dsh-vision-hub

Vision hub plugin for DeepSeek Harness. Pixel-level vision tooling, a five-tier provider fallback chain, automatic image-model routing, and a chat-side image bridge with inline previews — in one plugin.

dsh-vision-hub is the merged successor of dsh-vision-router (14 vision_* tools, provider resilience) and dsh-tool-vision (image bridge + inspect_image) — both original codebases are preserved and maintained here.

Features

14 pixel-level vision tools

vision_describe · vision_ground · vision_detect · vision_crop · vision_pixel_diff · vision_ocr · vision_trace · vision_colors · vision_extract_foreground · vision_html_screenshot · vision_screenshot · vision_long_screenshot_ocr · vision_present · vision_materialize

Plus the inspect_image tool (dsh-tool-vision compatible) for ad-hoc image inspection with your configured vision model.

Five-tier provider fallback chain

1. User-configured models (your vision_* model settings)
2. Local Ollama
3. Local LM Studio
4. Custom HTTP endpoints (OpenAI-compatible)
5. OVH anonymous free endpoint as a last-resort fallback

The resilience layer (lib/vision-resilience.js) adds a failure taxonomy, deadline budgets, a circuit breaker, per-turn memory, and structured errors so a flaky endpoint degrades gracefully instead of failing the call.

Content-hash caching & downscaling

Identical images are served from cache; oversized inputs are downscaled before being sent to a provider.

Image bridge + inline previews

When the active model is text-only (bridgeTextOnly, default on):

• Pasted images are exported to disk and replaced with a durable text hint stamped with an invisible \u200b[bridge] marker (the hint survives in chat history and keeps request reconstruction intact).
• The browser client scans for marker-stamped hints and renders inline thumbnails via a loopback route, with a lightbox on click, a configurable rescan interval, and a hide-hint-on-load option.
• Multimodal models listed in multimodalModels keep receiving raw image blocks.

Settings card

A single settings card in the plugin settings UI covers both the router tooling (progressive tools, provider chain, caching) and the image bridge (export dir, preview, scan interval). The native EAC settings page is untouched.

Install

dsh plugin add xing666173/dsh-vision-hub

Requires a DeepSeek Harness profile with the dsh-settings / dsh-tools / dsh-llm-deepseek / dsh-anonymous-user-id services (standard in desktop profiles). Restart the harness after installing.

Quick start

• Ask for the model to inspect an image: the bridge exports it and the hint points at inspect_image.
• Use vision_describe / vision_ocr etc. directly when you want a specific tool.
• If every provider is unreachable, the OVH free fallback keeps the call working with a clear warning.

Configuration

See the dsh-vision-hub settings card. Highlights:

Key	Meaning
progressiveTools	Expose tools progressively (opt-in; recommended off for long sessions)
bridgeTextOnly	Bridge images to text hints on text-only models (default true)
bridgeExportDir	Where bridged images are exported (default <tmp>/dsh-vision-bridge)
bridgePreview	Render inline previews of bridged images (default true)
bridgePreviewScanIntervalMs	Rescan interval for new hints (default 2000)
bridgePreviewHideHint	Hide the text hint once the preview loads
inspectModel / baseURL / apiKey	inspect_image provider settings (dsh-tool-vision compatible)
multimodalModels	Model IDs that keep receiving raw image blocks

Development

pnpm install
pnpm test        # 590 tests (583 pass / 7 environment-skipped)

The test suite covers provider chain construction, resilience behavior, caching, bridge export semantics, preview-route security (containment, size cap, host checks, missing-file 404, NTFS-ADS legacy filenames), client marker scanning, manifest/bundle invariants, and the rc6/rc7 contract seam.

License

MIT — three-way attribution: ysr666 (dsh-vision-router) · Scorp1o117 (dsh-tool-vision) · xing666173 (merge & maintenance). See LICENSE.

---

v0.1.0 — see CHANGELOG.md.