Module thinking 

Thinking/Reasoning Level Control

Allows users to control how deeply the model reasons per message, trading speed for depth. Levels range from Off (fastest, most concise) to Max (deepest reasoning, slowest).

Users can set the level via:

• Inline directive: /think:high at the start of a message
• Agent config: [agent.thinking] section with default_level

Resolution hierarchy (highest priority first):

1. Inline directive (/think:<level>)
2. Session override (reserved for future use)
3. Agent config (agent.thinking.default_level)
4. Global default (Medium)

Structs

ResolvedThinking

ThinkingConfig

Configuration for thinking/reasoning level control.

ThinkingParams

Parameters derived from a thinking level, applied to the LLM request.

Enums

ThinkingLevel

How deeply the model should reason for a given message.

Functions

apply_thinking_level

Convert a ThinkingLevel into concrete parameters for the LLM request.

apply_thinking_level_with_config

Convert a ThinkingLevel into parameters, resolving native extended thinking from the provided config.

clamp_temperature

Clamp a temperature value to the valid range [0.0, 2.0].

parse_thinking_directive

Parse a /think:<level> directive from the start of a message.

resolve_thinking_from_message

resolve_thinking_level

Resolve the effective thinking level using the priority hierarchy:

validate_thinking_config

Validate thinking config at startup. Call once during agent initialization to warn about unrecognized budget_tokens keys.