dsh-image-bridge

dsh-image-bridge

Read images through a dedicated vision subagent whose model is configured independently of the main agent's model.

Many text-first models (and any route the LLM runtime declares as text-only) cannot see pixels: the runtime projects every image block to placeholder text before the request leaves the process. This plugin gives such an agent a tool that looks at the image for it.

What it does

• Registers one model-facing tool, recognize_image by default.
• On a call it resolves the images the model pointed at (file paths, image attachment ids, or the most recent image already in the conversation), stores them through the Harness attachment service, and starts a child agent under ctx.subagents.
• The child runs on its own provider/model from this plugin's config — never the parent's — with the images plus the caller's question, and its answer comes back as plain text for the main model to use.
• The child runs at maxDepth: 1 (it cannot spawn further agents of its own), runs with delivery: 'caller' (no completion notice leaks into the main conversation), and is always disposed after one question.
• While enabled, a short system-prompt line tells the main model that this tool exists, so a model whose runtime strips images knows to route them here.

Configuration

Set these in the profile's cordis.patch.yml entry vision-bridge, or edit the *-marked fields live from the Settings UI.

Field	Default	Meaning
provider	(required)	Child model provider id, e.g. openai, anthropic, deepseek.
model	(required)	Child vision model id on that provider, e.g. gpt-4o-mini, qwen-vl-max.
reasoningEffort*	''	Child-route reasoning effort id; empty uses the model default.
maxTokens*	4096	Output cap per child call.
subagentProvider	spawn	Backend on ctx.subagents; spawn ships with dsh base.
subagentLabel	vision	Label prefix for each vision child; the question summary is appended.
labelQuestionChars	40	Characters of the question appended to the child label so panel rows stay distinguishable; 0 keeps the bare label.
toolName	recognize_image	Model-facing tool name.
maxDepth	1	Absolute delegation-depth cap; a child of the main agent is depth 1, so 1 admits the vision child but denies it grandchildren.
timeoutMs	180000	Per-call timeout.
systemPrompt*	''	Replace the child's built-in analyst instruction.
guideline*	true	Add the "call this tool for images" prompt line.

The bundled cordis.patch.yml leaves provider/model empty on purpose — you must fill them in with a route in your profile that accepts image input. The plugin refuses to run until both are set, and reports a clear error if a call is attempted while they are still empty. It also checks the configured model's declared inputModalities before starting the child and fails early with a clear message if that route does not accept images, instead of silently sending placeholder text to a text-only model.

Tool arguments

• question (required) — what to answer from the image(s).
• images (optional) — image file paths (absolute or workspace-relative) or image attachment ids. Several images are compared in one call. Omit it to read the most recent image sent in the conversation.

Installing

This package is published to npm as dsh-image-bridge and installed as a DSH profile bundle:

dsh plugin --profile <profile> add dsh-image-bridge

or through the DSH plugin hub / plugin manager by package name. It ships plain JavaScript (lib/index.js) and needs no build step.

Runtime packages (@deepseek-ai/cordis, @deepseek-ai/dsh-tools, @deepseek-ai/dsh-subagent) are declared as peer dependencies and supplied by the Harness installation at load time — the profile's module resolver routes these bare specifiers to the packages bundled inside the app, so this package ships with no node_modules/ of its own and stays tiny. Only the Harness needs to provide them; install nothing extra.

Development

The pure helpers (lib/pure.js) have no external imports, so the unit tests run under bare Node without the Harness runtime:

npm test        # node --test