fix(mcp-server): reject text between tags, which draw.io cannot open

draw.io reads any text inside a page as compressed page data, so a stray
text node makes the whole page fail with an atob error. gpt-5-mini sends
new cells with a literal "\n" between the tags; the edit card said
Complete while draw.io showed the error and kept the old diagram.

Validation now reports text between tags, and auto-fix turns a literal
\n, \t or \r between tags into whitespace. Other text goes back to the
model as an error. The compressed data directly under <diagram> is fine.

packages/mcp-server/src/xml-validation.ts

@@ -404,6 +404,21 @@ function findOrphanMxPoints(
  *   <mxPoint>s. Used for XML the model wrote, not for files or browser state.
  * @returns null if valid, error message string if invalid
  */
+/** The first non-blank text under el, skipping a page's compressed data */
+function findTextBetweenTags(el: Element | null): string | null {
+    if (!el) return null
+    for (const node of Array.from(el.childNodes)) {
+        if (node.nodeType === 1) {
+            const text = findTextBetweenTags(node as Element)
+            if (text) return text
+        } else if (node.nodeType === 3 && el.tagName !== "diagram") {
+            const text = node.textContent?.trim()
+            if (text) return text.slice(0, 40)
+        }
+    }
+    return null
+}
+
 export function validateMxCellStructure(
     xml: string,
     opts: { strict?: boolean } = {},
@@ -429,6 +444,15 @@ export function validateMxCellStructure(
                 return `Invalid XML: Found nested mxCell (id="${id}"). Cells should be siblings, not nested inside other mxCell elements.`
             }
         }
+
+        // draw.io reads any text inside a page as compressed page data and
+        // then fails to open the page
+        if (!doc.querySelector("parsererror")) {
+            const text = findTextBetweenTags(doc.documentElement)
+            if (text) {
+                return `Invalid XML: Found text "${text}" between tags. Labels belong in the value attribute; remove any other text between tags.`
+            }
+        }
     } catch (error) {
         console.warn(
             "[validateMxCellStructure] DOMParser threw unexpected error, falling back to regex validation:",
@@ -542,6 +566,15 @@ export function autoFixXml(xml: string): { fixed: string; fixes: string[] } {
         fixes.push("Fixed JSON-escaped XML")
     }
 
+    // 0b. Literal \n, \t or \r between tags, from escaping the XML twice
+    const unescaped = fixed.replace(/>(?:\s|\\[nrt])+</g, (gap) =>
+        gap.replace(/\\n/g, "\n").replace(/\\t/g, "\t").replace(/\\r/g, ""),
+    )
+    if (unescaped !== fixed) {
+        fixed = unescaped
+        fixes.push("Replaced literal \\n between tags with line breaks")
+    }
+
     // 1. Remove CDATA wrapper
     if (/^\s*<!\[CDATA\[/.test(fixed)) {
         fixed = fixed.replace(/^\s*<!\[CDATA\[/, "").replace(/\]\]>\s*$/, "")

packages/mcp-server/tests/edit-diagram.test.ts

@@ -110,6 +110,19 @@ describe("editDiagram", () => {
         expect(out.ok).toBe(true)
     })
 
+    it("fixes a literal \\n between tags, as gpt-5-mini sends it", () => {
+        const newXml = `<mxCell id="c" value="Reset password" vertex="1" parent="1">\\n  <mxGeometry x="0" y="0" width="80" height="40" as="geometry"/>\\n</mxCell>`
+        const out = editDiagram(
+            DOC,
+            [{ operation: "add", cell_id: "c", new_xml: newXml }],
+            {},
+        )
+        expect(out.ok).toBe(true)
+        if (!out.ok) return
+        expect(out.xml).toContain('value="Reset password"')
+        expect(out.xml).not.toContain("\\n")
+    })
+
     it("reports a missing page as a page-level error", () => {
         const out = editDiagram(DOC, [{ operation: "delete", cell_id: "a" }], {
             page_id: "nope",

packages/mcp-server/tests/xml-validation.test.ts

@@ -212,7 +212,7 @@ describe("autoFixXml keeps valid tags", () => {
 
     it("removes a stray <a> without touching <Array> waypoints", () => {
         const edge = `<mxCell id="e" edge="1" parent="1"><mxGeometry relative="1" as="geometry"><Array as="points"><mxPoint x="1" y="2"/></Array></mxGeometry></mxCell>`
-        const r = validateAndFixXml(model(`<a>x</a>${edge}`))
+        const r = validateAndFixXml(model(`<a></a>${edge}`))
         expect(r.valid).toBe(true)
         expect(r.fixed).toContain('<Array as="points">')
         expect(r.fixed).toContain('<mxPoint x="1" y="2"/>')
@@ -277,3 +277,34 @@ describe("validateAndFixXml strict checks", () => {
         expect(r.fixed).toContain('<mxPoint x="1" y="2"/>')
     })
 })
+
+describe("text between tags", () => {
+    // draw.io reads any text inside a page as compressed data, so the whole
+    // page fails to open with an atob error
+    it("turns a literal \\n between tags into a line break", () => {
+        const cell = `<mxCell id="3" value="Reset password" vertex="1" parent="1">\\n  <mxGeometry x="0" y="0" width="80" height="40" as="geometry"/>\\n</mxCell>`
+        const r = validateAndFixXml(cell)
+        expect(r.valid).toBe(true)
+        expect(r.fixed).toBe(
+            `<mxCell id="3" value="Reset password" vertex="1" parent="1">\n  <mxGeometry x="0" y="0" width="80" height="40" as="geometry"/>\n</mxCell>`,
+        )
+    })
+
+    it("rejects other text between tags", () => {
+        const r = validateAndFixXml(
+            model(
+                `<mxCell id="3" vertex="1" parent="1">Reset password<mxGeometry as="geometry"/></mxCell>`,
+            ),
+        )
+        expect(r.valid).toBe(false)
+        expect(r.error).toMatch(/Reset password/)
+    })
+
+    it("accepts a compressed page", () => {
+        expect(
+            validateAndFixXml(
+                `<mxfile><diagram id="p" name="P">dZHBDoIwDIafhjtsGPWM6MkTB8/LVmBxrGQMQZ/eLRuIUS/bv/VfmybF</diagram></mxfile>`,
+            ).valid,
+        ).toBe(true)
+    })
+})