HTML-to-Word still showed the generic convert error after the walker fix because DOMPurify’s HTML profile deleted SVG/MathML text and any Packer/image throw aborted the whole document.

Date: 2026-09-07 21:30 Summary: HTML-to-Word still showed the generic convert error after the walker fix because DOMPurify’s HTML profile deleted SVG/MathML text and any Packer/image throw aborted the whole document. Visibility: people

[question] https://onlinefreetools.org/zh/tools/convert-html-web-pages-to-word-document 这段 HTML 没能生成 Word 文件，请检查标记后再试。

Tool links (English)

convert-html-web-pages-to-word-document

[try to solve]

Context

The Chinese Convert HTML web pages to a Word document page still showed tool_convert_html_web_pages_to_word_document_err_convert (“这段 HTML 没能生成 Word 文件，请检查标记后再试。”) after the earlier header/walker fix was already live (processNode + PAGES_CACHE_VERSION 4.56). Load sample succeeded on production, so the docx vendor and the sample heading/list/table path were not the failure.

Process

Fetched the live zh page with Accept: text/html (without that header the Worker returns 404). Confirmed the recursive mapper was in the HTML (document.body walk, SKIP, processNode). Playwright against production: sample, header-wrapped documents, Wikipedia URLs, nested tables, and Word-exported HTML all enabled Download.

The same generic string is used for two different throws: htmlToBlocks → Error('empty') when zero blocks, and any later exception (Document / Packer.toBlob). Mapped cases that still failed on the live mapper:

SVG with <text> and MathML: DOMPurify.sanitize(..., { USE_PROFILES: { html: true } }) returned "" (profile strips those trees including their text).
Image-only markup with alt but a non-http src, or CORS-blocked fetch(src, { mode: 'cors' }): no blocks, same error.
:scope-unsafe nested querySelectorAll('tr') was already tolerated for simple nested tables; new Table() was still able to abort the whole Promise chain if it threw.

Local prerender after the fix: SVG text, MathML, <form> wrappers, image alt fallback, and sample all produced a .docx. Truly empty <div><span></span></div> still errors, which is correct.

Root cause / analysis

The 18:36 walker fix made ordinary HTML documents work. Real paste still died in two ways that look identical to the user:

Sanitize too strict. The HTML profile is meant for a safe subset. It removes svg and math entirely (KEEP_CONTENT does not preserve that text). A chart, formula, or SVG export then maps to zero blocks, and the catch shows “check the markup” even though the source had readable text.
One node fails, whole Convert fails. Images were fetched with browser CORS (most CDNs fail), always stamped type: 'png', and a thrown Table / Packer rejected the only Promise. Text that had already been mapped never reached Download.

Error('empty') was not mapped to the empty-input string; it used the same convert error as Packer, so “no blocks” and “library threw” were indistinguishable.

Solution

Loosen sanitize: forbid script / iframe / object / embed / frame / frameset only, and explicitly allow SVG/MathML/content tags. Walk form / svg / math / template.content; treat br/hr as empty paragraphs; use alt/title when an image cannot be embedded. Fetch remote images through /api/tools/convert-html-to-pdf/asset (same proxy as HTML-to-PDF), sniff PNG/JPEG/GIF/BMP, skip other bytes. Wrap table mapping; if Packer.toBlob throws with images on, retry without images; if blocks are still empty, strip tags from the raw HTML for a plain-text fallback. Strip XML 1.0 illegal characters from runs.

Verified in Chrome against the prerendered zh HTML: sample, SVG text, MathML, header, form, image alt, and text+failed image all enable Download.

Bumped PAGES_CACHE_VERSION to 4.57 so the Worker Cache API key changes after git push (HTML s-maxage=86400).

Notes / boundaries

Truly empty markup (no text, no alt, no tags that survive) still shows the convert error. CSS layout is still not mapped. URL fetch of this site’s own hostname can 400 from the HTML fetch API (blocked/non-HTML handling); that path uses the URL error string, not the convert string. Production tool HTML updates only after upload:r2 plus a Worker deploy that picks up 4.57.

[actions]

Updated src/pages/convertHtmlWebPagesToWordDocumentPage.ts sanitize, image proxy, table/Packer fallbacks.
Bumped catalog updatedAt; prerendered and gzipped this slug.
Bumped wrangler.jsonc PAGES_CACHE_VERSION to 4.57.