dsh-vision

DSH (DeepSeek Harness) vision plugin: image recognition and generation for text-only models, with a multi-engine failover chain and full UI integration.

Built for personal use — engines ride what you already have (your logged-in Codex / ChatGPT account), with free Zhipu and optional Gemini fallbacks. Every engine parameter is editable in the harness Settings page, and all temporary data is delete-after-use (用完即焚).

Tools

Tool	What it does	Engine chain
describe_image	Read a local image, return a description. Default mode is structured (JSON evidence: description + OCR lines with pixel boxes + layout regions + entities); the agent integrates it into a natural-language answer. mode: text for plain description only	Codex → Gemini → Zhipu GLM
generate_image	Generate an image from a prompt; shown inline in the conversation (produced-file card + lightbox + download + reveal in Finder). quality: auto (Nano Banana 2 Lite, cheapest) / hd (Nano Banana Pro 1K/2K) / 4k (Nano Banana Pro 4K)	Codex (GPT Image) → Gemini Nano Banana (paid key) → Zhipu CogView

Image input (paste / drag → path)

The GUI blocks image paste for text-only models. This plugin intercepts paste and drag at the window level, uploads the image to ~/.dsh/generated-images/uploads/ (content-addressed: identical images are stored once), and inserts the file path as text into the composer — the model only ever sees text. A thumbnail rail above the input shows the images (preview, horizontal scroll, per-image delete that also removes the path from the draft). After sending, the images are injected back into the conversation beside your user message.

Storage discipline (用完即焚)

• Every Codex call runs --ephemeral (no session files in ~/.codex/sessions) with --sandbox workspace-write.
• Generated images: gen-<hash>-<内容>-<引擎>.<ext> + a .meta.json sidecar (engine label for the caption). The old plain-hash path is kept as a symlink so historical images keep working.
• Uploads: deduplicated by content hash; auto-cleaned after uploadRetentionDays (default 7, 0 = never); orphan .meta.json sidecars are cleaned automatically.
• Existing ~/.codex data is never touched.

Settings (Settings page → dsh-vision)

Key	Default	Meaning
codexPath	auto	Codex CLI path (blank = auto-detect)
uploadRetentionDays	7	Upload retention (0 = keep forever)
describeEngines	["codex","gemini","zhipu"]	Recognition chain order
generateEngines	["codex","gemini","zhipu"]	Generation chain order
engines.codex.*	gpt-5.6-luna / max / priority	Codex model, effort, speed tier, timeout
engines.gemini.*	aliases + Nano Banana models	Gemini describe chain + generation models
engines.zhipu.*	glm-4.6v-flash / cogview-3-flash	Zhipu models, size, timeout

Credentials (~/.dsh/.credentials.yaml)

• DEEPSEEK_API_KEY — DeepSeek (harness)
• GEMINI_API_KEY — Gemini recognition (free tier; Pro degrades to Flash)
• GEMINI_IMAGE_API_KEY — generation only, paid, separate project
• ZHIPU_API_KEY — Zhipu fallback (free)

Engine registry (adding/removing models)

Engines live in the ENGINES registry in lib/index.js:

const ENGINES = {
  codex: { id, label, describe(bytes, cfg, prompt, signal, runtime), generate(prompt, cfg, signal, runtime) },
  gemini: { ... },
  zhipu: { ... },
  // future: openai: { ... }, ollama: { ... }
};

Add = one registry entry + one settings config object + a default. Remove = delete those. Chain order and per-engine params are editable in Settings.

Install

dsh plugin --profile web add /path/to/dsh-vision

Then restart dsh web. Add keys to ~/.dsh/.credentials.yaml for the fallback engines.

Requirements

• Node.js with the DSH harness (dsh web)
• Codex CLI (npm: @openai/codex) logged in with a ChatGPT account
• Optional: Zhipu / Gemini keys for fallback engines

License

MIT