dsh-effort-slider

dsh-reasoning-effort-slider

English | 中文

A DeepSeek Harness client (web) bundle that injects a reasoning-effort slider (Codex GPT-6 style) into the conversation input dock, shown when the selected model declares reasoning effort levels.

• Per-session: sliding only changes the current session's effort (calls modelDirectories.directoryFor(session).select({provider,model,reasoningEffort})); other sessions are unaffected.
• Levels ordered along the harness seven-step off/minimal/low/medium/high/xhigh/max; only the levels the selected model supports are shown.
• Models with one or zero levels (e.g. always-thinking kimi-k3) hide the slider.
• Zero build tooling: lib/client.js is written directly in window.__ModuleLoader__.load({id, factory}) form using react/jsx-runtime jsx, for easy static audit.

Keeping the effort across model switches

DSH's own model picker only carries reasoningEffort over when you re-pick the same model; switching models falls back to the target model's defaultEffort, which is undefined when the route configures no reasoning. That undefined is not "the default level": under the qwen dialect it is equivalent to enable_thinking: false, so thinking is silently turned off (verified by capturing the real request body with a mock fetch).

This plugin therefore does two things:

1. Remembers per model: the last level you explicitly picked for each model (provider/model → level).

2. Carries it over and writes it back on a model switch, clamping to the target model's ladder:

   • exact match → keep it;
   • otherwise the nearest level at or below (previous max, target tops out at xhigh → use xhigh);
   • nothing lower → the target's lowest level.

   Since the slider never shows off, this rule cannot silently demote "thinking" to "not thinking".

If you have never picked a level (fresh state), the plugin writes nothing and shows a grey "默认" instead, honestly reflecting "unset".

Install

dsh plugin --profile web add dsh-reasoning-effort-slider

Relation to the Bailian plugin

General-purpose (any model declaring reasoningEfforts); can also be an optional dependency of dsh-bailian-models.

License

MIT