Skip to content

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

TypeScript

fix(mcp-host): read the package XML in one pass instead of stripping tags

marcoteb5354c6ab9d96
diff --git a/mcp-host/src/__tests__/artifacts/officePackageText.test.ts b/mcp-host/src/__tests__/artifacts/officePackageText.test.ts
index 4e25822a2..842b60d32 100644
--- a/mcp-host/src/__tests__/artifacts/officePackageText.test.ts
+++ b/mcp-host/src/__tests__/artifacts/officePackageText.test.ts
@@ -49,6 +49,16 @@ describe('officePackageHasText', () => {
expect(officePackageHasText(xlsx, holds)).toBe(true)
})
+ it.each([
+ ['an attribute value holding a >', `<a v="x > ${needle}"/>`],
+ ['character data', `<t><![CDATA[${needle}]]></t>`],
+ ['a comment', `<a/><!-- ${needle} -->`],
+ ['a value left open', `<a v="${needle}`],
+ ])('reads %s', async (_label, xml) => {
+ const part = await pack({ 'xl/worksheets/sheet1.xml': xml })
+ expect(officePackageHasText(part, holds)).toBe(true)
+ })
+
it('decodes character references', async () => {
const docx = await pack({ 'word/document.xml': '<w:t>&#51;141592&#x36;53 &amp; more</w:t>' })
expect(officePackageHasText(docx, holds)).toBe(true)
diff --git a/mcp-host/src/artifacts/officePackageText.ts b/mcp-host/src/artifacts/officePackageText.ts
index f0a41d6c5..d6c72b283 100644
--- a/mcp-host/src/artifacts/officePackageText.ts
+++ b/mcp-host/src/artifacts/officePackageText.ts
@@ -23,8 +23,6 @@ const CENTRAL_DIRECTORY_ENTRY = 0x02014b50
const LOCAL_FILE_HEADER = 0x04034b50
const TEXT_PART = /\.(?:xml|rels|vml)$/i
const EMBEDDED_PACKAGE = /\.(?:docx|docm|xlsx|xlsm|pptx|pptm)$/i
-const TAG = /<[^>]*>/g
-const ATTRIBUTE_VALUE = /\s[^\s=<>"'/]+\s*=\s*(?:"([^"]*)"|'([^']*)')/g
const CHARACTER_REFERENCE = /&(?:#(\d{1,7})|#[xX]([0-9a-fA-F]{1,6})|(lt|gt|amp|quot|apos));/g
const NAMED_REFERENCES: Record<string, string> = {
lt: '<',
@@ -81,14 +79,51 @@ function searchPackage(
return false
}
+/** Sections whose content is read as it stands, up to their end mark. */
+const RAW_SECTIONS: ReadonlyArray<readonly [start: string, end: string]> = [
+ ['<![CDATA[', ']]>'],
+ ['<!--', '-->'],
+]
+
+/**
+ * Whether one XML part's text satisfies `matches`, read in a single pass: its
+ * text nodes, character data and comments run together, and each quoted
+ * attribute value on its own, so a '>' a value holds does not end its tag. A
+ * value or section left open runs to the end of the part, which is read too.
+ */
function xmlHasText(xml: string, matches: (text: string) => boolean): boolean {
- if (matches(decodeReferences(xml.replace(TAG, '')))) return true
- // One pass over the whole part: text that merely looks like an attribute is
- // text the first check already read, so reading it again costs nothing.
- for (const attribute of xml.matchAll(ATTRIBUTE_VALUE)) {
- if (matches(decodeReferences(attribute[1] ?? attribute[2] ?? ''))) return true
+ const text: string[] = []
+ let at = 0
+ while (at < xml.length) {
+ const open = xml.indexOf('<', at)
+ if (open < 0) {
+ text.push(xml.slice(at))
+ break
+ }
+ text.push(xml.slice(at, open))
+ const raw = RAW_SECTIONS.find(([start]) => xml.startsWith(start, open))
+ if (raw) {
+ const close = xml.indexOf(raw[1], open + raw[0].length)
+ const end = close < 0 ? xml.length : close
+ text.push(xml.slice(open + raw[0].length, end))
+ at = end + raw[1].length
+ continue
+ }
+ let i = open + 1
+ while (i < xml.length && xml[i] !== '>') {
+ const quote = xml[i]
+ if (quote !== '"' && quote !== "'") {
+ i++
+ continue
+ }
+ const close = xml.indexOf(quote, i + 1)
+ const end = close < 0 ? xml.length : close
+ if (matches(decodeReferences(xml.slice(i + 1, end)))) return true
+ i = end + 1
+ }
+ at = i + 1
}
- return false
+ return matches(decodeReferences(text.join('')))
}
function decodeReferences(text: string): string {