dsh-model-extension

The built-in Model page in DeepSeek Harness only handles basic model selection. Fine-grained parameters like reasoning depth, input modality, and compatibility protocol have no dedicated visual interface, and hand-editing settings.yaml is fragile and error-prone. dsh-model-extension solves exactly that: alongside the untouched built-in page, it introduces an independent Model+ navigation panel where expansion settings can be declared per model and persisted through the native DSH settings.mutate channel directly to $DSH_HOME/settings.yaml.

Three field groups form the core. Reasoning Efforts (reasoningEfforts) lets you declare which of the seven thinking levels a model supports, from off to max, along with the corresponding wire values. Input Modality (input) tags a model as text or image capable. The Compatibility block (compat) exposes the supportsReasoningEffort toggle and the thinkingFormat wire protocol. Every write goes through DSH settings.mutate, keeping the plugin fully in sync with the host configuration system. A strict version-anchor check runs at runtime: if the host version does not match the build anchor, the plugin refuses to register and logs a warning, preventing silent breakage from version drift.

This plugin is ideal for developers running DeepSeek Harness 0.1.2-alpha.2 or later who want a structured, per-model way to tune reasoning depth and modality support. The repository ships with pre-built artifacts for immediate use, and can also be built from source to lock onto a new Harness version anchor when upgrading.