rich conversation

TimelineChat renders agent-authored prose as safe, compact Markdown with KaTeX mathematics through one small renderer boundary.

raw source

The terminal-free conversation should render the formats coding agents naturally write: Markdown for structured answers and LaTeX for mathematics. Rich output must remain readable and safe when the agent emits malformed, hostile, or very large content, without turning TimelineChat into a second rendering framework.

expanded spec

One RichText component is the whole rendering boundary. TimelineChat sends every human-authored prose field through it: the originating prompt, sent messages, and authored status notes. One module-scoped parser handles them all; the timeline never constructs a parser per row and never grows its own Markdown branches.

The supported language is compact agent Markdown: headings, emphasis, links, images (including remote image URLs), blockquotes, ordered and unordered lists, fenced and inline code, tables, strikethrough, and soft line breaks. KaTeX renders inline and display math written with $...$ / $$...$$ or \(...\) / \[...\]. Math inside code stays code. An invalid expression stays visibly readable rather than blanking the message or throwing through React. Inline math never crosses a code span or line break, so ordinary currency, shell variables, and escaped dollar signs remain prose.

The renderer keeps markdown-it and KaTeX's default untrusted-source behavior: source HTML is text and unsafe URL schemes do not become active content. It adds no sanitizer, image allowlist, remote-content policy, or other custom content filter. Generated HTML has one audited insertion point inside RichText; callers never use dangerouslySetInnerHTML themselves.

Rendered mathematics remains ordinary selectable conversation content. KaTeX's accessibility and visual DOM representations identify one owning math token. A selection intersecting any portion of that token copies the formula atomically as its complete authored TeX source exactly once, rather than exposing a glyph-level rule tied to KaTeX internals or concatenating MathML, annotation, and visual fallback text. All non-math portions keep the browser Range's native text serialization, so this does not become a second prose parser or change the conversation's focus-preserving custom-highlight selection model.

Layout remains chat-dense. Markdown blocks remove browser-default outer margins, images shrink to the message width while retaining their aspect ratio, code preserves whitespace and scrolls horizontally, tables and display equations scroll within the message instead of widening the session pane, and long unbroken prose still wraps. KaTeX's stylesheet rides the lazy TimelineChat module, so graph and review pages do not download math styling. No syntax highlighter, HTML sanitizer, general math plugin, Markdown editor, or second message model is added.