Module reasoning_parser 

Reasoning parser — model-agnostic thinking/reasoning block detection.

Each served model family delimits its chain-of-thought differently. This module abstracts that behind the ReasoningParser trait so the server can extract reasoning vs. final content for any model:

Format	Delimiters	Models
ReasoningFormat::Qwen	<think> / </think>	Qwen3.5/3.6/Next/VL
ReasoningFormat::DeepSeekR1	<think> / </think>	Nemotron-H / Nano-3 / Super (nano_v3)
ReasoningFormat::MiniMax	<think> / </think>	MiniMax M2 / M2.7
ReasoningFormat::Mistral	[THINK] / [/THINK]	Mistral Small 4 / Magistral
ReasoningFormat::Gemma4	`<	channel>/<channel

The Qwen / DeepSeek-R1 / MiniMax families all use <think> tags with the same extraction contract — the chat template injects the opening tag into the prompt, so output begins inside the reasoning block — and share one implementation, configured per-family with a distinct identity. Mistral differs structurally: the model emits its own [THINK], so the parser does not assume an open block. Gemma 4 uses a channel format entirely unlike <think> tags and has its own parser.

Follows the same trait + enum + TOML-auto-detect pattern as ToolCallParser.

Enums

ReasoningFormat

Supported reasoning-block formats. One variant per model family — see the module docs for the delimiter/contract of each.

Traits

ReasoningParser

Parses reasoning/thinking blocks from completed model output.