Skip to content

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

TypeScript

refactor(mcp-host): measure DOCX and slide text through one body-face measure

marcotebc0824cf9b637
diff --git a/mcp-host/src/workflow/docxTable.ts b/mcp-host/src/workflow/docxTable.ts
index f57febdc8..96842a30b 100644
--- a/mcp-host/src/workflow/docxTable.ts
+++ b/mcp-host/src/workflow/docxTable.ts
@@ -1,4 +1,3 @@
-import { type SKRSContext2D, createCanvas } from '@napi-rs/canvas'
import {
BorderStyle,
Paragraph,
@@ -17,7 +16,7 @@ import {
type DocxPalette,
docxHex,
} from './docxStyle'
-import { CHART_FONT_FAMILY, ensureFontsReady } from './fonts'
+import { CJK_WIDE_CHAR, bodyFaceEm } from './fonts'
import { inlineSpans } from './inlineMarkup'
import { columnNames, headerText } from './tableRows'
@@ -50,34 +49,24 @@ const LONG_TOKEN_TWIPS = 2800
/** Added to every column: Word wraps a word that fills its cell exactly. */
const SLACK_TWIPS = 40
-/**
- * CJK characters, counted one em wide as every CJK face sets them, and the
- * characters Roboto has no glyph for, counted at a generous 0.6 em, so neither
- * depends on the faces the runtime image happens to ship.
- */
-const WIDE_CHAR =
- /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}\u3000-\u303F\uFF01-\uFF60\uFFE0-\uFFE6]/u
+/** Characters Roboto has a glyph for; any other is counted, not measured. */
const MEASURED_CHAR = /[\u0000-\u052F\u1E00-\u206F\u20A0-\u20BF]/u
-let context: SKRSContext2D | undefined
const advances = [new Map<string, number>(), new Map<string, number>()]
/**
* Advance of one character at 1pt in ems. Characters in MEASURED_CHAR are
* measured once each in Roboto, which runs about 7-10% wider than Calibri
* (Medium against Bold included), so a sum of advances errs toward fitting.
+ * CJK counts one em and the rest a generous 0.6, so neither depends on the
+ * faces the runtime image happens to ship.
*/
function advanceEm(ch: string, bold: boolean): number {
- if (!MEASURED_CHAR.test(ch)) return WIDE_CHAR.test(ch) ? 1 : 0.6
+ if (!MEASURED_CHAR.test(ch)) return CJK_WIDE_CHAR.test(ch) ? 1 : 0.6
const cache = advances[bold ? 1 : 0]
const known = cache.get(ch)
if (known !== undefined) return known
- if (!context) {
- ensureFontsReady()
- context = createCanvas(4, 4).getContext('2d')
- }
- context.font = `100px "${bold ? `${CHART_FONT_FAMILY} Bold` : CHART_FONT_FAMILY}"`
- const em = context.measureText(ch).width / 100
+ const em = bodyFaceEm(ch, bold)
cache.set(ch, em)
return em
}
diff --git a/mcp-host/src/workflow/fonts.ts b/mcp-host/src/workflow/fonts.ts
index 4dba6a2ce..881218108 100644
--- a/mcp-host/src/workflow/fonts.ts
+++ b/mcp-host/src/workflow/fonts.ts
@@ -133,6 +133,28 @@ function errorText(err: unknown): string {
return err instanceof Error ? err.message : String(err)
}
+/**
+ * CJK characters, which every CJK face sets one em wide. Text measures count
+ * them so rather than measure them, since the image may have no CJK face.
+ */
+export const CJK_WIDE_CHAR =
+ /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}\u3000-\u303F\uFF01-\uFF60\uFFE0-\uFFE6]/u
+
+let measuring: SKRSContext2D | undefined
+
+/**
+ * Width of `text` in ems of the bundled body face (CHART_FONT_FAMILY), as the
+ * canvas sets it. Each document format scales it to the face it is shown in.
+ */
+export function bodyFaceEm(text: string, bold: boolean): number {
+ if (!measuring) {
+ ensureFontsReady()
+ measuring = createCanvas(4, 4).getContext('2d')
+ }
+ measuring.font = `100px "${bold ? `${CHART_FONT_FAMILY} Bold` : CHART_FONT_FAMILY}"`
+ return measuring.measureText(text).width / 100
+}
+
/**
* Families a PDF can be built from, widest coverage first.
*
diff --git a/mcp-host/src/workflow/pptxText.ts b/mcp-host/src/workflow/pptxText.ts
index 069694e84..4cb085209 100644
--- a/mcp-host/src/workflow/pptxText.ts
+++ b/mcp-host/src/workflow/pptxText.ts
@@ -4,9 +4,8 @@
* another slide, so the generator lays text out itself. It measures with the
* Roboto faces fonts.ts registers, within a few percent of Arial, the deck font.
*/
-import { type SKRSContext2D, createCanvas } from '@napi-rs/canvas'
import { isRtlText } from './docxScript'
-import { CHART_FONT_FAMILY, ensureFontsReady } from './fonts'
+import { CJK_WIDE_CHAR, bodyFaceEm } from './fonts'
/** The one family the deck uses, for slides and charts alike. */
export const PPTX_FONT_FACE = 'Arial'
@@ -27,29 +26,12 @@ const WIDE_LINE_HEIGHT = 1.35
export const TEXT_INSET_X = 0.2
export const TEXT_INSET_Y = 0.1
-/**
- * CJK characters, counted one em wide: every CJK face sets them so, and the
- * runtime image may have no CJK face to measure them with.
- */
-const WIDE_CHAR =
- /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}\u3000-\u303F\uFF01-\uFF60\uFFE0-\uFFE6]/u
-
/** Runs of spaces, single CJK characters, and words: the points where a line may break. */
const BREAK_TOKENS = new RegExp(
- `\\s+|${WIDE_CHAR.source}|[^\\s]+?(?=\\s|${WIDE_CHAR.source}|$)`,
+ `\\s+|${CJK_WIDE_CHAR.source}|[^\\s]+?(?=\\s|${CJK_WIDE_CHAR.source}|$)`,
'gu'
)
-let context: SKRSContext2D | undefined
-
-function canvasContext(): SKRSContext2D {
- if (!context) {
- ensureFontsReady()
- context = createCanvas(4, 4).getContext('2d')
- }
- return context
-}
-
const widthCache = new Map<string, number>()
/** Longest text whose width is kept, and how many widths are kept at once. */
const CACHED_TEXT_LENGTH = 64
@@ -60,15 +42,13 @@ function widthEm(text: string, bold: boolean): number {
const key = `${bold ? 'b' : 'r'}${text}`
const cached = widthCache.get(key)
if (cached !== undefined) return cached
- const ctx = canvasContext()
- ctx.font = `100px "${bold ? `${CHART_FONT_FAMILY} Bold` : CHART_FONT_FAMILY}"`
let wide = 0
let narrow = ''
for (const ch of text) {
- if (WIDE_CHAR.test(ch)) wide++
+ if (CJK_WIDE_CHAR.test(ch)) wide++
else narrow += ch
}
- const measured = narrow ? ctx.measureText(narrow).width / 100 : 0
+ const measured = narrow ? bodyFaceEm(narrow, bold) : 0
const em = measured * (bold ? BOLD_WIDTH_FACTOR : REGULAR_WIDTH_FACTOR) + wide
// Only short texts repeat (words, labels); a whole cell is measured once.
if (text.length <= CACHED_TEXT_LENGTH) {
@@ -143,7 +123,7 @@ export function wrapLines(text: string, sizePt: number, widthIn: number, bold =
}
export function lineHeightIn(sizePt: number, text: string): number {
- return ((WIDE_CHAR.test(text) ? WIDE_LINE_HEIGHT : LINE_HEIGHT) * sizePt) / 72
+ return ((CJK_WIDE_CHAR.test(text) ? WIDE_LINE_HEIGHT : LINE_HEIGHT) * sizePt) / 72
}
/** Height, in inches, of `paragraphs` set in a box `widthIn` wide, without the box insets. */