Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

fix(mcp-host): lay out a PDF paragraph of many runs in linear time

marcoteb9b9c731d4090
diff --git a/mcp-host/src/workflow/__tests__/internalTools.pdf.layout.test.ts b/mcp-host/src/workflow/__tests__/internalTools.pdf.layout.test.ts
index 65456d3e9..bd28268c3 100644
--- a/mcp-host/src/workflow/__tests__/internalTools.pdf.layout.test.ts
+++ b/mcp-host/src/workflow/__tests__/internalTools.pdf.layout.test.ts
@@ -721,6 +721,30 @@ describe('PDF input that would stall the host', () => {
expect(pastMargin(pages)).toEqual([])
}, 60_000)
+ it('lays out a paragraph, list item, quote and cell of many styled runs', async () => {
+ // 100,000 runs of alternating emphasis in each; pdfmake alone takes about
+ // half a minute on one such line when it is laid out as one paragraph.
+ const line = '*a'.repeat(100_000)
+ const started = performance.now()
+ const r = await tool.execute(
+ {
+ filename: 'runs.pdf',
+ body: `${line}\n\n- ${line}\n\n> ${line}`,
+ tables: [{ headers: ['A'], rows: [[line]] }],
+ },
+ outputDir
+ )
+ expect(r.success, r.error).toBe(true)
+ expect(performance.now() - started).toBeLessThan(budgetMs)
+ }, 120_000)
+
+ it('keeps every word of a paragraph cut into pieces', async () => {
+ const { pages } = await render({ filename: 'pieces.pdf', body: '*a b* c '.repeat(1_500) })
+ const text = allText(pages)
+ expect(text.match(/c/g)).toHaveLength(1_500)
+ expect(text.match(/a/g)).toHaveLength(1_500)
+ }, 60_000)
+
it('lays out a long run of words joined by no-break spaces', async () => {
const started = performance.now()
const { pages } = await render({ filename: 'nb.pdf', body: 'a\u00a0'.repeat(20_000) })
diff --git a/mcp-host/src/workflow/internalTools.ts b/mcp-host/src/workflow/internalTools.ts
index 1ff15c347..54317db1b 100644
--- a/mcp-host/src/workflow/internalTools.ts
+++ b/mcp-host/src/workflow/internalTools.ts
@@ -1689,6 +1689,68 @@ function parseInlineMarkdown(line: string, onImage: (image: MarkdownImage) => vo
return { text: runs }
}
+/** Characters, and styled runs, one pdfmake paragraph holds before the next piece starts. */
+const PARAGRAPH_PIECE_CHARS = 20_000
+const PARAGRAPH_PIECE_RUNS = 1_000
+
+/**
+ * `paragraph` in pieces of up to PARAGRAPH_PIECE_CHARS characters and
+ * PARAGRAPH_PIECE_RUNS runs, each cut after a space where there is one.
+ * pdfmake lays a paragraph out in time that grows with the square of its words
+ * and runs, so one 200 KB line of alternating emphasis held the event loop for
+ * half a minute; pieces set one under another read on as the same paragraph,
+ * with a line ending at each cut. No paragraph a person writes reaches either
+ * limit.
+ */
+function inPieces(paragraph: ContentText): ContentText[] {
+ if (
+ !Array.isArray(paragraph.text) ||
+ (paragraph.text.length <= PARAGRAPH_PIECE_RUNS &&
+ plainText(paragraph).length <= PARAGRAPH_PIECE_CHARS)
+ ) {
+ return [paragraph]
+ }
+ const pieces: ContentText[][] = [[]]
+ let size = 0
+ for (const run of paragraph.text as ContentText[]) {
+ if (pieces[pieces.length - 1].length >= PARAGRAPH_PIECE_RUNS) {
+ pieces.push([])
+ size = 0
+ }
+ let text = String(run.text ?? '')
+ while (size + text.length > PARAGRAPH_PIECE_CHARS) {
+ const room = PARAGRAPH_PIECE_CHARS - size
+ let cut = text.lastIndexOf(' ', room - 1) + 1
+ if (cut <= 0 && size > 0) {
+ pieces.push([])
+ size = 0
+ continue
+ }
+ if (cut <= 0) {
+ // No space to cut after: cut inside the word, never inside a surrogate pair.
+ cut = room
+ const code = text.charCodeAt(cut - 1)
+ if (code >= 0xd800 && code <= 0xdbff) cut--
+ }
+ pieces[pieces.length - 1].push({ ...run, text: text.slice(0, cut) })
+ pieces.push([])
+ size = 0
+ text = text.slice(cut)
+ }
+ if (text) {
+ pieces[pieces.length - 1].push({ ...run, text })
+ size += text.length
+ }
+ }
+ return pieces.filter(piece => piece.length > 0).map(piece => ({ text: piece }))
+}
+
+/** `parsed` as it is, or as a stack of its pieces when it is that long. */
+function asPieces(parsed: ContentText): ContentText | { stack: ContentText[] } {
+ const pieces = inPieces(parsed)
+ return pieces.length === 1 ? pieces[0] : { stack: pieces }
+}
+
/** The text a parsed line prints, for measuring it. */
function plainText(parsed: ContentText): string {
const text = parsed.text
@@ -1791,7 +1853,7 @@ function buildTableNode(
widths: fit.widths,
body: [
headerCells.map((h, col) => ({
- ...h,
+ ...asPieces(h),
...(col === 0 ? { id: `${TABLE_HEADER_ID}${++env.tablesBuilt}` } : {}),
bold: true,
color: '#ffffff',
@@ -1800,7 +1862,7 @@ function buildTableNode(
})),
// pdfmake accepts a `text` array of inline runs as a cell — use that
// so **bold** / *italic* / `code` inside cells render correctly.
- ...cells.map(row => row.map((cell, col) => ({ ...cell, ...align(col) }))),
+ ...cells.map(row => row.map((cell, col) => ({ ...asPieces(cell), ...align(col) }))),
],
},
layout: pdfTableLayout(layout, palette),
@@ -1924,7 +1986,7 @@ function buildList(
const isOrdered = ORDERED_RE.test(lines[i].trimStart())
// Switching between bullets and numbers at the same indent starts a different list.
if (isOrdered !== ordered) break
- items.push(parseInlineMarkdown(stripListMarker(lines[i]), onImage))
+ items.push(asPieces(parseInlineMarkdown(stripListMarker(lines[i]), onImage)))
i++
}
@@ -1984,7 +2046,7 @@ function bodyToContent(body: string, palette: PdfPalette, env: PdfBodyEnv): Cont
const style = level === 1 ? 'h1' : level === 2 ? 'h2' : 'h3'
const top = level === 1 ? 16 : level === 2 ? 12 : 8
out.push({
- ...parseInlineMarkdown(withoutClosingHashes(heading[2]), stray),
+ ...asPieces(parseInlineMarkdown(withoutClosingHashes(heading[2]), stray)),
style,
headlineLevel: 1,
margin: [0, top, 0, level === 1 ? 6 : 4],
@@ -2065,7 +2127,12 @@ function bodyToContent(body: string, palette: PdfPalette, env: PdfBodyEnv): Cont
// Default: paragraph with inline markdown, then any images it references.
const images: MarkdownImage[] = []
const paragraph = parseInlineMarkdown(line, image => images.push(image))
- if (plainText(paragraph).trim() !== '') out.push({ ...paragraph, margin: [0, 0, 0, 4] })
+ if (plainText(paragraph).trim() !== '') {
+ const pieces = inPieces(paragraph)
+ pieces.forEach((piece, p) =>
+ out.push({ ...piece, margin: [0, 0, 0, p === pieces.length - 1 ? 4 : 0] })
+ )
+ }
for (const image of images) {
const block = env.image(image.src)
if (block) out.push(block)
@@ -2128,10 +2195,13 @@ function buildBlockquote(
body: [
[
{
- stack: paragraphs.map((text, n) => ({
- ...parseInlineMarkdown(text, onImage),
- margin: [0, 0, 0, n < paragraphs.length - 1 ? 4 : 0],
- })),
+ stack: paragraphs.flatMap((text, n) => {
+ const pieces = inPieces(parseInlineMarkdown(text, onImage))
+ return pieces.map((piece, p) => ({
+ ...piece,
+ margin: [0, 0, 0, p === pieces.length - 1 && n < paragraphs.length - 1 ? 4 : 0],
+ }))
+ }),
italics: true,
color: palette.muted,
margin: [10, 4, 6, 4],