dsh-autovision

Vision for text-only models inside DeepSeek Harness — paste an image, and a configured multimodal model transcribes it to text automatically. No model switching, no built-in keys, no relay.

dsh-autovision gives text-only models (DeepSeek, GLM, …) real image support in the DeepSeek Harness web UI. It registers a transparent twin provider for every pure-text model, routes image-bearing requests to a multimodal model you configure yourself, and feeds the transcription back as text — so the text model "sees" the image without you switching models or touching the request.

⭐ If this plugin saves you time, please star the repo — it helps other dsh users find it.

Why

DeepSeek Harness only lets a model receive images when that model declares image input (inputModalities). Pure-text models (e.g. deepseek-*, glm-*) reject image messages — pasting a screenshot into a session either fails silently or errors out.

Existing workarounds made you switch models, use a third-party relay, or hardcode a key. dsh-autovision keeps your setup: the plugin never ships a key, never proxies through a relay, and never touches your model config. It simply borrows the multimodal model you already configured in dsh settings to transcribe images to text.

Features

• Zero-friction, transparent — every pure-text model gets a <provider>-autovision twin registered at runtime. agent/request auto-redirects each request to the twin, so you never switch models and never edit settings.yaml.
• Paste → text, automatically — attach an image in the composer; it is transcribed by your configured vision model and injected into the text model's context. The original image stays visible in the UI (thumbnail + message), and the durable log keeps the original.
• Clean model selector — the twin's listModels returns [], so the model picker shows only your real models. No noise.
• Agent-callable autovision_read_image tool — the model can actively read an image file during a run, with its own per-task prompt (e.g. "transcribe every word", "describe the UI state").
• No built-in credentials — the recognition engine is whatever multimodal model you configure as the default vision model in the plugin settings (e.g. opencode-go, minimax-m3). No API key, no relay URL, nothing hardcoded.
• Survives dsh upgrade — pure plugin implementation, zero patches to dsh core, zero config rewrites.

Install

Requires dsh web ≥ 0.1.0-rc.6.

dsh plugin --profile web add @iroam2375/dsh-autovision

The npm package is published as @iroam2375/dsh-autovision (the bare name dsh-autovision is unavailable on npm — too similar to the existing dsh-auto-vision). The plugin itself is still addressed by its bundle id dsh-autovision.

Restart dsh web, open 设置 → 插件 (plugin settings) → Autovision, and pick a default vision model (any multimodal model available in your LLM providers, e.g. minimax-m3 / opencode-go). That model does all the transcribing; nothing else is configured.

If you develop locally, the standard bundle wiring is used: add "dsh-autovision" to dsh.profile.bundles in your profile's package.json. Do not also manually insert it into cordis.patch.yml — that produces duplicate loader entry id: autovision at boot.

Usage

…