ThinkingConfig

ThinkingConfig controls extended thinking behavior for Claude-family models, enabling them to show internal reasoning before generating a final answer. It belongs to the Anthropic config surface and is also used when Anthropic Claude is accessed through Bedrock.

Struct Definition

pub struct ThinkingConfig {
    pub enabled: bool,
    pub budget_tokens: u32,
    pub adaptive: bool,
}

Fields

enabled

Enable extended thinking in legacy mode with a fixed token budget. Ignored when adaptive is true.

budget_tokens

Maximum tokens for reasoning. Only used when adaptive is false and enabled is true.

• Minimum: 1024
• Must be less than max_tokens

Deprecated for Claude Opus 4.6+; use adaptive thinking instead.

adaptive

Enable adaptive thinking for Claude Opus 4.6 and newer. When true, Claude dynamically decides when and how much to think based on request complexity. Automatically enables interleaved thinking (Claude can think between tool calls). Control thinking depth with EffortLevel instead of budget_tokens.

Defaults to false for backward compatibility.

Constructors

ThinkingConfig::adaptive()

pub fn adaptive() -> Self;

Create an adaptive thinking config for Opus 4.6+. Adaptive thinking lets Claude dynamically decide when and how much to think. Pair with EffortLevel to control thinking depth.

use appam::llm::anthropic::ThinkingConfig;

let thinking = ThinkingConfig::adaptive();
// enabled: false, budget_tokens: 0, adaptive: true

ThinkingConfig::enabled(budget_tokens)

pub fn enabled(budget_tokens: u32) -> Self;

Create an enabled thinking config with the specified budget. Use for models before Opus 4.6 (Opus 4.5, Sonnet 4.5, etc.).

use appam::llm::anthropic::ThinkingConfig;

let thinking = ThinkingConfig::enabled(16000);
// enabled: true, budget_tokens: 16000, adaptive: false

ThinkingConfig::disabled()

pub fn disabled() -> Self;

Create a disabled thinking config.

use appam::llm::anthropic::ThinkingConfig;

let thinking = ThinkingConfig::disabled();
// enabled: false, budget_tokens: 0, adaptive: false

Model Differences

Provider Support

• Anthropic: Native extended thinking support. Thinking content streamed as StreamEvent::Reasoning(text).
• Bedrock: Supported through the Anthropic Messages API (same as direct Anthropic).

Incompatibilities

Extended thinking cannot be used with:

• temperature parameter
• top_k parameter
• Forced tool use (tool_choice: Any or tool_choice: Tool)

Builder Usage

use appam::prelude::*;
use appam::llm::anthropic::ThinkingConfig;

// Adaptive thinking for Opus 4.6+
let agent = AgentBuilder::new("thinker")
    .model("claude-opus-4-6")
    .thinking(ThinkingConfig::adaptive())
    .system_prompt("Reason carefully before answering.")
    .build()?;

// Fixed-budget thinking for older models
let agent = AgentBuilder::new("thinker")
    .model("claude-sonnet-4-5")
    .thinking(ThinkingConfig::enabled(16000))
    .max_tokens(32000)
    .system_prompt("Reason carefully before answering.")
    .build()?;

Streaming Events

When thinking is enabled, reasoning content appears in the stream:

• StreamEvent::Reasoning(text) -- Emitted during the thinking phase
• Stored in ChatMessage::reasoning field after completion

Source

Defined in src/llm/anthropic/config.rs.