API Reference
ReasoningProfile
Defined in: providers/reasoning-kwargs.ts:59
What a model family's chat template understands about reasoning.
Properties
| Property | Type | Description | Defined in |
|---|---|---|---|
defaultOn | boolean | The family's default thinking state, used when the mode is "auto". | providers/reasoning-kwargs.ts:70 |
interleaved | boolean | The model reasons between tool calls (interleaved thinking). Informational: it is enabled by default — either a server-side parser, or the chat template's own last_query_index retention (Qwen3) — rather than by the toggle kwarg. The caller must keep prior reasoning in the message history rather than strip it. A family can be both interleaved AND have a toggle (Qwen3, Kimi). | providers/reasoning-kwargs.ts:84 |
reasoning | boolean | Whether this model reasons at all. false → never inject kwargs. | providers/reasoning-kwargs.ts:61 |
toggleKey | "enable_thinking" | "thinking" | null | The chat_template_kwargs key that toggles thinking, or null when there is no runtime toggle (always-on reasoners like DeepSeek R1, or interleaved models with no switch like MiniMax-M2). Interleaved ≠ null toggle: Qwen3 and Kimi interleave AND expose a toggle. | providers/reasoning-kwargs.ts:68 |
whenOn? | Record<string, unknown> | Extra kwargs merged in WHEN THINKING IS ON — continuity of prior-turn reasoning across tool calls (e.g. GLM clear_thinking:false, Kimi preserve_thinking:true). | providers/reasoning-kwargs.ts:76 |