dsh-resoning-effort

dsh-reasoning-effort

Adds configurable thinking-strength (reasoning-effort) levels to custom providers in DeepSeek Harness (hand-declared llm-pi-ai routes).

Built-in deepseek-official already offers Off/High/Max, but custom providers only expose levels when each model explicitly declares reasoningEfforts. This plugin adds a settings page that writes those declarations straight into llm-pi-ai — the composer's model picker and the request path pick them up immediately.

Highlights

• Custom (hand-declared) routes only; built-in/catalog routes are never touched.
• Explicit opt-in: only models you tick are written.
• Changes auto-save (600 ms debounce).
• Apply known model: search and apply the reasoning levels of ~420 known reasoning models (embedded catalog, regenerable via scripts/dump-catalog.mjs && scripts/embed-known.mjs).
• Provider cards collapsed by default.

Install

dsh plugin --profile web add "link:<path>/dsh-reasoning-effort"

(Or copy the package into <DSH_HOME>/profiles/node_modules/dsh-reasoning-effort and add "dsh-reasoning-effort" to the profile's dsh.profile.bundles.)

Restart the web profile afterwards.

Use

Settings → Thinking strength: expand a custom provider, tick the levels each model should offer (off / minimal / low / medium / high / xhigh / max; off omits the reasoning parameter), optionally set the provider default level and thinking format, or search an input field to apply a known model's levels. Changes write directly to llm-pi-ai's reasoningEfforts.

How it works

Settings page → settings.mutate("llm-pi-ai")
  → providers.<route>.models[].reasoningEfforts (persisted in settings.yaml)
  → dsh-llm-pi-ai hot-reload
  → resolveModelInfo().reasoning → picker shows levels, request path validates/dispatches

Limitations

• Hand-declared (declared === true) routes only; catalog modelOverrides injection is out of scope for v1.
• Requires @deepseek-ai/dsh-llm-pi-ai to be mounted.
• scripts/ holds the catalog regeneration tools and is not published with the npm package.