xuanfenglajiao/dsh-effort-menu

A replacement model menu for the DeepSeek Harness composer: pick a model and set its reasoning effort in the same popup.

English | 简体中文

In the shipped menu, effort lives in a submenu: after switching models, the "Effort" row opens a second list with Default, Low, High and Max. Following zcode's approach, this plugin puts the levels on a slider at the bottom of the menu instead. The slider has one node per level the current model supports, so the effort can be dragged right after picking a model.

The three most recent models sit at the top, and hovering the provider row opens every model in a flyout. Reads and writes go through the shipped per-session model directory, so the native selector and the /model popup share the same state.

Screenshots

The model menu, with the reasoning-effort slider at the bottom.

Light	Dark

Hovering the provider row opens a flyout with every provider's models.

Features

• The current level carries over on a model switch; if the new model lacks it, its highest level is used.
• A model without its own default gets a leading "Default" node that leaves the choice to the provider.
• A model list that fails to load is reported in the menu with a retry button.
• Keyboard support. Arrow keys move between items and adjust the slider, Esc closes the menu.
• Follows the app's light and dark themes; interface text in English and Chinese.

Install

Requires DSH 0.2.2-rc.1 or newer, with a build that ships the dsh-client-ui-conversation, dsh-client-ui-model-selection and dsh-client-locale client packages (see Compatibility).

Web

With a CLI-managed profile, one command installs it into web.

dsh plugin --profile web add dsh-effort-menu

There is no build step. The browser half is the committed client.js. Restart dsh web, or reload the page.

Desktop

The desktop profile is managed by the app. Fully quit the app before using the CLI, or keep it running and install the checkout directory through Plugin Manager.

# CLI install, with the desktop app fully quit
dsh plugin --profile desktop add dsh-effort-menu

# Or clone it and point Plugin Manager at the directory
git clone https://github.com/xuanfenglajiao/dsh-effort-menu.git

Restart the app after installing.

Manual

If you would rather not use the CLI, edit the profile's own package.json.

Web profile. Edit ~/.dsh/profiles/web/package.json and add the package to dependencies.

"dependencies": {
  "dsh-effort-menu": "^1.0.0"
}

Then append it to the dsh.profile.bundles array.

"bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "dsh-effort-menu"]

Run pnpm install in that directory and restart dsh web.

Desktop profile. Edit ~/.dsh/profiles/desktop/package.json and add the same entry to dependencies only. The desktop profile does not list plugins in bundles.

"dependencies": {
  "dsh-effort-menu": "^1.0.0"
}

Run pnpm install and restart the app.

Uninstall with dsh plugin --profile web remove dsh-effort-menu, or remove it in Plugin Manager on the desktop app.

Compatibility

• Supports DSH 0.2.2-rc.1 and newer. Developed and tested on DeepSeek Harness Desktop 44.7.0 (DSH 0.2.2-rc.1) on Windows 11; earlier DSH versions are not supported.
• The current model must expose reasoning.efforts for the slider to have anything to show. Otherwise the menu says the model provides no levels.
• Renders nothing in subagent sessions.

Known limitations

• One occupant of the model slot. The plugin takes over the single slot conversation.input.model at priority -1, shadowing the shipped selector. Another plugin that replaces the same slot competes for it, so install only one. Plugins that restyle the shipped selector from outside may not apply to this menu.
• The model label sub-slot is not extensible. conversation.input.model.label is declared by the shipped model-selection package, and a second declaration makes the whole client entry fail to load. This plugin therefore renders model names itself, and labels that other plugins contribute to that sub-slot do not appear here.
• No automated tests yet. Changes are checked by hand in a running app.