"action":"remove","name":"bad-source"}'
# 然后用正确的 Key 重新添加

8. 请求超时

大图或慢厂商可能需要 30-120 秒。

插件使用 120 秒请求超时。如果厂商较慢，可重试一次——生图本身需要时间。

9. 设置面板不显示

REST API 正常工作但设置卡片不显示。

Client bundle 从 lib/client.js 加载。确保：

• package.json 有 "client": { "platform": "web" }
• 重启 DSH 后浏览器硬刷新（Ctrl+Shift+R）
• 如果仍然不显示，插件的服务端功能仍然正常——通过 REST API 配置即可。

10. draw-config.json 损坏

崩溃后持久化文件损坏。

删除后重启——通过 cordis.patch.yml 配置的源会重新加载，运行时添加的源重新添加即可。

🗂️ 项目结构

dsh-draw-router/
├── lib/
│   ├── index.js      # 服务端：模型发现、绘图逻辑、agent 工具、REST API
│   └── client.js     # 客户端：设置面板卡片（可选）
├── package.json      # 包元数据 + DSH bundle 声明
├── cordis.patch.yml  # 静态配置示例
├── draw-config.json  # 运行时持久化（已 gitignore）
├── README.md
├── CHANGELOG.md
└── LICENSE

🏷️ 标签 (Topics)

deepseek-harness · dsh · dsh-plugin · draw · image-generation · text-to-image · sensenova · stepfun · agnes · qwen · wanxiang · tongyi · flux · imagen · stable-diffusion · hunyuan · seedream · jimeng · doubao · gpt-image · dall-e · openai · midjourney · kling · cogview

📄 许可证

MIT License — 自由使用、修改、分发。请勿将你的真实 API Key 提交到公开仓库。