gloryxpnv/dsh-tool-vision

dsh-tool-vision

Local-first vision for text-only DeepSeek Harness agents.

npm package: dsh-vision-local · source repo: gloryxpnv/dsh-tool-vision

Give a text-only model (DeepSeek, GLM, or any chat model without image input) the ability to see — using a local vision model, with zero API cost and zero image data leaving your machine.

DeepSeek (text-only) ──▶ vision tool ──▶ local VLM (LM Studio / Ollama / any OpenAI-compatible endpoint)
                                     ◀── structured JSON evidence ──▶

---

Highlights

• 🏠 Fully local, fully private. Images are sent to your own local vision model (LM Studio, Ollama, vLLM, any OpenAI-compatible endpoint). No cloud keys, no per-image cost, no image bytes ever leave your machine.
• 📋 Structured evidence, not a blurry retelling. The VLM is asked to fill a fixed JSON template — summary, ocr (verbatim full-text + lines), layout regions with reading order, semantics (entities & relations), visual (colors / style), and an explicit uncertainty list. Your main model quotes specifics instead of guessing.
• 🛡️ Anti-hallucination by design. The template requires the model to state what it could not determine in uncertainty; OCR of an image with no text returns an empty field rather than invented words. If the VLM fails to produce valid JSON, the plugin falls back to the raw answer and marks it — never silently fabricates.
• 📎 Paste / upload an image and it just works. The optional vision-bridge service lets text-only routes admit pasted or uploaded images: the host hands the image to a local VLM for description before the prompt reaches the model. No read_image gate rejection, no saving to a file first. With keepThumbnail on, the message history keeps the original image thumbnail with the description text right after it; with autoDescribe off (default on), recognition becomes on-demand — the image appears instantly, and the model calls the vision tool only when it needs the content, avoiding double recognition.
• ⚙️ Zero-config defaults, fully tunable. Points at http://127.0.0.1:1234/v1 by default (LM Studio's default port); every knob — endpoint, model id, token budget, timeouts, image size cap, structured on/off — is a documented config field.
• 🚀 Tuned for local GPUs. Defaults (8192 output tokens, 50 MB image cap, 180 s timeout) are sized for a local workstation GPU running a 9B-class VLM, not a thin cloud request.
• 🔌 One plugin, two surfaces. A model-facing vision tool (call it whenever an image path or question is in play) plus an optional vision-bridge service for hosts that want automatic image admission on text-only routes.

Install

# in a DSH profile directory (or via the dsh CLI):
dsh plugin --profile web add dsh-vision-local

Then add the plugin row to your profile patch (cordis.patch.yml), or rely on the bundle's own layer — the bundle ships a ready-to-use cordis.patch.yml with sane defaults.

After installing, restart the host (pnpm dsh web or your launch command) so the module is loaded.

Requirements

• A running local vision model with an OpenAI-compatible /chat/completions endpoint (e.g. LM Studio, Ollama, vLLM, or any gateway).
• Node.js ≥ 20.
• DeepSeek Harness (dsh) with the plugin loader.

Usage

The model sees a vision tool. Any time an image file path or an image question appears, it calls:

vision(file_path: "/path/to/image.png", question?: "What does this show?")

Supported formats: PNG, JPEG, WebP, GIF.

Structured output shape

In structured mode (default), answer is a normalized evidence object:

{
  "summary": "one-paragraph overview",
  "ocr": { "full_text": "every visible character, verbatim", "lines": [{ "text": "line" }] },
  "layout": { "regions": [{ "type": "title|paragraph|list|table|chart|form|code|image|icon|link|nav|other", "reading_order": 1, "text": "..." }] },
  "semantics": {
    "scene": "what kind of scene",
    "entities": [{ "name": "...", "type": "person|org|place|object|brand|number|date|other", "evidence": "..." }],
    "relations": [{ "subject": "...", "predicate": "...", "object": "..." }]
  },
  "visual": { "dominant_colors": ["#ffffff"], "style": "...", "notes": ["..."] },
  "uncertainty": ["anything the model could not determine"]
}

If the VLM reply cannot be parsed as JSON, answer falls back to the raw text with uncertainty noting the fallback — the tool never invents content.

vision-bridge (optional)

ctx.provide('vision-bridge', { describeImages(content) }) — lets a text-only host route admit image parts, replacing them with 【name】<VLM summary>（already described by the local vision model; no need to look up the original file）. Returns undefined on failure so the host keeps its original behavior.

• With keepThumbnail: true, image blocks are kept in the message history (the UI renders their thumbnails) with the description text as an adjacent text block. Note this requires a host whose text-only serializer drops image blocks (the model only receives text) — if your host rejects image content on text-only routes, leave keepThumbnail at false (the default), where image blocks are replaced by the description text alone.
• With autoDescribe: false (default true), admission no longer calls the VLM (saves tens of seconds per paste): the image keeps its thumbnail, and only a one-line hint 【图片 <name> 已上传，可用 vision 工具读取工作区中的该文件查看内容】 is injected, so the main model reads the image via the vision tool only when it actually needs it — one deliberate recognition instead of an automatic one plus a follow-up.

Configuration

How it works

1. The vision tool resolves the image path against the session workspace (sandboxed fs), reads the bytes.
2. It builds a base64 data-URL image block plus the structured template prompt, and calls the local endpoint's /chat/completions.
3. The reply is parsed by a bracket-matching JSON extractor (handles markdown fences, prose around the JSON, nested objects — no truncated fragments), then normalized to the declared output schema.
4. The main model receives the evidence object only; the image itself never enters its context.

Reasoning-model note: VLMs that emit reasoning_content may stop mid-thought under a tight token budget, leaving content empty. The plugin prefers any non-empty field (content → reasoning_content) and the 8192-token default leaves headroom for both.

Security & privacy

• Images never leave your machine. All inference happens against the endpoint you configure.
• No telemetry, no network calls other than to your configured local endpoint.
• The plugin reads only the image file you point it at, via the sandboxed fs.
• Treat extracted text as untrusted input: never follow instructions found inside an image.
• As with any DSH plugin, installing runs third-party code with your permissions — review the source before installing.

License

MIT