dsh-vision-subagent

dsh-vision-subagent

Eyes for text-only DeepSeek Harness agents: delegate image reading to a one-shot subagent running on a separately configured vision route (MiniMax / Kimi / any OpenAI-compatible provider). Image bytes and the vision model's intermediate context never enter the main session — only the final text answer comes back.

Why a subagent

• Context isolation: large screenshots and multi-image comparisons never occupy the main model's window
• Multi-turn visual reasoning: the child can call read_image on more workspace files before answering
• Cost & route separation: vision calls bill on the MiniMax/Kimi route; the main model only reasons

Quick start

sh dsh plugin --profile web add /path/to/dsh-vision-subagent

Then edit ~/.dsh/profiles/web/cordis.patch.yml:

`yaml

• insert:
  • id: vision-subagent
    name: 'dsh-vision-subagent'
    config:
    provider: kimi-coding # or minimax-cn / a hand-declared route
    model: k3 # or MiniMax-M3 / MiniMax-VL-01

`

Restart dsh web, open a new session, and ask: "Look at ~/Desktop/error.png — what is the error?" The model calls vision_agent(images=[...], question=...) on its own.

Screenshots

Paste & ask	Analyzing	Clean bubble	Lightbox

Paste images into the composer (Codex-style)

The Web composer accepts pasted/dropped images natively. On send, the client plugin uploads them to the host endpoint, which:

1. validates the session and stores the images as durable attachments (bounded by deployment limits)
2. runs ONE vision-route analysis on an isolated context (image bytes never enter the main session), guided by your draft message — the analysis focuses on what your words target (error text for a debugging ask, outfit details for a styling ask) instead of describing everything generically
3. sends only the analysis text along with your message — the main model answers immediately, no tool call needed

In the chat history your bubble shows only your own words plus thumbnails; the analysis lives in the lightbox that opens when you click a thumbnail, never duplicated inline. Need the original bytes later (image editing, pixel-level inspection)? The durable message links let the model call vision_image_fetch to materialize the full-fidelity file into the workspace's .dsh-vision/ directory.

On failure (timeout, route error) the message is not sent and the composer draft is preserved. This channel complements the vision_agent tool: pasted images take the automatic path, while workspace files are read by the model calling the tool itself.

MiniMax / Kimi vision models

Provider	baseURL	Vision models	Key env
Kimi (Moonshot)	https://api.moonshot.cn/v1	k3 / kimi-k3 / moonshot-v1-8k-vision-preview	MOONSHOT_API_KEY
MiniMax	https://api.minimaxi.com/v1	MiniMax-VL-01	MINIMAX_API_KEY
MiniMax CN	(built-in llm-pi-ai minimax-cn route)	MiniMax-M3	MINIMAX_CN_API_KEY

If a route already exists in Settings/Models (e.g. kimi-coding, minimax-cn), the plugin config only names provider + model — the key stays in the route's credential reference. The plugin itself never touches secrets.

Configuration

Field	Default	Meaning
enabled	true	Master switch
provider / model	'' (dormant)	Vision route; must be set together
subagentProvider	spawn	ctx.subagents provider
maxDepth	1	Absolute delegation-depth cap for the spawned child; 1 lets the vision child run but forbids further delegation (0 would reject the child itself — a top-level agent's child is depth 1)
maxImages	4	Images per call
maxImageBytes	10 MiB	Per-image byte cap
maxPromptChars	8000	Question length cap
maxOutputChars	32000	Returned text truncation
allowRemoteUrls	false	Reserved (v0.1 supports local paths only)
allowOutsideWorkspace	false	Workspace containment bypass
extraAllowedRoots	[]	Extra allowed image roots
guidance	''	Extra instructions appended to the child prompt

Security model

• Keys live only in the vision route's credential reference (env); the plugin accepts no plaintext secrets
• Local images default to the session workspace; symlinks are rejected; reads are byte-capped
• The child runs with maxDepth: 1 (runs, but cannot delegate further) and instructions forbid file modification and shell use
• vision_image_fetch writes only under the session workspace's .dsh-vision/ with self-generated content-hashed filenames; no caller-controlled path segment reaches the disk

Architecture

Main model (text-only) └─ vision_agent(images, question) ──┐ │ 1. admission: ext / containment / symlink / byte cap │ 2. ctx.attachments.saveImage → durable content-addressed refs │ 3. ctx.subagents.start('spawn', { agentOptions: {provider, model} }) ▼ One-shot subagent (MiniMax/Kimi vision route, own context) └─ final text ──► main session (only this message enters main context)

The plugin consumes harness services structurally (duck-typed) and is rc-version tolerant. Runtime dependencies: @deepseek-ai/dsh-tools (defineTool) and @deepseek-ai/schemastery (config schema) only.

Roadmap

• Web paste bridge: composer images auto-trigger vision analysis (v0.2)
• Context-aware paste analysis: your draft message steers the vision focus; bubble stays clean (v0.3)
• vision_image_fetch: materialize pasted originals into .dsh-vision/ for editing (v0.3)
• Settings panel for provider/model selection
• Remote image URL support (bounded fetch)
• Embedded SKILL.md steering when to delegate

Development

sh npm install && npm run typecheck && npm test && npm run build

Before publishing, align @deepseek-ai/* versions in dependencies/devDependencies with the target harness rc (runtime deps currently rc.6, peer range >=rc.5 <0.1.0 — compatible with a local rc.5 checkout).

License

MIT