dsh-vision: a vision tool plugin that gives text-only models (deepseek-v4-flash) image recognition

[sjakdhasdh]

dsh-vision 📷 — a vision tool plugin for text-only models

I built a DeepSeek Harness plugin that adds a vision tool to models without native image input — like deepseek-v4-flash.

What it does

• Takes a local image path or http(s) URL
• Delegates the image to an OpenAI-compatible vision API (default: Alibaba Cloud Bailian qwen3.7-flash, ~cheap per-call)
• Returns a text description to the model, so the agent can "see" images in plain text

Demo: a test image with a red circle, a blue rectangle and the text "Vision Test 123" → the model correctly described all three.

Install

pnpm install && pnpm run build
dsh plugin --profile web add ./dsh-vision

Config is trivial — env vars or a profile patch layer:

- id: dsh-vision
  config:
    apiKey: sk-xxx
    model: qwen3.7-flash-2026-07-15

Zero extra runtime deps (uses built-in fetch). It also pairs with a tiny patch to dsh-llm-deepseek that lets users paste images directly (image blocks render as [图片附件: sha256:...] markers) — documented in the repo.

Repo & License

https://github.com/sjakdhasdh/dsh-vision — MIT

Ecosystem is still early — would love feedback, issues and PRs! ⭐

[sjakdhasdh]

I wrote a detailed write-up about how this plugin was built (the 4-line adapter patch, the vision tool, and 6 real pitfalls) — available in Chinese on Juejin:

📖 https://juejin.cn/post/7673561909976039450

Feedback and PRs welcome! ⭐

[sjakdhasdh]

Update: the full write-up is now also available on:

• 掘金 Juejin: https://juejin.cn/post/7673561909976039450
• CSDN: https://blog.csdn.net/qq_64659497/article/details/163757944
• 知乎 Zhihu: https://zhuanlan.zhihu.com/p/2071669323702219654