dsh-model-autoconfig

dsh ships with an incomplete model catalog: some models have missing or incorrect thinking levels, vision capabilities, and context-length fields. In practice, this means thinking mode fails to toggle properly, long outputs get silently truncated, and image-capable models cannot accept images. dsh-model-autoconfig fixes these gaps automatically when you add a model, eliminating the need to hand-edit settings.yaml.

The plugin does four things. First, it corrects thinking modes—declaring whether a model supports thinking, whether it can be disabled, and which levels (Low, High, Max) are valid—and remembers your last chosen level per model in the browser. Second, it auto-configures parameters such as context length and max output length based on the vendor's official documentation. Third, it fills in vision capabilities like image-input support. Fourth, it detects and blocks invalid configurations—spurious thinking levels, missing vision flags, or parameters that dsh would reject. It covers official APIs, SiliconFlow, Volcano Ark, Alibaba Bailian, OpenRouter, and common relay stations with zero manual intervention.

This plugin is for developers and teams running dsh against multiple model backends who keep hitting mismatches in thinking levels or missing capability fields. After installing and restarting dsh web, the plugin handles the rest the moment you add a model service, giving you accurate, vendor-correct model definitions out of the box.