Introduction

The DeepSeek Harness (DSH) ecosystem adopts an “everything is a plugin” philosophy. Many text-only LLMs have limitations when processing images. The dsh-modlens plugin is designed to complement these models with visual capabilities. It converts images into structured JSON evidence (including OCR, layout, and semantic information), enabling pure text models to “see” and process images.

Plugin Overview

This plugin is maintained by YZz-S and is a fork of liustack/modlens. It preserves all upstream behavior while adding multi-engine support and the ability to select an engine per invocation. It currently includes configuration schemes for engines such as Gemini, Volcengine Ark/Doubao, and Claude Code, with tested recipes especially for Doubao/Ark.

Installation and Enablement

Install it in a DSH environment by running the following command:

dsh plugin --profile web add github:YZz-S/dsh-modlens

After installation, restart DSH. The system will show a (modlens vision) model entry and provide the modlens_read_image tool.

Configuring Engines

The configuration file is located at ~/.modlens/config.json (shared with upstream). Run modlens doctor to check which engines are ready.

Gemini

modlens config set gemini-api.apiKey <key>
modlens config set provider gemini-api

Doubao / Volcengine Ark

modlens config set openai.baseUrl https://ark.cn-beijing.volces.com/api/coding/v3
modlens config set openai.apiKey <ARK_CODING_PLAN_API_KEY>
modlens config set openai.model doubao-seed-2.1-turbo

Selecting an Engine per Invocation

When a model needs to read an image and multiple engines are configured, the tool description instructs the model to first ask the user to choose. When calling modlens_read_image, specify the engine via the provider parameter, for example gemini-api or openai (Doubao/Ark). If the parameter is omitted, the default engine and its failover chain are used.

Verification

Local diagnostics do not consume any quota:

modlens doctor

An end-to-end image read consumes one API call:

modlens -i image.png

Notes

  • The license is MIT. Copyright of the original work: 2026 Leon Liu. Copyright of branch changes: 2026 YZz-S.
  • Upstream ModLens is under active development. This branch primarily adds provider pass-through and Doubao/Ark recipes.
  • DSH plugin directories (such as skillhub.cn) have no official affiliation with DeepSeek or Huanfang.