Reshape the stream while it's still arriving.

The problem

Reasoning models don't hand you clean text.

They interleave a hidden thought-stream with the answer, stutter into repetition, and emit messy whitespace — all arriving token by token.

A black-box response opens only at the end. A stream you can shape per chunk.

01 Separate

Thinking on its own channel.

Every chunk is routed to answer or thinking. Show the reasoning, style it, or hide it — the answer stays clean either way.

Google Vertex parts[].thought  ·  OpenRouter delta.reasoning

02 Collapse

Repetition loops, gone.

Near-duplicate lines — a common stutter artifact — are detected with midstream's temporal-compare edit-distance similarity, then dropped.

Real crate: midstreamer-temporal-compare (DTW / LCS / EditDistance).

03 Normalize

Whitespace, tidied.

Runs of spaces and stray blank lines collapse to one clean pass — per completed line, as it streams.

Three dialects, one stream

Write your handler once.

Google Vertex, OpenRouter, and meta-llm each speak a different SSE dialect. Adapters normalize every one into a single neutral chunk stream.

How it works

One streaming pass. Nothing buffered.

It never waits for the whole response. Each chunk is transformed the moment it lands — then handed straight to your UI.

Get started

Drop it into your stream loop.

const { Reformatter } = require('llm-stream-reformat');

const rf = new Reformatter();
for (const line of sseLines) {          // your provider's raw data: {...} lines
  for (const ev of rf.pushSse('google', line)) {
    if (ev.channel === 'thinking') showReasoning(ev.text);
    else appendAnswer(ev.text);
  }
}
rf.finish().forEach(ev => appendAnswer(ev.text)); // flush the tail

import { init, Reformatter } from 'llm-stream-reformat/web';

await init();                     // auto-fetches the wasm
const rf = new Reformatter();
// then push provider lines exactly as above:
for (const ev of rf.pushSse('openrouter', line)) { /* … */ }

use stream_reformat::{Reformatter, Provider, Channel};

let mut rf = Reformatter::new();
for line in sse_lines {
    for ev in rf.push_sse(Provider::Google, &line) {
        match ev.channel {
            Channel::Thinking => show_reasoning(&ev.text),
            Channel::Answer   => append_answer(&ev.text),
        }
    }
}

Each event is { channel: 'answer' | 'thinking', text }. Providers: 'google', 'openrouter', 'metallm'.