dsh-image-modlens-bridge

A DeepSeek Harness plugin that lets text-only model routes "see" user-uploaded images. When a user uploads an image in a session whose model does not declare image input, the plugin intercepts it at agent/pre-step, saves the image to <cwd>/.dsh_images/, runs the external vision provider modlens, and replaces the image block with text evidence (summary + OCR + layout + semantics) before the message is persisted. The model request therefore never carries an image, while the model can still reason over the image content.

给纯文本模型路由用的图片桥接插件：用户上传图片时，在 agent/pre-step 拦截，把图片落盘到会话工作目录下的 .dsh_images/，调用 modlens 外部视觉通道识别，再把图片 block 替换为文本证据（summary + OCR 全文 + 版面 + 语义），因此模型请求里永远不会携带图片——纯文本模型也能"看到"图片内容。

Requires a working modlens setup (e.g. the modlens skill's run.ps1/run.sh launcher, or a modlens CLI on PATH). Vision-capable model routes are never touched — they keep the native image path.

Features / 特性

• Text-only routes only: gate on resolveModelInfo().inputModalities; routes that declare image input pass through untouched.
• Clean interception point: agent/pre-step runs before the loop persists decision messages, so neither durable history nor the request ever contains the image block.
• Fail-open by contract: modlens failure, attachment read failure, or any plugin exception degrades to an explanatory text block (with the saved path when available). A request is never blocked because of the bridge.
• Zero runtime dependencies: only Node built-ins; the modlens launcher is spawned as an external process.
• 只对声明不支持 image 输入的模型路由生效；视觉模型路由保持原生图片路径，不做任何干预。
• 拦截点在 agent/pre-step（step 1）：决策消息在此被持久化进 session，替换发生在持久化之前，历史与请求都不含图片 block。
• modlens 分析失败或附件读取失败时 fail-open：替换为说明性文本（含保存路径），绝不让请求因插件失败而卡死。

Install / 安装

# 1. clone anywhere (an ASCII path avoids Windows CJK-path issues)
git clone https://github.com/wendou-chen/dsh_image-modlens-bridge.git
# e.g. move to C:\Users\admin\.codex\dsh-image-modlens-bridge\

2. Add the dependency to your profile's package.json:

{
  "dependencies": {
    "dsh-image-modlens-bridge": "file:C:/Users/admin/.codex/dsh-image-modlens-bridge"
  }
}

3. In the profile directory (e.g. C:\Users\admin\.dsh\profiles\web), install:

pnpm install

4. Register the plugin in the profile's cordis.patch.yml:

- insert:
    - id: image-modlens-bridge
      name: 'dsh-image-modlens-bridge'

5. Restart DeepSeek Harness (plugins load at startup).

Config / 配置（profile cordis.patch.yml 的 config 字段）

Usage / 使用

1. Start a session with a text-only model (e.g. deepseek-v4-flash).
2. Upload an image in the composer.
3. The plugin automatically:
   • saves the image to <cwd>/.dsh_images/upload-<timestamp>-<id>.<ext>
   • runs modlens against it (external vision provider)
   • feeds the model the text evidence instead of the image
4. The model answers based on the vision evidence.

验证：上传后检查 <cwd>/.dsh_images/ 是否出现图片文件；模型回复应引用图片内容（摘要/OCR 文本等），而不是提示"无法查看图片"。

Troubleshooting / 排查

• 图片没有保存、模型仍说看不到：确认重启了 DSH；确认模型路由确实不声明 image（视觉模型不会被插件处理）；看 DSH 日志是否有插件报错。
• modlens 分析失败：单独运行启动器验证：powershell -ExecutionPolicy Bypass -File <launcher> -i <image>，检查 modlens 配置（modlens config show / doctor）。
• 想恢复原生行为：config.enabled: false 或移除 patch 条目。

License

MIT © 2026 wendou-chen. See LICENSE.