Vision Sub-agent

When the main model does not accept image input (read_image refuses), use a workflow sub-agent pinned to a vision-capable model to "see" the figure — for checking chart visual quality (blank/overlapping regions, missing axis labels, whether a figure supports a conclusion).

Principle

DeepSeek Harness's workflow tool lets you specify provider and model independently on agent(). The key prerequisite is choosing a model in the current deployment that actually accepts image input (its inputModalities includes image), so the sub-agent can read and review the image.

Registered vision model names differ across deployments — do not hardcode one. Probe first.

How to find an available vision model

Use the llm service's listProviders() / resolveModelInfo(provider, model) to iterate providers and pick a model whose inputModalities includes image:

// Idea: iterate provider -> listModels/resolveModelInfo -> filter inputModalities includes 'image'
// Pick a model that accepts images and get its { provider, model }
const vision = /* a {provider, model} whose inputModalities contains 'image' */
await agent(prompt, { provider: vision.provider, model: vision.model })

Usage template

Dispatch the vision sub-agent via workflow (replace provider/model with what you probed):

await agent(
  'Use the read_image tool to read <absolute image path>, review the chart, and output a structured verdict: ' +
  'title, axis ticks/labels, legend, data lines, any blank or overlapping regions, and whether it is acceptable.',
  { provider: '<probed provider>', model: '<probed vision model>' }  // the vision model
)

Validated example (this host)

On the deployment hosting this preset, kimi-coding / kimi-for-coding (MiMo V2.5, inputModalities text+image, 1M context) was used and validated as the vision model:

• The sub-agent successfully read a PNG image
• It returned a structured review: title, chart type, axes, data lines, legend, and defects (e.g. "axis missing tick labels")
• The result can be used as evidence for chart QA / paper figure verification / review gates

⚠️ This is one validated candidate, not a name every deployment uses. On a different environment, re-probe as described above; on this host the economy vision model is kimi-coding/kimi-for-coding; if it is unavailable, use any other model whose inputModalities includes image.

Notes

• Do not hardcode provider / model — probe with llm.resolveModelInfo before use to confirm the model supports image input
• If the main model (e.g. deepseek-v4-flash) declares no image input, a direct read_image throws — that is expected; use this approach instead