poiuyjie/dsh-vision-opencode

DeepSeek can't "see" images? OpenCode multimodal workaround: give your text-only main model a configurable vision model.

What it does

• Image pasted in chat → the vision model (e.g. MiMo-V2.5) converts it to text first, then your main model replies as usual — no model switch needed
• "Vision model" dropdown next to the input box, auto-listing every image-capable model across your providers
• Manage models under Settings → Vision; vision_read_image tool / vision-image-analysis skill for OCR, charts, screenshots
• Resilience: 60s per-call timeout, 1 retry, and a placeholder fallback so the turn never dies

Install

Option 1 (native DSH, recommended):

dsh plugin --profile web add -w github:poiuyjie/dsh-vision-opencode

Option 2 — one-click script (install.sh Ubuntu / install.ps1 Windows):

curl -fsSL https://raw.githubusercontent.com/poiuyjie/dsh-vision-opencode/main/scripts/install.sh | bash

Restart dsh, then pick a vision model in the dropdown.

Uninstall: dsh plugin --profile web remove -w dsh-vision-opencode (or uninstall.sh).

Back up image conversations before uninstalling — old image chats may no longer reach a text-only main model afterwards.

Configuration

Edit ~/.dsh/settings.yaml (or use Settings → Vision):

vision-opencode:
  provider: ''       # vision model provider; empty = not chosen
  model: ''          # vision model id; empty = not chosen
  autoConvert: true  # auto-convert toggle; set false if problems

Text-only main routes are auto-detected and get image conversion; native multimodal routes keep the stock DSH path.

Turning off reasoning for the vision model

Each model has a "Reasoning" row in Settings → Vision:

Very few models have a real "Off" entry; the rest show "Default / Force off" with a "not guaranteed" note. Turning thinking off usually cuts first-token latency and cost (MiMo is verified to stop thinking with reasoning_effort:"none").

⚠️ Providers declare "turning off thinking" very inconsistently (off:"none" / off:null / no field). The plugin can only tell "Off" from "Force off" per each catalog and best-effort a disabling param — no guarantee every provider can actually disable thinking.

FAQ

• Disable auto-convert only (keep tool + selector): set vision-opencode.autoConvert: false and restart
• Conversion fails / selector missing: usually no vision model picked or a version mismatch — check the browser console and file an issue
• Text-only vs native multimodal routes are auto-distinguished; no config sync on provider/model switches

Development

• Always tag before pushing: create and push a version tag for every push (e.g. git tag v0.4.0 && git push origin v0.4.0) so every remote update carries a traceable version marker.

License

MIT