⚡ dsh-model-auto-hot-switch

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.

中文说明

What it does

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:

• Vision task (the step's messages carry an image) → routed to the discovered vision model, e.g. deepseek-v4-flash-vision-exp
• Everything else → the exact default model you selected, configuration returned untouched

Why it costs nothing

• Zero extra tokens. Task classification is pure local code (does the step contain an image block?). No classifier model call, no prompt rewrite, no duplicated requests.
• No context disturbance. Only the provider/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.
• No overhead on normal tasks. Non-image steps restore the session's default route in one field swap when a previous step left the vision model in the held request header (and return the configuration untouched otherwise) — no classifier call, no extra request.

Installation

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 routing target for vision tasks (the first discovered model whose capability declares image input),
• the default model all other tasks keep using,
• the most recent hot-switch event.

The toggle persists in config.json beside the package and survives restarts.

Prerequisites

• A vision-capable model in your provider catalog. The plugin auto-discovers it through 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.
• The 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.

How it works

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.

Configuration

None. The single toggle is the whole surface:

The preferred vision model id is deepseek-v4-flash-vision-exp; if absent, the first discovered image-capable model wins.

Compatibility

• Works with dsh Web profiles (routes under /plugins/dsh-model-auto-hot-switch/state).
• Headless profiles: routes are skipped, the hot-switch logic still runs from the persisted toggle.
• Does not depend on any official @deepseek-ai/* runtime package; no peerDependencies.

Development

npm test          # package-shape + syntax smoke checks

License

MIT — see LICENSE.