Struct StreamTranslator 

pub struct StreamTranslator { /* private fields */ }

Pure streaming translator state (DW-077).

Implementations

impl StreamTranslator

pub fn new(id: String, model_alias: String, created: u64) -> Self

New translator stamping chunks with id (the response id), model_alias (what the client asked for — never the provider model), and created (unix seconds, fixed at stream start).

pub fn feed( &mut self, bytes: &[u8], adapter: &dyn ProviderAdapter, ) -> (Vec<ClientFrame>, usize, bool)

Feed provider bytes; returns the client frames for every COMPLETE provider frame these bytes complete (may be empty — bytes that only continue a partial frame). Also reports how many delta chunks were forwarded and whether one of them was the FIRST forwarded chunk of the stream (the first-token signal).

pub fn flush_partial( &mut self, adapter: &dyn ProviderAdapter, ) -> (Vec<ClientFrame>, usize, bool)

Flush a final provider frame not terminated by a blank line (tolerant providers exist; well-formed streams yield nothing here — the body’s clean end drives Self::finish).

pub fn finish(&mut self) -> Vec<ClientFrame>

The terminal frames for a CLEAN stream end: the accumulated usage chunk (when the provider reported any usage) followed by the gateway’s own data: [DONE]. Returns them once; later calls return empty (idempotent tail).

pub fn abort_tail(&mut self, message: &str) -> Vec<ClientFrame>

The terminal frames for a MID-STREAM ABORT (the provider body errored after frames were forwarded): an OpenAI-shaped error chunk then the terminator. Already-forwarded content stands.

pub fn is_ended(&self) -> bool

Whether the provider signalled end-of-stream (its terminator or the body’s clean end already consumed).

pub fn usage(&self) -> Usage

The accumulated provider-reported usage so far.

pub fn id(&self) -> &str

The response id stamped on every client chunk (AI-07: the gen_ai.response.id value for streaming, where the gateway synthesizes the id the provider does not send as one frame).