dsh-vision-helper

English | 中文

Persistent vision plugin for DeepSeek Harness.

Registers the vision_analyze tool so agents can analyze images (OCR, UI/error screenshots, charts, general image description) through any configurable multimodal model, plus a settings-page UI in the Web app.

Zero dependencies by design — the host module has no imports, so installation is a plain folder copy plus one patch row. No npm install, no node_modules surgery, no junctions.

Features

• vision_analyze tool — accepts a local image path or a data: URI, sends it to the configured multimodal model, returns the analysis text.
• Configurable model — provider + model picked in 设置 → 视觉助手 (or directly in the data-directory dsh-vision-helper.json, see "Configuration"); empty fields auto-select a multimodal model.
• Robustness — strips invisible Unicode characters from pasted paths (U+202A etc.), guards images by longest edge (maxEdge, default 4096 px), dedupes streamed text, friendly error messages.
• Automatic invocation guidance — injects a system-prompt section so agents call vision_analyze on their own when a task involves images.
• Persistent & global — host composition plugin: survives restarts, available to every session.

Requirements (recipient side)

• A DSH web deployment (the settings page is Web-only; the tool itself works in any profile that mounts the plugin).
• A multimodal model route configured in the deployment's llm-pi-ai settings (the Models page) — e.g. mimo-v2.5, or any OpenAI-compatible vision endpoint you add. The plugin ships no API keys.

Install

The package is a DSH bundle (declares dsh.bundle.patch), published on npm. Prerequisite: pnpm on the machine (Node ships corepack — run corepack enable pnpm to activate it).

Method 1: npm install (recommended)

npx @deepseek-ai/dsh plugin --profile web add dsh-vision-helper

Pulls the prebuilt package from the npm registry (plain JS — no prepare script, no build authorization); the bundle layer activates automatically.

Method 2: GitHub install

npx @deepseek-ai/dsh plugin --profile web add github:Yuuz12/dsh-vision-helper

Fetches the repository source (works directly — no build step either). Prefer Method 1 when the network to GitHub is unreliable.

Method 3: manual (fallback)

Copy this dsh-vision-helper folder into <profile>/node_modules/ and add the row to <profile>/cordis.patch.yml:

- insert:
    - id: dsh-vision-helper
      name: 'dsh-vision-helper'

> Maintainer dev mode: dsh plugin --profile web add ./dsh-vision-helper (link: to a local source checkout) — changes take effect on next restart, no reinstall.

Either way the module is self-contained — zero dependencies to resolve. Verify without booting, then restart:

npx @deepseek-ai/dsh --profile web --dump-config   # shows the dsh-vision-helper layer
# then restart dsh web

Configuration

Via the settings page (设置 → 视觉助手): provider, model, temperature, max output tokens, max image edge. The equivalent config document dsh-vision-helper.json lives in the data directory, which depends on the install mode:

• registry / tarball / GitHub installs: the hosting profile directory <profile>/dsh-vision-helper.json (e.g. $DSH_HOME/profiles/web/) — user-owned and stable across plugin updates;
• local link: installs (maintainers): falls back to the plugin source checkout <repo>/dsh-vision-helper.json (when the module path carries no /profiles/ marker).

The plugin detects the location automatically from the /profiles/ marker in its module path:

{
  "provider": "opencode-go",
  "model": "mimo-v2.5",
  "temperature": 0.2,
  "maxTokens": 1024,
  "maxEdge": 4096,
  "mode": "auto"
}

provider/model empty = auto-select; maxEdge limits the longest image edge in px. mode controls the automatic-invocation guidance injected into the agent's system prompt:

• auto (default) — smart guidance: if the main model is multimodal (image delivered in conversation, or readable directly via read_image), analyze it natively with no extra cost; call vision_analyze only when the main model cannot view images (text-only model); ask for the path when missing.
• force — mandatory: image tasks MUST use the tool, no direct answering or refusal. Use this only when you explicitly want the dedicated vision model (e.g. a text-only main model, or a specialized vision model) — with a multimodal main model it incurs extra API calls.
• off — no guidance injected; the tool stays registered but the agent decides on its own.

Changes are read per use — no restart needed (the guidance section re-registers on save).

Uninstall

Standard (bundle install)

npx @deepseek-ai/dsh plugin --profile web remove dsh-vision-helper

This removes the dependency and its bundle layer (the plugin row goes with it). Then delete the config document and restart DSH:

• registry/tarball installs: <profile>/dsh-vision-helper.json (e.g. $DSH_HOME/profiles/web/)
• local link: install: <repo>/dsh-vision-helper.json (the source checkout — remove only deletes the node_modules link, the source repository is left untouched)

No other cleanup is needed — the plugin is self-contained.

Manual (fallback install)

Remove the plugin row from <profile>/cordis.patch.yml, delete <profile>/node_modules/dsh-vision-helper, delete the config document, then restart DSH.