dsh-vision-bridge

Vision bridge for DeepSeek Harness (dsh) — a self-contained replacement for dsh-vision-router.

When the chat model has no vision (e.g. deepseek-v4-flash) and a message contains an image, the image never reaches the text-only model. Instead, the plugin substitutes an automatic text description from the vision model you choose (Hermes-style). The text model just reads text and keeps the conversation going; the image stays in the session log and in the UI.

• describe_image — a tool for when you need more precise details on an image (ask a model).
• Description cache by contentHash of the bytes — an image is described once, then the previous description is reused on later turns.
• You pick the vision model in Settings → Vision (a top-level settings section alongside General / Models / Plugins).

Install

# From npm after publishing:
dsh plugin --profile web add @goodandready/dsh-vision-bridge

# From GitHub:
dsh plugin --profile web add github:GooDAnDReaDY/dsh-vision-bridge
# Locally from a checkout:
dsh plugin --profile web add /path/to/dsh-vision-bridge

Restart the Web UI, open Settings → Vision, pick a provider and model (or leave empty — auto-picks the first vision-capable model).

Replacing dsh-vision-router

This does the same job but under your control and with your own vision model. If dsh-vision-router is installed, remove it:

dsh plugin --profile web remove dsh-vision-router

How it works (Hermes-style)

1. An image (from you or from a tool like generate_image) enters the chat and is shown by the UI.
2. On every LLM request, at two points:
   • agent/pre-step, which sees the messages claimed from the inbox — that is, the images you attach. Rewriting them here puts the description into the session history, so the model still remembers the image on later turns.
   • llm/stream, which sees the whole outgoing request. This is the net under everything else: a tool result is appended straight to the session and never passes through pre-step, so an image a tool produced (generate_image, for one) would otherwise reach the adapter untouched and fail the turn with does not support image input.
3. At either point the rule is the same:
   • if the chat model is vision-capable (inputModalities includes image) — images go through as-is;
   • if the model is text-only — for each image block:
     • already cached description (by attachmentId or content hash) → reuse it;
     • otherwise automatically call the vision model via ctx.llm.stream, get a description, cache it, and inject it as [The user attached an image. Here is what it contains: ...].
4. The text model reads text, not pixels. Turns never fail with UNSUPPORTED_CONTENT.
5. describe_image stays available for when you need more precise details on an image (ask a model directly).

Settings

Settings → Vision (top-level section):

• Provider — the vision model's provider (from the LLM catalog).
• Model — the model (only vision-capable ones are shown).
• Empty → auto-pick the first vision model.

In settings.yaml:

dsh-vision-bridge:
  visionProvider: ""   # empty = auto-pick
  visionModel: ""      # empty = auto-pick
  sanitizeImages: true
  maxImageBytes: 20971520
  timeoutMs: 120000

Structure

dsh-vision-bridge/
├── package.json            # dsh bundle/plugin metadata + peerDependencies
├── cordis.patch.yml        # bundle layer: inserts the plugin row
├── lib/index.js            # host: agent/pre-step sanitizer + describe_image + cache + model list
├── lib/client.js           # browser: top-level Settings → Vision section
├── README.md
└── LICENSE                 # MIT

License

MIT