Custom (OpenAI-compatible) provider models cannot be given reasoning-effort levels from the GUI

[TeaDrink666]

Summary

When connecting models through a custom (OpenAI-compatible) provider — e.g. a relay gateway fronting ChatGPT or DeepSeek — there is no way in the web GUI to make reasoning effort selectable for those models. The composer's model picker never shows the 推理等级 (reasoning effort) pane for them, so the user cannot choose a thinking intensity, even though the engine fully supports it.

Expected

After adding a custom provider (or editing its models) in Settings → Models, the user should be able to declare which reasoning-effort levels each model supports (e.g. off / low / medium / high), and the composer's model picker should then offer those levels exactly as it does for the shipped DeepSeek models.

Actual

• The model-list editor for custom providers only edits id, name, contextWindow, maxTokens (packages/client/ui-settings-models/src/client/ModelListEditor.tsx).
• Models fetched from the gateway's /models endpoint carry no reasoning metadata at all (DiscoveredModelView is only id/name/contextWindow/maxTokens in packages/host/apiproxy/src/api/llm.ts).
• As a result resolveModel returns no reasoning field for these models (reasoningInfo() in packages/llm/llm-pi-ai/src/adapter.ts), and ModelSelect.tsx hides the effort pane entirely when model.reasoning is undefined.

Root cause

Reasoning effort is deliberately a per-model capability: the engine supports declaring models[].reasoningEfforts (level → wire spelling, e.g. { off: null, high: 'high', max: 'ultra' }) in settings.yaml for llm-pi-ai providers, and the composer surfaces it correctly — proven by the apps/web/tests/declared-reasoning.e2e.ts test. The gap is that the GUI offers no way to set that field: neither the custom-provider create card nor the model editor exposes it, and the fetch dialog cannot discover it.

Workaround

Hand-edit settings.yaml:

providers:
  my-gateway:
    models:
      - id: gpt-4o
        reasoningEfforts: { off: null, low: 'low', medium: 'medium', high: 'high' }

Suggested fix

Add a per-model reasoningEfforts editor to the custom-provider model rows (alongside contextWindow/maxTokens in the advanced fold), and optionally surface reasoning hints in the fetch dialog. The settings schema (packages/llm/llm-pi-ai/src/config.ts) already accepts the field, so this is primarily a UI gap.

[hytime]

This is the exact gap: hand-declared (llm-pi-ai) third-party models get no reasoningEfforts, so the composer never offers a reasoning-effort picker for them.

I built a small dual-face plugin for this: dsh-thinking-effort (MIT)

• Host half: auto-declares default levels (Off/High/Max, official DeepSeek style) for any model missing them — idempotent, never overrides your custom values, new models get filled in automatically.
• Browser half: adds a Settings → 思考强度档位 page to pick levels per model and set arbitrary wire values (e.g. high → ultra), plus apply-to-all presets.

Install:

dsh plugin --profile <profile> add github:hytime/dsh-thinking-effort

AI-friendly install doc: https://github.com/hytime/dsh-thinking-effort/blob/main/INSTALL.md
Promo thread: #1817