dsh-official-reasoning-effort

Reasoning-effort control for gateway / custom providers — using the official composer UI, not a re-implementation of it.

What it solves

Many gateway APIs (e.g. deepseek-v4.1-flash on a relay) expose no reasoning effort knob of their own, yet the request still succeeds upstream. The official composer only renders the Effort row when the model declares reasoning levels — so gateways without a declaration show no such option at all.

How it works (official components only)

This plugin does not draw its own UI. It has two small halves:

1. Host half (data fill): every model in llm-pi-ai that lacks reasoningEfforts gets off / low / high / max added — additively; an existing declaration (including false) is never overwritten. Once the model carries the field, llm-pi-ai reports reasoning: true with the level list, and the official ModelSelect composer renders its own 模型 / 推理等级 rows with the drill-down level picker by itself. Nothing is re-implemented.

2. Browser half (one hint line): the model picker panel is found by its official aria-label, and when the currently selected provider+model is one whose levels were supplied generically, a CSS ::after pseudo-element adds a single small line at the bottom of the official panel:

   上游未声明思考强度，此处为通用档位，上游可能忽略或拒绝。

   Upstream does not declare reasoning effort; these are generic levels and may be ignored or rejected.

   It is hidden for models whose levels are genuinely declared upstream. The match is on the provider + model pair (data-provider-id / data-model-id), so the same model id served by a different, reasoning-capable provider is never flagged.

Levels

• Default four, identical to the official control: Off / Low / High / Max. These are the ids off / low / high / max; llm-pi-ai renders each level's label by capitalising its id, so the official picker shows exactly Off / Low / High / Max with no UI changes.

Why reasoning: high is written too

Declaring reasoningEfforts alone yields five rows (an extra Default). The official picker builds its list as:

[...(reasoning.defaultEffort === undefined
     ? [{ effort: undefined, label: t('effort.providerDefault') }]  // ← the extra Default
     : []),
 ...reasoning.efforts.map((effort) => ({ ... }))]

defaultEffort is derived by llm-pi-ai from the provider profile's reasoning field (reasoningInfo(model, profile.reasoning)). A provider that declares no reasoning has no defaultEffort, hence Default + four levels = five rows. The official dsh-llm-deepseek adapter always sets one (falling back to high), which is why the built-in provider shows four rows.

The plugin therefore writes reasoning: high for any provider whose levels it filled (High is the official "default balance for most tasks"). An explicitly configured reasoning is never overwritten, and providers needing no fill are left alone.

• A model that natively publishes more (or fewer) levels keeps its own list — the official picker renders whatever the model declares;
• On send the level maps to the wire value following the pi-ai adapter (defaulting to the level name; overridable in config).

Install

dsh plugin --profile <your-profile> add github:Ayowait/dsh-official-reasoning-effort

lib/ is prebuilt and committed, so installing from GitHub needs no local build. No client bundling step is required either — the browser half is a plain side-effect module.

Configuration

Zero-config by default: it discovers every existing llm-pi-ai provider and fills generic levels for models that declare none. Override only for special gateways / exclusions / per-model handling:

- insert:
    - id: dsh-official-reasoning-effort
      name: dsh-official-reasoning-effort
      config:
        efforts:
          high: reasoning_high
        providers:
          special-gateway:
            efforts:
              max: max_reasoning
            models:
              non-reasoning-model:
                disabled: true
          excluded-provider:
            disabled: true

Fields

Field	Meaning
auto	Auto-discover all providers (default true); when false, only providers listed under providers.
efforts	Override the wire-value map for all providers (default off/low/high/max → same name).
defaultEffort	Provider default effort written when filling levels (default high, the official default).
syncDefaultEffort	Set false to skip writing the provider default (the official menu then shows an extra Default row).
providers.<id>.efforts	Override wire values for one provider.
providers.<id>.disabled	Exclude a provider.
providers.<id>.models.<model>.efforts	Override wire values for one model.
providers.<id>.models.<model>.disabled	Mark a model as non-reasoning (reasoningEfforts: false).

Safe defaults

• Existing reasoningEfforts (including false) is never overwritten;
• Re-injection is idempotent: a value equal to the generic set this plugin would write is recognised as its own prior write (the hint keeps showing) while a genuine upstream declaration is left alone;
• Migration: the {low, medium, high} set written by earlier 0.2.x releases is recognised and upgraded once to off / low / high / max;
• Other fields are untouched; no provider credentials, requests, tools, or telemetry are added;
• The host half is dependency-free (no @deepseek-ai/* imports) — those specifiers are not resolvable from a profile plugin's install directory and importing them would take the whole host bundle down;
• The read-only route GET /dsh-official-reasoning-effort/api/injected reports the provider\0model keys that were filled, so the browser can decide whether to show the hint.

Develop / test

npm test

License

MIT