dsh-reasoning-effort

Give custom DeepSeek Harness providers a reasoning-effort vocabulary they can actually use.

dsh-reasoning-effort is a small, standalone DSH Bundle that fills missing reasoning-effort declarations in the llm-pi-ai settings namespace. It lets you define provider defaults, model-specific mappings, and protocol wire values without replacing the settings that are already there.

Configuration overview: the bundle supplies settings; DSH remains the runtime UI and model adapter.

The plugin keeps the model chooser and reasoning levels together in one focused popover.

The GIF previews the slider moving through discrete levels; the maximum level switches to a purple glow, edge flash, and sweeping tail.

Why it exists

Different providers describe the same idea in different dialects. One accepts low / medium / high; another needs a provider-specific value; a third needs a model-level exception. This plugin gives those models one stable Harness-facing set of levels while keeping the wire-value mapping configurable.

The result is deliberately narrow:

• provider-aware defaults for common DSH adapter protocols;
• per-provider and per-model effort mappings;
• a provider-level default reasoning setting;
• explicit model opt-out with disabled: true;
• safe, additive settings updates that preserve existing declarations.
• a provider-aware slider beside the composer send button;
• one shared selection path with DSH's built-in model menu;
• a clear purple max-level effect when the highest advertised effort is active.

Install

Install the published Bundle into the DSH web profile, then restart the running Harness:

dsh plugin --profile web add -w --config.auto-install-peers=false dsh-reasoning-effort
dsh web

The package is prebuilt on npm, so normal installs do not need to compile the plugin from source.

Configure

The Bundle ships with examples for cliproxyapi and jyld. To configure another provider, add an override for the reasoning-effort loader in your final Cordis patch:

- insert:
    - id: reasoning-effort
      name: dsh-reasoning-effort
      config:
        providers:
          my-provider:
            api: openai-responses
            reasoning: medium
            efforts:
              low: low
              medium: medium
              high: high
              xhigh: xhigh
              max: max
            models:
              my-reasoning-model:
                efforts:
                  high: reasoning_high

Configuration rules

Field	Meaning
api	Selects the default wire-value map for a provider protocol.
reasoning	Sets a provider-level default reasoning level when one is missing.
efforts	Overrides the values sent for the generic Harness levels.
models.<id>.efforts	Overrides the map for one model.
models.<id>.disabled	Explicitly marks one model as non-reasoning.

The available Harness levels are off, minimal, low, medium, high, xhigh, and max. The built-in protocol maps cover the five common levels from low through max; custom gateways can provide their own values.

Providers must already exist in the llm-pi-ai settings. The plugin enhances an existing provider; it does not recreate one that has been removed.

Safe by default

This plugin is designed to sit beside existing DSH configuration:

• existing reasoningEfforts values are never overwritten, including false;
• model fields unrelated to reasoning effort are preserved;
• settings are retried briefly when llm-pi-ai registers late;
• settings updates trigger a fresh additive sync;
• no provider credentials, requests, tools, or telemetry are added.

In other words, it prepares the settings contract and lets DSH's existing model controls do the rest.

Composer control

The client half reads the active model's advertised effort list, renders those levels as a discrete slider, and sends the selected value through DSH's shared model directory. During a drag, the label, fill, and maximum-level effect update locally in real time; only the released position is committed to the backend. Models without reasoning metadata simply do not get an empty control.

Local development

Run the published-style Bundle through the local overlay:

pnpm install --config.auto-install-peers=false
npm run build
dsh web --patch ./cordis.yml

cordis.yml mounts src/index.ts with the same example configuration used by the Bundle patch.

License

MIT