Moving it out of the build left it with no way in. Settings can only install from the community catalogue, so a first-party extension that isn't bundled has a working update channel and no first copy for anyone to update — the mechanism was all there and the front door was missing. So it goes back beside screenshot, aegis and pdf-editor: seeded into every profile by the build, listed under "Built into Theseus", and kept current between releases by the operator-signed channel at theseus.x/extensions/docx-editor/. That is the arrangement docs/ADDON-UPDATES.md describes, and the one the signing script was written for. About 400 KB compressed in the installer, most of it the vendored editor libraries — next to the ~4 MB of pdf.js that pdf-editor already ships, the weight argument for keeping it out didn't survive contact with the numbers. The end-to-end driver goes back to checking that a fresh profile seeds it, which is the property that actually matters now.
6.9 KiB
What survives a round trip
The editor rebuilds a document's body from what you see on screen and carries the rest of the original package across. This is the ledger of what that costs.
Measured over 66 real-world Word documents found on a working machine (CVs, contracts, invoices, forms, letters — Greek, Russian, German and English):
clean round-trip : 65
content drift : 1 (a 7 MB WMF picture, see "Dropped")
invalid package : 0
threw : 0
"Clean" means: read the file, save it, read it again, and the two editor documents are identical — same blocks, same attributes, same marks on the same text. The saved package is also checked for well-formed XML, no dangling relationship ids and no relationships pointing at parts that aren't there.
Reproduce with:
cd addon-build/docx-editor
npm install && npm run build
node test/roundtrip.mjs # the built-in fixture
node test/corpus.mjs <a folder of .docx> # a real corpus
How the two halves work
Reading uses mammoth, but not its
HTML. mammoth's converter is deliberately semantic, and HTML has nowhere to put
a run's colour or a paragraph's line spacing, so it drops them. We take its
parsed document model instead, through the public transformDocument hook,
and walk that into the editor's model. See addon-build/docx-editor/patches.mjs
for the six properties we taught that model to carry: run colour, paragraph
spacing, paragraph bottom border, image display size, numbering format and
numbering id.
Writing uses docx, which always builds a
brand-new package. Anything living outside the document body would therefore
vanish, so lib/pkg.js grafts it back: headers, footers, footnotes, endnotes,
the style catalogue, the theme and the page setup, re-wiring relationship ids
and content types as it goes.
Kept
| How | |
|---|---|
| Headers and footers | The parts are copied across with their own relationships and images, and re-referenced from the new sectPr. |
| Footnotes and endnotes | The markers survive in the body as their own node; footnotes.xml is copied wholesale so the ids still match. |
| Page size, orientation, margins, gutter, title page | Read off the first sectPr and handed to the builder. |
| The document's styles | The original styles.xml is merged over the builder's. Where both define a style id the original wins — it is what the document actually looked like. docDefaults comes across too, so unstyled paragraphs don't shift. |
| Theme, fonts | theme1.xml is copied. |
| Title, author, subject, keywords | From docProps/core.xml. |
| Bold, italic, underline, strike, super/subscript, all-caps, small-caps | |
| Font family, size, colour, highlight | Highlight is Word's 15-value enum, not a hex colour, so it passes through exactly. |
| Alignment, indent, line spacing, space before/after | |
| Headings 1–6, quotes, code blocks | Code blocks ride on a SourceCode paragraph style. |
| Bulleted and numbered lists, nested, with their numbering format | Format is kept per level, so a list that is decimal at the top and lettered underneath stays that way. |
| Where one list ends and the next begins | Tracked by Word's numId, so a second list still restarts at 1. |
| Tables, including merged cells | Both directions. A 12-row vertical merge comes back as a 12-row vertical merge. |
| Images | At the size Word was displaying them, to EMU precision, not the file's natural size. |
| Links, internal anchors | |
| Page breaks, horizontal rules | A rule is Word's empty paragraph with a bottom border, and is written back as one. |
Dropped
These are detected when the file opens and named in a banner before any
editing, and again in the About dialog. The original file on disk is never
overwritten — a save downloads <name>-edited.docx.
- Tracked changes. mammoth renders insertions as ordinary text and drops deletions, so a save would silently accept every pending revision. A document with them opens read-only until you explicitly choose "Accept all and edit".
- Comments. Same gate as tracked changes.
- Equations (OMML), shapes, text boxes and WordArt, content controls.
- Fields — page numbers, tables of contents, cross-references. The text Word last calculated is kept; the field code that would recalculate it is not.
- Bookmarks.
- Section breaks and multi-column layout. Only the first section's page setup is kept.
- Metafile pictures (WMF/EMF). Word's vector picture format: browsers can't display it and the builder can't write it. This is the single drift in the corpus above — one CV with a 7 MB WMF.
Kept, but not exactly
- Paragraph borders and shading. Only the rule under an empty paragraph round-trips. A box around a paragraph, or a shaded paragraph, is lost.
- Custom tab stops. Tab characters are kept; the stop positions are not.
- Exact line spacing.
atLeastandexactline rules are read but the editor has no control for them, so they are written back as-is only when the paragraph is untouched. - Table borders. mammoth doesn't report the borders it read, so every table is written with a plain single-line border. A borderless table gains lines.
- List indentation depth. A list Word started at level 2 with no level 0 or 1 above it becomes a top-level list, and is written at level 0.
- An empty paragraph after a table inside a cell. OOXML forbids a cell that ends with a table, so every such document carries a paragraph the author never typed. It is dropped on read and put back on write.
Decisions worth knowing about
Why not preserve the dropped features as raw XML? docx can embed raw
OOXML (ImportedXmlComponent), so it is technically possible. What makes it
expensive is position: mammoth silently discards the elements it can't model
and reports no location for them, so anchoring a passthrough node in the right
place needs a second OOXML reader running alongside mammoth purely to recover
block order. That is a large amount of machinery whose failure mode is a subtly
corrupt package, which is worse than an honest warning. The grafting approach
gets headers, footers, notes, page setup and styles — the things most real
documents actually have — without that risk.
Why patch mammoth instead of using it as shipped? Colour, line spacing and numbering format are all editable in this editor's ribbon. Shipping without the patches would mean the editor shows a control for something it silently eats on the next save. The patches are six string replacements applied at build time and each one asserts its anchor, so a mammoth upgrade that moves the code fails the build instead of quietly shipping a lossy reader.
Why is the original file never overwritten? Because of everything on the "Dropped" list. A save is a download of a new file, so the original is always still there to fall back on.