1helloman1/dsh-vision-fallback
Keeps your text-only main model and routes dropped images to a fixed vision model, so chat attachments just work.
Overrides the pre-send capability check so a text-only model can receive image-bearing messages, then on the pre-step hook sends the image, the latest question and recent context to the configured OpenAI-compatible vision endpoint and gets back a factual text observation. A model-only surface replacement swaps the image for that observation before the request reaches the main model, so the main model reads a description while the UI keeps showing your original image as a normal attachment. Observations are reused across compaction and restarts, and a usage log records what was sent where.
Install
dsh plugin --profile web add dsh-vision-fallbackThe npm line; the README's other forms are a local checkout path and a source build, plus a note that pre-release peers may need the pnpm release-age flag bypassed. Then restart dsh web — patch and bundle layers are not hot-reloaded. A vision model has to be configured before images are useful.
Compatibility
The README states the compatibility window it was tested against — DSH 0.1.2-alpha.3, alpha.4 and alpha.5 — rather than claiming an open range. The image-capability override is an entry-point compatibility layer, and the README is explicit that it does not claim the main model gains native visual reasoning: the image reaches the harness so the bridge can convert it to text. Switching the main model (DeepSeek, Kimi, MiniMax) never changes the fixed vision model. MCP, ACP and PTC durable attachments are read by attachment id rather than temporary file paths, and nested PTC subtool images are converted the same way.
Details
- Repo: 1helloman1/dsh-vision-fallback
- Category: Vision & Multimodal
- Stars: 1
- Version: 0.10.0 (npm)
Alternatives
Harvey-Will/dsh-vision-analysis · sprainJinyu/dsh-vision-link · lijian-ui/dsh-vision-toggle