Module normalize 

Universal local-LLM consistency layer.

This module rewrites model-specific output dialects into the strict OpenAI-shaped crate::domain::agent::LlmStreamEvent sequence that the rest of the codebase expects. Adapters wrap the LLM stream once at the port boundary; every downstream surface (Axum, CLI, Tauri, proxy) consumes the canonical form.

Module map

• tags — format:* constants used to pick a parser.
• error — non-fatal error::NormalizationError surfaced from parsers.
• parser — the parser::ToolCallParser trait + parser::ParserOutput.
• parsers — concrete parser implementations, one file per dialect.
• registry — the single dispatch site that maps tags to parsers.

Adding a new dialect

1. Add a pub const FORMAT_* to tags.
2. Drop a new file under parsers.
3. Add one match arm to registry::get_parser.

The registry is the only place that knows the full set of parsers, by design — see the module docs there.

Re-exports

pub use error::NormalizationError;

pub use error::NormalizationErrorKind;

pub use history::strip_thinking_debt;

pub use parser::ParserOutput;

pub use parser::ToolCallParser;

pub use registry::get_parser;

pub use stream::NormalizingStream;

Modules

error

Non-fatal error reporting from normalization parsers.

history

Cross-turn “thinking debt” removal for chat history.

parser

The ToolCallParser trait and its companion ParserOutput.

parsers

Submodule index for concrete super::parser::ToolCallParser implementations.

registry

Tag-driven parser dispatch.

stream

NormalizingStream — the single wrap point that canonicalises an LLM event stream.

tags

Format-tag constants used to select a normalization parser.