dsh-design-qa：DeepSeek Harness でデザイン QA と画像読み取り補助｜導入ガイド - DeepSeek Harness (dsh) プラグインの使い方・インストール

作者の根拠に基づく DSH プラグイン記録。デザイン QA と画像読み取り補助の導入、動作、制約をまとめています。

DSH 統合

ネイティブ実行

作者による申告

安全性監査

未監査

最終検証日

2026-09-06

ライセンス

MIT

01どんなタスクに使えるのか？

Let a text-only model inside DeepSeek Harness read images and judge whether the implementation matches the mock
A design-fidelity verdict that traces back to evidence, plus a benchmark (4 fixtures, 23 injected defects, four pass lines) with every raw model transcript
Developers and design engineers running product design on DSH who need a text-only model to act as the referee in the fidelity loop
a760f72e965e…a760f72e965e…
Borrow multimodal image-reading for any text-only model wired into DSH
The model calls `deepseek_vision` itself via tool calling; the image never enters the main model context and costs nothing unless the model decides to look
Users who bring their own OpenAI-compatible text endpoints into DSH and want those models to handle images
a760f72e965e…a760f72e965e…
Step aside automatically once DeepSeek ships native vision
No changes needed: `deepseek_vision` self-rejects at runtime and the installer skips the core patch; the plugin can stay installed as-is
DSH users who care about plugin lifecycle and maintainability
a760f72e965e…

02DeepSeek Harness への導入方法

前提条件

A DeepSeek Harness that can already hold a normal conversation — workspace selected and the main model key configured
Node `^22.19 || >=24` (same as DSH)
A BAILIAN_API_KEY from Alibaba Cloud Bailian (the plugin only handles the image-reading link)

インストール手順

01
Run `dsh plugin --profile web add dsh-design-qa`
$ dsh plugin --profile web add dsh-design-qa
02
cd into the installed folder: `${DSH_HOME:-$HOME/.dsh}/profiles/web/node_modules/dsh-design-qa`
03
Set the key: `export BAILIAN_API_KEY=<your key>`
04
Run `node install.mjs --route-only`
05
Run `node install.mjs --restart` (cold start; HMR is off)

導入成功の確認

Paste an image into the chat box and ask "what is this"
Images go in by paste or drag — DSH's input has no separate upload button

ロールバック

Run `node install.mjs --revert-patches` to restore the patched DSH core files
Then `dsh plugin --profile web remove dsh-design-qa`

03DSH 統合と能力の範囲

DSH 統合ネイティブ実行

Installed through the official dsh CLI (`dsh plugin add`); registers a `deepseek_vision` tool and patches the local DSH core so text-only models can read images.

deepseek_vision image-reading tool
image path / http(s) URL / attachment id→an answer to a question about the image
a760f72e965e…a760f72e965e…
Image-paste core patch (3 DSH core edits)
running `node install.mjs --route-only`→text-only routes accept pasted images
Edits three files in the DSH core on disk (auto-backed-up; restorable via `node install.mjs --revert-patches`)
a760f72e965e…a760f72e965e…
Design-fidelity benchmark
mock and implementation images→a fidelity verdict plus 4 fixtures, 23 injected defects and four pass lines
a760f72e965e…a760f72e965e…

04誰に向いているのか？使うべきでない場面は？

向いている用途

Developers and design engineers running product design on DSH who need a text-only model to act as the referee in the fidelity loop
Users who bring their own OpenAI-compatible text endpoints into DSH and want those models to handle images
DSH users who care about plugin lifecycle and maintainability

不向きな用途

The targeted model must support tool calling — otherwise it only sees a `[图片 …]` pointer and cannot call the tool. Confirm the endpoint accepts the `tools` parameter before wiring the route.
Installation edits three DSH core files on disk. Uninstall must explicitly run `node install.mjs --revert-patches` first — otherwise the leftover patch tells the model to call a tool that no longer exists.

05互換性・メンテナンス・セキュリティ上の注意

The targeted model must support tool calling — otherwise it only sees a `[图片 …]` pointer and cannot call the tool. Confirm the endpoint accepts the `tools` parameter before wiring the route.
Installation edits three DSH core files on disk. Uninstall must explicitly run `node install.mjs --revert-patches` first — otherwise the leftover patch tells the model to call a tool that no longer exists.
`deepseek_vision` reads one image at a time — multi-image comparison is not covered; side-by-side mock-vs-implementation needs to be assembled by the caller. Images behind login (Feishu, Notion) return 401/403 on a direct URL.

2026-08-172026-08-25v0.1.3

MIT · latest release v0.1.3 (2026-08-20)