Harvey-Will/dsh-vision-analysis

---

✨ Why DSH Vision Analysis?

Your text-only agent can finally "see" — with a free vision source built in: install the plugin, paste an image, ask. No API key, no model swap, no local-file dance.

• 🆓 Built-in FREE vision source — ships pointed at OVHcloud AI Endpoints' anonymous tier (Qwen2.5-VL-72B). Zero cost, zero key, zero config.
• 🖼️ Image bridge for text-only models — paste or send images directly in conversation; the plugin routes them to vision automatically (native multimodal routes stay untouched).
• 🔁 Rate-limit failover — when one vision model is throttled, the next in the chain answers; if everything is exhausted you get clear recovery guidance instead of a failure.
• 🧾 Structured output — chart-data and ocr return machine-readable JSON (rows, lines, …) your agent can consume directly.
• 8 analysis modes out of the box — describe, ocr, ui-review, chart-data, object-detect, compare, code-gen, debug — each with a tuned instruction template.
• Any vision endpoint — OpenAI chat/completions or Anthropic messages wire formats. MiMo, Step, SiliconFlow, OpenRouter, Gemini (OpenAI-compat), GPT-4o, Claude, Qwen-VL, or a local Ollama / LM Studio / vLLM.
• Any input — absolute local path, http(s) URL, or base64 data: URL; up to 4 images per call with built-in comparison.
• Privacy-first by design — image bytes never enter the session log or reach your main model; only the vision model's text comes back. The debug report never reveals your API key (fully masked).
• Production-grade plumbing — result caching, retry with exponential backoff, live configuration from Settings → 插件配置.
• Dependency-light — just @deepseek-ai/schemastery + @deepseek-ai/dsh-settings at runtime.

---

🖼️ Demo

Paste an image, ask a question, get a real answer — even on a text-only model. The image is routed to your configured vision endpoint and the analysis lands straight in the conversation:

_{In the screenshot: a pasted image plus the question "这是谁?" — the vision endpoint identifies the DeepSeek fan-art character and walks through its reasoning, all without switching models or saving files locally.}

More scenarios — real outputs from the free vision models

Three everyday capabilities, each answered by a different free vision model automatically (when one is rate limited, the plugin fails over to the next).

1. OCR — pull text out of documents and screenshots

Weekly Ops Report — 2026-W33 Item 01 · Pending action: review queue / escalate blocker Item 02 · Pending action: review queue / escalate blocker … (all lines transcribed verbatim)

2. Charts → structured data your agent can use

{ "title": "Monthly Revenue — Q1–Q3", "rows": [["Jan","82"],["Feb","95"],…] }

3. UI review — a designer's eye on your interface

• Inconsistent button styling across "Add to cart" and "Checkout" (High) • Product name and price lack visual hierarchy (Medium) • Cart items unstructured; subtotal not visually distinct (Medium)

---

🚀 Quick start

# From GitHub (no npm needed)
dsh plugin --profile web add github:Harvey-Will/dsh-vision-analysis

# Or one-click from the plugin market inside the Harness

Restart the web profile and ask your agent to analyze an image by path or URL:

"Use analyze_image to OCR /tmp/screenshot.png and tell me what it says."

That works with zero configuration: the plugin ships pointed at a free anonymous vision endpoint (OVHcloud AI Endpoints, Qwen2.5-VL-72B) — no API key required.

Two ways to use it

1. analyze_image tool (zero config) — the agent reads a local path, an http(s) URL, or a data URL. Works immediately after install.

2. Paste images straight into the conversation (image bridge) — requires two setup steps:

• add the model to bridgeModels in the plugin config;
• declare image in that model's inputModalities in settings.yaml (this is what lets the Harness admit image prompts for it).

# ① ~/.dsh/settings.yaml — under llm-deepseek.models, for each text-only model:
#    inputModalities: [text, image]
# ② plugin config:
bridgeModels: [deepseek-v4-flash]

Bring your own endpoint (optional)

config:
  apiFormat: openai          # or anthropic
  baseURL: https://api.siliconflow.cn/v1
  apiKey: your-key           # leave empty for anonymous/local endpoints
  model: Qwen/Qwen2.5-VL-72B-Instruct
  fallbackModels: [Qwen3.5-9B]   # same-endpoint alternates tried on HTTP 429

---

🧭 Choose the right mode

Mode	What it does	Built-in tokens / temp
describe	General understanding (default)	4096 / 0.7
ocr	Exact text extraction	4096 / 0.0
ui-review	Design review with score	4096 / 0.5
chart-data	Tables + trend from charts	4096 / 0.0
object-detect	Objects, people, activities	4096 / 0.5
compare	Two+ images side by side	4096 / 0.5
code-gen	HTML+CSS from a UI shot	4096 / 0.3
debug	Endpoint connectivity report	4096 / 0.7

🔧 The tool

analyze_image(image?, images?, mode?, prompt?)

• image — absolute path, http(s) URL, or data:image/...;base64, URL
• images — up to maxImages (default 2, max 4) for multi-image calls
• mode — one of the eight above; describe by default
• prompt — your precise instruction overrides the mode template

A targeted prompt beats a generic description: prompt: "Extract the table as CSV" >> prompt: "Describe this".

⚙️ Configuration

- id: vision-analysis
  name: dsh-vision-analysis
  config:
    apiFormat: openai          # openai | anthropic
    baseURL: https://api.siliconflow.cn/v1
    apiKey: ''                # empty → UNIVERSAL_VISION_API_KEY → local model
    model: Qwen/Qwen2.5-VL-72B-Instruct
    defaultMode: describe
    maxImages: 2              # 1-4
    maxBytes: 10485760        # per-image cap (10 MB)
    timeoutMs: 120000
    maxTokens: 4096
    temperature: 0.7
    modes:                    # per-mode overrides
      ocr:
        temperature: 0.0

All fields are editable live from Settings → 插件配置 (API key field is masked).

---

🔒 Security & privacy

• Your images stay private: local files are read by the tool and sent base64-embedded only to your configured endpoint; the raw bytes never enter the session log or reach the main model.
• Your key stays secret: never embedded in requests to the main model; the debug report only says configured / not configured — no prefix, no characters.
• Prefer the environment: keep keys out of cordis.yml — use UNIVERSAL_VISION_API_KEY or the masked secret field in Settings.
• Endpoints are not sandboxed by tool approvals — only point the tool at endpoints you control, and only reference http(s) image URLs you trust the endpoint to fetch.
• Installing a plugin runs its code with your permissions — review the source before installing.

---

🧩 Compatibility

	Supported
DeepSeek Harness	0.1.0-rc.x (verified on rc.8)
Node.js	^22.19 || >=24
Vision wire formats	OpenAI chat/completions, Anthropic messages
Image formats	PNG, JPEG, GIF, WebP, BMP (local / URL / data URL)

⚠️ Community plugin — not an official DeepSeek product. The Harness API is in developer preview and may break between versions.

---