dsh-model-switch
English | 中文
Use the right model for each stage of work in DeepSeek Harness.
The main conversation, subagents, and plan execution serve different purposes. dsh-model-switch lets each of them use a suitable model without changing the normal DSH workflow.
Why use it
- Match models to tasks — keep your preferred model in the main conversation and use another model for subagents or plan execution.
- Avoid repeated switching — save defaults once instead of changing models before every task.
- Confirm before execution — review or change the model directly in the plan approval panel.
- See what subagents use — display the model, context window, and reasoning effort, such as
GPT-5.6 Sol 1M Max.
- Keep the main model unchanged — after plan execution, DSH restores the previous main-session model.
Preview
Configure model defaults
Model switch settings
Choose the model for plan execution
See the model used by a subagent
Open a subagent session to see its model, context window, and reasoning effort in the session header.
Subagent model details in the session header
Install
Use the GitHub spec. This plugin is not published as an npm package, so do not use @lincong1987/dsh-model-switch.
dsh plugin --profile web add github:lincong1987/dsh-model-switch
Replace web with the profile DSH is actually running. DeepSeek Harness Desktop typically uses desktop.
Restart DSH after installation.
Update DSH
This plugin has been typechecked and built against DeepSeek Harness 0.2.0-rc.2. The manifest declares the compatible peer range ^0.2.0-rc.2; later releases are not runtime-tested by this change.
If you upgrade DSH, use the normal channel for that installation. Desktop updates its bundled DSH through the app updater; updating npm or npx does not replace Desktop's copy. Follow DSH release notes and back up ~/.dsh (Windows: %USERPROFILE%\.dsh) before crossing a release that changes on-disk formats.
After a DSH upgrade, reinstall this plugin with --force as described below.
Reinstall / after a DSH update
Reinstall the plugin when this repository changes, or after you upgrade DeepSeek Harness. This change was verified against 0.2.0-rc.2; the package peer range is ^0.2.0-rc.2.
- Update DSH first (see Update DSH; skip this step if only the plugin changed).
- Reinstall with
--force so DSH refreshes the GitHub source:
dsh plugin --profile web add --force github:lincong1987/dsh-model-switch
- Restart DSH.
On Desktop, run one plugin add at a time. Fully Exit from the tray (closing the window is not enough), wait until Desktop finishes starting, then reinstall, then Exit and reopen again. If install fails with another plugin install recovery transaction is pending, quit Desktop, wait, and retry once. If it is still pending, quit Desktop, delete %APPDATA%\DSH Desktop\plugin-install-recovery, reopen Desktop, then run a single add --force.
To bypass the Desktop wrapper, run the matching DSH CLI in a normal terminal:
dsh plugin --profile desktop add --force github:lincong1987/dsh-model-switch
Use
- Open Settings → Model switch.
- Configure the Subagent model and Plan execution model:
- Follow main model uses the current conversation model.
- Custom lets you choose a model and reasoning effort.
- When reviewing a plan, confirm or change the execution model before selecting Approve.
The model selected in the plan review panel takes priority over the saved plan-execution setting.
Compatibility
Built and typechecked against DeepSeek Harness 0.2.0-rc.2 (peer range ^0.2.0-rc.2). Settings are the plugin entry's validated Config, persisted through DSH's settings Remote under the loader entry id model-switch.
Subagent model switching applies to in-process subagents such as spawn and fork. External-process agents may manage their models independently.
DSH 0.1.2-alpha.1 also introduced a native, agent-driven model selection for subagents (the agent picks provider/model/effort within its authorization). This plugin is complementary: it enforces a fixed, user-configured routing policy instead, and adds the plan-execution model switch, restore, and session model badges.
Community
WeChat
Links
License
MIT
