[UX] Models without explicit input capability are silently treated as text-only; image attachments fail with no warning

[malosusu]

Summary

DSH treats any model entry without an explicit input field in settings.yaml as text-only, silently. Users attach images to a vision-capable model and the images never reach the API - it looks like the model has no vision at all. On a real deployment this silent default caused 7 vision-capable models to be mis-declared as text-only.

Environment

• dsh-tether 1.0.0, @deepseek-ai/dsh 0.1.3-alpha.2 (dsh web)
• Models wired via llm-pi-ai.providers (openai-completions)

Steps to reproduce

1. Add a multimodal model (e.g. glm-5.3-flash) under providers.

   .models without an input field

2. Point agent-default-model at it
3. Attach an image in chat, or call an image-dependent tool

Actual behavior

Image is silently dropped / tool reports the model does not accept images. No warning, no documentation of the default.

Expected behavior (any of these would help)

• a) Document the default: missing input field = text-only
• b) Warn in UI/CLI when an image is sent to a model without image in input, with a fix hint
• c) A dsh doctor models style capability lint, flagging known multimodal families declared text-only

Evidence from a live deployment

Probing method: POST a 48x48 solid-red PNG asking for the color, accept only the literal answer red. HTTP 200 alone is NOT proof (one endpoint returned 200 but stripped the image; the model replied it sees no image).

Models mis-declared text-only due to the silent default, all confirmed vision-capable and fixed by adding input: [text, image]:

• glm-5.3-flash (default model)
• qwen3.8-flash
• glm-5.2 (on one aggregator channel; note the same model id is text-only on the official channel - capability is channel-specific)
• grok-4.6
• gpt-5.6-luna
• sensenova-6.8-flash-lite
• Qwen3.8-Flash-Next

中文说明

DSH 以 settings.yaml 中模型条目的 input 字段作为多模态能力门控，字段缺失时默认按纯文本处理且无任何提示，用户会误以为模型本身不支持识图。实测中 7 个已知多模态模型因此被误配，补上 input: [text, image] 后全部恢复正常。建议补充文档、在发送图片时给出警告，或提供能力声明校验命令。