dsh-model-auto-hot-switch
DeepSeek Harness (dsh) 按任务自动切换模型：包含图像的任务自动路由至视觉模型，其他任务继续使用默认模型。不增加额外 token，不干扰上下文。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:SHUJILAI/dsh-model-auto-hot-switch#2a3a3ff90046fcd45ac3b0add6d90bf7cfca5bd1Model Auto Hot Switch 以 dsh-model-auto-hot-switch 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →Automatic per-task model hot-switching for DeepSeek Harness (dsh). Image-aware tasks route to the vision model automatically; every other task keeps the default model you chose. Zero extra tokens, zero context disturbance, one click to enable.
DeepSeek Harness models do not all accept images. When you ask the agent to read a screenshot, an OCR job, or any image file, the request should go to a vision-capable model — while plain chat and coding keep the fast default model. This plugin makes that switch automatic and per-task, inside the same session:
deepseek-v4-flash-vision-expprovider/model fields of the frozen per-step call configuration are replaced, through the harness's own agent/request extension point. Messages, system prompt, tools, and session state are never rewritten.Requires DeepSeek Harness (dsh) with a Web profile.
# from npm
dsh plugin --profile web add dsh-model-auto-hot-switch
# or straight from GitHub
dsh plugin --profile web add github:SHUJILAI/dsh-model-auto-hot-switch
Restart dsh web, then click the floating ⚡ button (bottom-right) to switch automatic hot-switching on. The readout shows:
The toggle persists in config.json beside the package and survives restarts.
llm.listModels / llm.resolveModelInfo — no configuration needed. For the official DeepSeek provider this is deepseek-v4-flash-vision-exp (available on the DeepSeek API). Without any vision model, the toggle still works but vision routing stays idle and the readout says so.read_image tool's own gate must also see the vision model, which requires the dsh build whose DeepSeek adapter declares image input for it.user sends an image task
│
▼
agent/pre-step ──► contains image? ── yes ──► mark this step as vision
│ │
│ no
▼ ▼
agent/request ──► return the default config return config with provider/model
(untouched, zero overhead) swapped to the vision model
The classification runs per (session, turn, step) and is consumed by the matching agent/request call, so a later step of the same turn re-classifies independently. Subagents get the same treatment under their own sessions.
None. The single toggle is the whole surface:
| Field | Meaning |
|---|---|
enabled (in config.json) | true routes image tasks to the vision model; false leaves every request on the default model |
The preferred vision model id is deepseek-v4-flash-vision-exp; if absent, the first discovered image-capable model wins.
/plugins/dsh-model-auto-hot-switch/state).@deepseek-ai/* runtime package; no peerDependencies.npm test # package-shape + syntax smoke checks
MIT — see LICENSE.