Streaming Markdown in a chat, token by token

Render Markdown incrementally as tokens arrive, powered by streaming-markdown. It parses only the new text and appends DOM nodes per chunk — no per-token re-parse, no innerHTML rebuild.

npm install @aparte/plugin-streaming-markdown @aparte/core streaming-markdown

@aparte/core and streaming-markdown are peer dependencies.

import { setupStreamingMarkdownProvider } from '@aparte/plugin-streaming-markdown';

setupStreamingMarkdownProvider();

Call it once at startup. It fills the aparteGlobalConfig.setStreamingMarkdownProvider seam, which the chat bubble uses while a message is streaming. That is all you need: a reply renders the same whether it streams or arrives complete — a conversation restored after a reload, a non-streaming provider’s answer — because with no one-shot provider registered, core runs the whole string through this parser too. Pairing it with a one-shot provider like marked is optional — register one and finished and restored messages are rendered through it, at that renderer’s full fidelity.

Security

The streaming path writes DOM nodes directly, so it bypasses the one-shot HTML sanitiser. To keep the same URL policy live, the plugin drops any href/src whose scheme fails core’s isSafeUrl as it streams — a [x](javascript:…) token never becomes a clickable javascript: link, even if the scheme is split across two chunks — and a link that resolves off-site (https://…, //host, a leading space, http:/host) gets target="_blank" rel="noopener noreferrer" live, the rule the sanitiser applies. At completion the settled message goes through the sanitiser in full: re-rendered by the one-shot provider when one is registered, re-sanitised in place otherwise.