ToolCallParser

gglib_core::normalize::parser

Trait ToolCallParser 

Source 
pub trait ToolCallParser: Send {
    // Required methods
    fn push_text(&mut self, chunk: &str) -> ParserOutput;
    fn push_reasoning(&mut self, chunk: &str) -> ParserOutput;
    fn finish(&mut self) -> ParserOutput;
}
Expand description

Stream-stateful parser that normalizes a single LLM dialect into canonical ParserOutput fragments.

Implementations MUST:

  • Be chunk-safe: dialect markers may straddle chunk boundaries, so the parser must internally buffer ambiguous trailing bytes and flush them on a later call or at Self::finish.
  • Be deterministic: feeding the same byte sequence in any chunking yields the same total output (modulo when individual bytes flush).
  • Never lose input bytes: every byte of input is either forwarded verbatim, consumed as part of a recognised marker, or surfaced via a NormalizationError.

Implementations are NOT required to be Send here — the NormalizingStream adapter erases the parser through Box<dyn …> and adds the Send bound at that boundary.

Required Methods§

Source

fn push_text(&mut self, chunk: &str) -> ParserOutput

Feed a chunk that arrived on the upstream text channel.

Source

fn push_reasoning(&mut self, chunk: &str) -> ParserOutput

Feed a chunk that arrived on the upstream reasoning channel.

Source

fn finish(&mut self) -> ParserOutput

Flush any buffered partial state at end-of-stream.

Called exactly once per stream, after the last push_* call and before the surrounding Done event is forwarded downstream. Implementations should emit any held-back bytes as text and surface any unfinished marker state as an NormalizationError.

Implementors§