The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
Rust agent-orchestrationai-agentsai-assistantdesktopllmlocal-firstmcppersonal-aiprivacyrustsecond-braintauri
/**
* The assistant's markdown on `/chat` must render code and maths the way the
* rest of the app already does — `AgentMessageBubble` has had syntax
* highlighting and KaTeX since long before the assistant-ui surface existed,
* and for a while the same message rendered richly in a tool result and as
* flat grey text in the answer above it.
*
* Every assertion here is on RENDERED OUTPUT. Asserting that the plugin array
* has four entries would pass against a component that renders nothing.
*/
import { TextMessagePartProvider } from '@assistant-ui/react';
import { act, render, waitFor } from '@testing-library/react';
import { describe, expect, it } from 'vitest';
import { MarkdownText } from '../markdown-text';
async function renderMarkdown(text: string) {
const { container } = render(
<TextMessagePartProvider text={text} isRunning={false}>
<MarkdownText />
</TextMessagePartProvider>
);
// `defer` hands parsing to a lower-priority render.
await waitFor(() => expect(container.querySelector('.aui-md')?.innerHTML).toBeTruthy());
return container;
}
describe('MarkdownText — syntax highlighting', () => {
it('updates an indented code block as streamed text grows and completes', async () => {
const prefix =
Array.from(
{ length: 80 },
(_, i) => `Filler line ${i + 1} — autogenerated to grow the scroll column.\n`
).join('') + '\n**BOLD-CANARY** is bold.\n\n';
const view = (text: string, isRunning: boolean) => (
<TextMessagePartProvider text={text} isRunning={isRunning}>
<MarkdownText />
</TextMessagePartProvider>
);
const { container, rerender } = render(view(`${prefix} C`, true));
await waitFor(() => expect(container.querySelector('pre code')?.textContent).toContain('C'));
const complete = `${prefix} CODE-CANARY-93b1\n line 2\n\nVisit [the docs](https://example.com/canary) for more.`;
for (let end = prefix.length + 5; end <= complete.length; end += 10) {
await act(async () => {
rerender(view(complete.slice(0, end), true));
await new Promise(resolve => setTimeout(resolve, 10));
});
}
rerender(view(complete, false));
await waitFor(() => {
expect(container.querySelector('pre code')?.textContent).toContain('CODE-CANARY-93b1');
expect(container.querySelector('pre code')?.textContent).toContain('line 2');
});
});
it('tokenises a fenced block that declares a language', async () => {
const container = await renderMarkdown('```ts\nconst x = 1;\n```');
const code = container.querySelector('code');
expect(code?.className).toContain('hljs');
// The specific token matters: `hljs` alone is on the wrapper even when
// nothing was tokenised, so assert a keyword span actually carrying "const".
const keyword = container.querySelector('code .hljs-keyword');
expect(keyword).not.toBeNull();
expect(keyword?.textContent).toBe('const');
});
it('keeps the language label and the copy button on a highlighted block', async () => {
// The regression this guards: `rehypeHighlight` turns the <code> element's
// string child into token spans, which routes the kit past the branch that
// renders its `CodeHeader` slot. Highlighting a block must not cost it its
// header.
const container = await renderMarkdown('```ts\nconst x = 1;\n```');
expect(container.querySelector('.aui-code-header-root')).not.toBeNull();
expect(container.querySelector('.aui-code-header-language')?.textContent).toBe('ts');
expect(container.querySelector('code .hljs-keyword')).not.toBeNull();
});
it('still renders a header for a fence with no language', async () => {
const container = await renderMarkdown('```\nplain text\n```');
expect(container.querySelector('.aui-code-header-root')).not.toBeNull();
expect(container.querySelector('pre')?.textContent).toContain('plain text');
});
it('leaves inline code unhighlighted', async () => {
const container = await renderMarkdown('use `const` here');
const inline = container.querySelector('code');
expect(inline?.textContent).toBe('const');
expect(inline?.className).toContain('aui-md-inline-code');
expect(container.querySelector('.hljs-keyword')).toBeNull();
});
});
describe('MarkdownText — LaTeX', () => {
it('renders display math written with $$', async () => {
const container = await renderMarkdown('$$\\frac{a}{b}$$');
// KaTeX emits `.katex`; its absence means the `$$` reached the DOM as text.
expect(container.querySelector('.katex')).not.toBeNull();
expect(container.textContent).not.toContain('$$');
});
it('renders math emitted with backslash delimiters', async () => {
// What models actually emit. `remark-math` understands only `$…$`, so this
// passes only if the text is normalised before parsing.
const container = await renderMarkdown('The value \\(\\frac{a}{b}\\) is small.');
expect(container.querySelector('.katex')).not.toBeNull();
// KaTeX keeps the original TeX in an <annotation>, so the source survives in
// textContent by design. What must NOT survive is the delimiter pair — if
// normalisation had not run, `\(` and `\)` would still be sitting in the
// prose, which is exactly the bug a reader sees.
expect(container.textContent).not.toContain('\\(');
expect(container.textContent).not.toContain('\\)');
});
it('does NOT treat prices as math', async () => {
// The reason math is gated rather than always on: `remark-math` would read
// "$10 vs $20" as one inline formula and swallow the words between them.
const container = await renderMarkdown('It costs $10 vs $20 elsewhere.');
expect(container.querySelector('.katex')).toBeNull();
expect(container.textContent).toContain('$10 vs $20');
});
});
describe('MarkdownText — GFM is unaffected', () => {
it('still renders tables', async () => {
const container = await renderMarkdown('| a | b |\n| - | - |\n| 1 | 2 |');
expect(container.querySelector('table')).not.toBeNull();
expect(container.querySelectorAll('td')).toHaveLength(2);
});
});