#Office Viewer

A read-only document viewer built with MoonBit and Rabbita. Office Viewer provides standalone DOCX, XLSX, and PDF previews and a reusable browser surface with structured selections for host applications and agents.

DOCX/XLSX interaction state, UI construction, source mapping, and selection logic live in MoonBit, using Rabbita's typed DOM interfaces and small browser bindings. See the browser runtime for the boundaries and verification workflow.

Live demo

#Supported formats

Format	Rendering	Selection
DOCX	office.mbt PageModel and paginated SVG	Native text selection with validated paragraph/run references where available
XLSX	office.mbt workbook reader and HTML renderer	Cells and rectangular ranges with worksheet names and canonical selectors
PDF	PDF.js canvas, text, and annotation layers	Native text selection with page/item ranges, document identity, geometry, and available tagged structure

The viewer supports zoom and search. PDF previews also provide page navigation, fit to width, rotation, and password prompts. Files are read locally in the browser; the standalone demo uses a mock file path in its selection metadata. Excel previews preserve workbook styles, merged cells, images, charts, and formula results. Supported uncached formulas are evaluated by office.mbt; unsupported formulas remain visibly marked as pending.

#Run locally

Use the official MoonBit toolchain 0.10.15+943464bfa and Node.js 24. Dependencies resolve from the Mooncakes registry; sibling checkouts are not required.

moon update
npm ci --ignore-scripts
moon run tools/prepare-browser-assets.mbtx
moon check --target js
moon test --target js
npx --no-install playwright install chromium
npm run test:browser

Start the development server:

moon install moonbit-community/warren@0.4.5
warren \
  -C "$PWD" dev --direct --browser-entry cmd/browser --server-entry "" --port 4311

Open http://127.0.0.1:4311/. With Warren installed as above, build the same release artifact as CI with:

npm run build:pages

The generated dist/ is ignored by Git; an optional first argument selects a different output directory inside the repository (required by Warren). The script uses Warren to compile/minify the browser entry and copy public/, including the existing bundled fonts. It fingerprints JS/CSS filenames and writes build-info.json with the source revision and full asset SHA-256 hashes. It does not regenerate fonts. Run moon run tools/prepare-browser-assets.mbtx after updating npm dependencies; the generated PDF.js worker, CMaps, fonts, codecs, and icon assets are copied from public/vendor/ into the release artifact.

#Continuous deployment

The Pages workflow checks, tests, and builds on pull requests and every push to main. PRs never deploy. After a successful main build (or a manual run on main), it uploads the generated temporary artifact and deploys it with GitHub Pages. Build output is not committed to the repository. The workflow pins the published Mooncakes dependencies, official compiler/core archives and their checksums, Warren (with its bundled minifier), and npm browser dependencies. Update these pins together after compatibility checks. The deployed build-info.json identifies the exact source commit and asset contents. PDF.js and icon assets are generated before the source build; they are served locally from the deployed directory, including under subpaths.

#Structured selections and annotations

The current selection layer is deliberately read-only: Word uses browser text selection inside positioned page SVG, and Excel uses rectangle selection over cells. Word's source metadata is built in the viewer by reading the same bytes through office.mbt's annotated DOCX reader and joining its paragraph/run projection with the PageModel text. The join is fail-closed, so an uncertain paragraph remains physical-only instead of receiving a guessed selector. Both formats produce the versioned office.selection.v1 JSON envelope shown in docs/selection-model.md; the inspector displays that envelope and can copy it through the browser clipboard for the later host application's Agent conversation.

Hosts can choose @viewer.materialize_selection_full, @viewer.materialize_selection_preview, or @viewer.materialize_selection_deferred to control how much selected content enters an Agent context. The APIs preserve the existing office.mbt file and selector fields and make UTF-16 omission counts explicit; see docs/embedding.md.

PDF selections use office.pdf.text-range.v1 references rather than office.mbt selectors. Agent materializations keep the selected text and a single reference with the file path, SHA-256, parser version, and text-item endpoints. Hosts can resolve that reference against the same file bytes using resolve_pdf_selection; the raw selection and resolver result retain detailed geometry and tagged metadata. The demo's Selection panel exposes Full, Preview, and Deferred output and PDF reference resolution. See PDF preview for the PDF contract and browser tests.

Existing DOCX comments/w15 replies and Excel traditional Notes are available in a document gutter and the optional Comments overview, and through @viewer.read_annotations or RenderedDocument.annotations. The office.annotations.v1 contract preserves missing metadata and explicit support limits. Word navigation shows only proven paragraph context; Notes beyond preview bounds remain in the list and JSON. The gutter opens alongside the document when space permits; on narrow layouts, open its heading or a source marker to read comments and replies. Quote a whole comment or highlight any body text to capture an office.selection.v1 annotation-body selection, including source identity and UTF-16 offsets. Preview/deferred modes retain the reference without leaking the unselected body. See docs/annotations-model.md for the API, unsupported modern-comment formats, and tested dependency versions.

The reusable surface lives in viewer/. Its renderer and Rabbita surface are host-neutral: a standalone host can provide browser file bytes, while OpenSeek can provide bytes and the real office.mbt file path from its native backend. The intended OpenSeek integration mounts this same JS/Rabbita surface inside the desktop frontend; it does not open a second webpage or iframe.

Word surfaces include a visible limitations banner and inert office.render.diagnostics.v1 JSON when the layout reader reports unsupported content. Warning coverage depends on the installed pagelayout release; an empty warning list does not certify complete OOXML rendering. Floating drawings, nested tables, and paragraph-level section geometry still need engine work.

Excel freezes authored rows/columns in a worksheet scroll area, including after zoom and worksheet switches. Split panes are not treated as frozen panes. Embedding hosts can set render_document(..., max_rows=..., max_cols=...); the defaults are 1000 rows and 256 columns with a 262144-cell per-sheet ceiling. The table, formula metadata, and Notes preview flags use the same bounds. Notes outside the preview remain available in the annotation list. See rendering support for version boundaries and known limitations.

Word and PDF hosts can add the reusable page-thumbnail list through render_page_thumbnails and page_thumbnails_command. Rendering, lazy caching, current-page synchronization and navigation belong to the viewer; the host chooses its panel layout and whether it starts open. See the embedding guide.

#Font policy

The document surface uses the same bundled font programs as the pagelayout metrics. They are exported into public/fonts from pagelayout/fontoutlines and loaded with @font-face, so a browser fallback cannot change the fixed SVG character positions. The bundled Carlito, Liberation, and Noto Sans SC files are distributed under the license notices in public/fonts/LICENSES. scripts/prepare-fonts.mjs adds browser naming and checksum metadata while preserving glyphs and metrics. Modified internal names and the changes are recorded in public/fonts/LICENSES/MODIFICATIONS.md.

System fonts remain an explicit local enhancement. The Aa button switches the browser's glyph source to installed Calibri/Arial/Times New Roman/Courier New or common CJK equivalents when available, while keeping the page model's bundled coordinates and pagination unchanged. Missing local faces fall back to the bundled files. The default mode never silently changes with the host machine; the browser runtime records both bundled loading and the selected font mode.

From the Office checkout, regenerate bundled fonts only after changing the upstream font data:

moon run --target native ./pagelayout/cmd/pagelayout -- \
  export-fonts ../office-viewer/public/fonts
node ../office-viewer/scripts/prepare-fonts.mjs ../office-viewer/public/fonts

PDF fonts, CMaps, and images are handled by PDF.js. Embedded fonts come from the opened document; standard fonts and CMaps are included with the PDF.js assets and their upstream license notices. Image-only PDFs display normally; OCR and inferred paragraph/table extraction are not included.

Site source Report issue Packages Build queue Skills Statistics