dsh-sight

Plug-in vision for text-only DeepSeek Harness (dsh) models — paste an image, get a text description through a built-in VLM backend, no model switching.

中文版 → README.zh-CN.md

Features

• Built-in VLM presets — OpenCode Zen (free, keyless) and Gemini Flash (free tier), plus a custom mode for any OpenAI-compatible endpoint. Pick one in the web settings page, done.
• Multi-image batch — the vision tool takes up to 10 paths/URLs and describes all of them in ONE request, labeled per image.

How it works

1. Prompt-admission override — dsh refuses image pastes for text-only models. dsh-sight wraps apiProxy.sessions.prompt: the paste is accepted, the bytes land in /tmp/dsh-sight/image{N}/{hash}.png, and the image block becomes a path hint before entering history. Works with any provider — no model variant to switch. 2. vision tool — the model calls it with the hint path (or any local path / http(s) URL); the plugin reads the bytes and answers through the configured OpenAI-compatible VLM backend. 3. System-prompt section — teaches the model the hint → vision tool flow. 4. Web settings page (Settings → Vision) — backend source (preset or custom endpoint), an effective-config preview showing the actual request target, API-key field, and advanced knobs. Saved through the standard settings RPC and applied live, no restart (hot-reload via the dsh-sight: section of $DSH_HOME/settings.yaml). 5. Cache cleanup — pasted images are stored under /tmp/dsh-sight/image{N}/ with MD5 dedup and an LRU cap (maxImages, default 200). A boot-time sweep deletes image* dirs older than 7 days (DSH_SIGHT_MAX_AGE_DAYS), touching only the plugin's own directories; the OS clears /tmp on reboot too. 6. Security — the API key is role('secret') and never rides a settings response. Local reads are capped at 25 MiB; URL fetches get a 30s timeout, a 25 MiB cap, and must claim an image/* content type. Remote bodies are downloaded and inlined — the vision API never receives your URLs (no SSRF surface). Only png/jpeg/webp/gif/bmp are accepted.

How to use

1. Install & configure — dsh plugin --profile web add dsh-sight, then open Settings → Vision, pick a preset (or a custom endpoint) and hit Save. 2. Paste an image — it is auto-saved under a plugin store directory and the image block becomes a hint carrying the exact path, e.g. [Image #1 auto-saved to /tmp/dsh-sight/image1/xxxx.png]. The store root is OS-dependent (/tmp on Linux, /var/folders/… on macOS, %TEMP% on Windows), but the hint always shows the real full path. 3. Or call vision directly — the paths array takes the hint path above, or any local path / http(s) URL, optionally with a question:

{ "paths": ["/tmp/dsh-sight/image1/xxxx.png"], "question": "What does this chart show?" }

4. Batch — up to 10 images per call, described in one request.

Demo

<p align="center"> <img src="https://raw.githubusercontent.com/Fu3rte/dsh-sight/master/assets/demo/demo1.png" width="300" alt="pasted screenshot 1" /> <img src="https://raw.githubusercontent.com/Fu3rte/dsh-sight/master/assets/demo/demo2.png" width="300" alt="pasted screenshot 2" /> </p>

<p align="center"> <img src="https://raw.githubusercontent.com/Fu3rte/dsh-sight/master/assets/demo/demo.png" width="640" alt="dsh-sight workflow" /> </p>

<p align="center"> <img src="https://raw.githubusercontent.com/Fu3rte/dsh-sight/master/assets/demo/result.png" width="640" alt="model description" /> </p>

The vision tool's paths array takes up to 10 images per call (local paths or URLs, 25 MiB each). One request, per-image labels:

--- Image 1 ---
<description>
--- Image 2 ---
<description>

Install

Via your AI agent (recommended) — copy this to your agent:

Install dsh-sight for me: https://raw.githubusercontent.com/Fu3rte/dsh-sight/master/install.md

Or manually (npm registry, recommended):

dsh plugin --profile web add dsh-sight

Or from GitHub:

dsh plugin --profile web add github:Fu3rte/dsh-sight

Or clone it yourself:

git clone https://github.com/Fu3rte/dsh-sight.git
cd dsh-sight && pnpm install
dsh plugin --profile web add ./

> GitHub downloads slow or unstable (e.g. mainland China)? Use the npm-registry install above. Point pnpm at a mirror and the whole install — package and dependencies — stays off GitHub: pnpm config set registry https://registry.npmmirror.com

Configure

Open dsh web → Settings → Vision:

1. Pick a backend source: - a preset (opencode-zen / gemini-flash) — model / base URL fill themselves; or - Custom endpoint — fill in model, Base URL (OpenAI-compatible), and API key yourself. 2. Check the effective config card — it shows the model / endpoint / key state the tool will actually use. 3. Paste the API key if one is needed, hit Save — applied immediately.

The keyless preset needs nothing but the save button. For any other OpenAI-compatible endpoint (Aliyun Bailian Qwen, OpenAI, local models, …), pick Custom endpoint and fill in model / Base URL / API key. If a preset's model or Base URL is edited by hand, the page warns that the preset is overridden and offers to switch the row to Custom endpoint with one click.

Headless / no-GUI fallback

Config layers (highest wins):

1. settings.yaml dsh-sight: section (hot-reloads on edit) 2. DSH_SIGHT_* env vars (DSH_SIGHT_PROVIDER, DSH_SIGHT_API_KEY, DSH_SIGHT_MODEL, DSH_SIGHT_BASE_URL, DSH_SIGHT_TIMEOUT_MS, DSH_SIGHT_MAX_TOKENS, DSH_SIGHT_MAX_IMAGES, DSH_SIGHT_CONFIG) 3. ~/.config/dsh-sight/config.json (re-read on mtime change) 4. plugin row config in the profile's cordis.patch.yml 5. preset defaults

The API key is role('secret'): it never rides a settings response; the UI renders a write-only field and reports whether one is stored.

Acknowledgements

Inspired by modlens and dsh-eyes.

DeepSeek Harness: official site · GitHub

License: MIT