Plain text master model
yuqingsh/dsh-image-subagent · skills
dsh-image-subagent enables text-only master models (such as deepseek-v4-pro) to also receive image attachments: after the image enters the session, it is projected as an explicit text placeholder, which is read by the visual subagent (multimodal model) entrusted by the master model.
★ 5GitHub stars
0Forks
2026-08-15Last updated
JavaScriptLanguage
MITLicense
Key features
- Allow text-only main models (such as deepseek-v4-pro) to also receive image attachments: after the image enters the session, it is projected as an explicit text placeholder, which is read by the visual subagent (multimodal model) entrusted by the main model.
- Attach the image and attach the question → the main model receives the placeholder (including id=sha256:…) → entrust the observer to read the image → the observer uses read_attachment to return the pixel-level description → the main model answers.
Requirements
- There is a visual subagent in the preset (such as observer), and its model declares an image input (such as MiniMax-M3)
- The subagent's toolset includes read_attachment / read_image
Install command
dsh plugin --profile web add github:yuqingsh/dsh-image-subagent