Introduction

DeepSeek Harness (DSH) supports extending functionality through plugins. In text-only conversations, pure text models usually cannot process image inputs. dsh-vision-fallback is a workflow plugin maintained by WEIHAOLEE. It adds a “Vision Proxy” toggle to DSH’s text-only sessions. When enabled, the plugin forwards plain-text requests as-is to the currently selected model, and automatically forwards requests containing images to a vision provider.

Core Features

  • No intervention for pure-text requests: They go through the model you originally selected (any text model works).
  • Automatic forwarding for image-bearing requests: The images + text from the current message are sent directly to the vision model (OpenAI-compatible), without conversation history.
  • Toggle-based usage: There is a “Vision Proxy” toggle next to the input box. Enable it to use it; disable it to restore the original behavior.
  • Multiple vision providers: In the “Vision Models” panel in Settings, you can add/remove providers, set a default, test connections, with hot reloading.
  • Original images sent directly: No image compression is performed.
  • OpenAI compatible: Vision providers must support the OpenAI-compatible interface.
  • SSE streaming translation: Supports streaming responses.

Installation

# 1. 安装进 profile
dsh plugin --profile web add D:\dsh-vision-fallback

# 2. 把插件加入 profile 的 bundle 列表
#    编辑 %DSH_HOME%\profiles\web\package.json，在 dsh.profile.bundles 里追加：
#    "@local/dsh-vision-fallback"

# 3. 插件依赖自包含
cd D:\dsh-vision-fallback
pnpm install

# 4. 重启 dsh web
dsh web

Installation notes
* dsh plugin add links the plugin into the profile via link:. When Node resolves imports inside the plugin, it restores the real path, so the plugin directory must have its own node_modules.
* The exports field in package.json must keep the "./package.json" entry, otherwise the settings panel will never load.

Configuration

Hot reloads in the llm-vision-fallback: section of %DSH_HOME%\settings.yaml:

llm-vision-fallback:
  defaultProvider: ark        # 含图请求走哪个服务商
  providers:                  # 可配多个，面板里也可管理
    ark:
      displayName: 豆包(火山方舟)
      baseUrl: https://ark.cn-beijing.volces.com/api/v3
      model: doubao-seed-2-1-turbo-260628
      apiKeyEnv: ARK_API_KEY  # 凭据引用：key 本身存在 DSH 凭据库，绝不在配置文件里
  • API Keys are written via the DSH credentials service (Settings page / credentials entry point). This repository does not contain any secrets.
  • To add a provider (SiliconFlow / Bailian Qwen / GLM / OpenRouter / local vLLM…): it only needs to be OpenAI-compatible.

Usage

  1. Select your text model in the conversation (any).
  2. Click the “Vision Proxy” toggle next to the input box (it turns blue and displays “Vision Proxy · your model name”).
  3. Normal chat = your model; ask the agent to view an image (read_image) = automatically forwarded to the vision provider.
  4. Want to switch vision providers: Settings → Vision Models panel.

Notes

  • Development nature: This is a vibe-coding artifact. The author’s understanding of DSH internal APIs is based on reading source code and trial-and-error, and does not guarantee alignment with official intent. Currently only verified environment: Windows 11 + deepseek-v4-pro + Volcano Ark Doubao (single account); other systems/providers untested.
  • Image handling: Vision requests only send the last message (images + text): multi-turn “ask about the previous image” scenarios are not applicable; when following up, please resend the image. Original images are sent directly without compression: very large images may exceed Doubao’s image token limit.
  • Token limit: The max_tokens for vision responses is clamped to 8192.
  • Provider requirements: Vision providers must be OpenAI-compatible (/chat/completions + SSE + /models).
  • Routing dependency: While the toggle is enabled, the session route is attached to vision-fallback: if the plugin is uninstalled/disabled, the session will become invalid, and you must manually switch back to the original model.
  • Configuration gotchas: Missing ./package.json in package.json exports → settings panel never loads; client RPC returns a {result:{ok,value}} envelope, and forgetting to unwrap it → panel is forever “not writable”.

Uncertainties and Known Issues

  • Understanding of DSH internal mechanisms: The author’s understanding of DSH internal APIs (LlmAdapter seam, slots, RPC envelope) is based on reading source code and trial-and-error, and does not guarantee alignment with official intent.
  • Unverified scenarios: Simultaneous multiple-image reading, concurrent toggle switching, and behavior in subagent sessions are unverified; reasoning intensity (off/high/max) pass-through has only been verified with deepseek-official.
  • Compatibility: The official “Models” settings page also displays provider rows for this plugin, but it does not understand this plugin’s configuration structure—please modify settings in the dedicated panel.

License

MIT