dsh-subagent-model-picker

In DSH, the main conversation and its subagents share a single model selection, so there is no built-in way to route subtasks to a lighter model for summarisation or a higher-reasoning model for planning without changing what the parent is running. dsh-subagent-model-picker adds a dedicated Sub: picker beside the normal model selector, letting you assign a specific provider, model, and reasoning effort to newly created subagents while the main conversation keeps its own model. Settings are stored per conversation with Inherited as the default; an explicit choice overrides the inherited policy and is saved immediately. Existing subagents retain the model they were captured with across continuation and restart, child views are read-only, and save or validation failures surface an error instead of silently switching a managed child.

This is useful when your workflow benefits from heterogeneous model selection — for example, running a strong reasoning model on the parent for orchestration while delegating retrieval, formatting, or draft generation to a faster or specialised model on child agents. The plugin controls standard in-process DSH subagents only; it does not extend to external tools such as Codex or Claude CLI, and cross-model vision support has not yet been independently qualified.

Released under the MIT licence, the project is community-maintained on a best-effort basis. It has been tested on Linux with Node 22/24 and DSH 0.1.2-rc.1, passing all 61 automated tests including a real archive roundtrip. Feedback and contributions are welcome.