Dsh Subagent Vision
niuniuaba/dsh-subagent-vision · model-adapters
A DeepSeek Harness bundle plugin that lets a text-only main agent (DeepSeek) read images in the same session — no model switching, no second session, no copy-paste.
★ 2GitHub stars
0Forks
2026-09-08Last updated
JavaScriptLanguage
MITLicense
Key features
- subagent-vision — this package's root plugin, which mounts two halves and does four things
- The shipped cordis.patch.yml pins qwen/qwen3.8-max as the vision processing model, so the tool works out of the box.
- The vision tool (tool.js)registers the subagent_vision delegation tool on the public ctx.tools / ctx.subagents seams, so this package implements the tool itself instead of inserting a second instance of another bundle's subagent tool.
- The row's visionTool config block carries the provider name (default spawn) and the child's agentOptions
- the tool appears when that provider appears and disappears with it.
Requirements
- A dsh installation whose base bundle mounts the subagent capability, tool-fs (read_image), attachments, and the Web surface (the stock web profile does
- the browser half needs the Web GUI).
Install command
dsh plugin --profile web add github:niuniuaba/dsh-subagent-vision