Reasoning

Extended thinking for supported models

Overview

Models that support reasoning can “think” before answering, producing higher-quality responses for complex questions. When a reasoning-capable model is available, ChatJS automatically creates two variants in the model selector so you can toggle reasoning on or off per conversation.

How it works

Automatic model splitting

When a model is tagged with reasoning by the gateway, ChatJS creates two entries in the model selector:

Both variants point to the same underlying API model. The -reasoning suffix controls whether reasoning options are sent with the request.

UI

When a model responds with reasoning content, the UI renders a collapsible “Thinking” section above the answer. You can expand it to inspect the model’s chain of thought.

While the model is still generating, the reasoning section streams in real time.

Cross-model compatibility

Reasoning parts are automatically stripped from message history before sending to the LLM. This prevents errors when a user switches between a reasoning and non-reasoning model mid-conversation.

Supported providers

Each provider activates reasoning differently. ChatJS handles this automatically based on the model’s owned_by field.

Customization

Adjusting reasoning budgets

Modify the token budgets in lib/ai/providers.ts inside the getModelProviderOptions function:

// Anthropic thinking budget
anthropic: {
  thinking: {
    type: "enabled",
    budgetTokens: 4096, // Increase for deeper reasoning
  },
}

// Google thinking budget
google: {
  thinkingConfig: {
    thinkingBudget: 10_000, // Increase for deeper reasoning
  },
}

Disabling reasoning variants

To hide reasoning variants for a specific model, add the base model ID to disabledModels in chat.config.ts. The reasoning variant is derived from the base model, so disabling the base model removes both variants.

To disable reasoning globally, you would need to modify buildAppModels in lib/ai/app-models.ts to skip the reasoning branch.