verify

LOK binding integration verification

End-to-end verification of github.com/docgo/libreofficekit against a real LibreOffice: load documents, render tiles/PDF, extract text selection, clipboard round-trip, and compare every extracted string against an independent oracle that parses the OOXML packages directly (no LibreOffice involved).

Layout

• main.go — Go harness. Loads every file in samples/ through the binding, renders, extracts, and writes out/<name>.json|.txt|.pdf|.png.
• oracle.py — independent expected-content extractor (zipfile + XML): docx body/footnotes/endnotes, xlsx cells (formula evaluation, date/percent number formats), pptx slides/notes, plus comment/reply counts.
• compare.py — token-level coverage/similarity of binding output vs oracle; pptx falls back to PDF text extraction (LO 7.3 Impress has no LOK text selection); xlsx numbers are normalized for display rounding.
• gen_samples.py — regenerates the python-docx/openpyxl/python-pptx samples (deterministic content).
• goexlib-repro/ — regression check for the cgo.CString contract: v0.0.9 returned Go heap memory (crash on Free), fixed upstream in v0.1.0 to allocate on the C heap; this module verifies the fixed contract.
• samples/ — 24 files: 10 docx (tables, headers/footers, lists, CJK, mixed, comments, change-tracking), 9 xlsx (types/number formats, formulas with oracle-side evaluation, merged cells, multi-sheet, 200-row sheet), 5 pptx (decks, CJK, 11-slide, notes).

Run

# 1. generate samples (optional; samples/ is committed)
python3 gen_samples.py          # needs python-docx openpyxl python-pptx

# 2. build & run the harness (uses this repo's binding)
go build -o out/verify ./verify
GODEBUG=asyncpreemptoff=1 ./out/verify
#   env: LOK_LO_PROGRAM=/path/to/libreoffice/program (default /usr/lib/libreoffice/program)

# 3. compare against oracle
python3 verify/compare.py

GODEBUG=asyncpreemptoff=1 is required: the LOK C callbacks are invoked on foreign threads and Go's async preemption otherwise aborts with a SIGURG fatal error.

Known LO 7.3 limitations (not binding bugs)

• Comment callbacks are never emitted by LO 7.3 — comment counts show 0 (comment-warn in the comparison; not counted as failure).
• Impress (pptx) has no text selection in LO 7.3 — pptx is verified via the rendered PDF instead.
• Calc: wait ~1.2 s after .uno:SelectAll before GetTextSelection, otherwise the pre-SelectAll anchor cell text is returned for large sheets.
• Calc SelectPart on LO 7.3 does not move the marked area to the new sheet (text selection keeps returning sheet 1 content), so multi-sheet spreadsheets are verified through the PDF export instead (LO exports all sheets).
• A second view created on a second document crashes LO 7.3, so the harness keeps a single view per document.

Binding bugs found and fixed along the way

• GetClipboard/SetClipboard passed cgo.CString(...) results to cgo.Free(...) — with goexlib/cgo v0.0.9 that aborted with free(): invalid pointer because CString returned Go heap memory. Upstream fixed the contract in v0.1.0 (C-heap allocation); the binding now uses cgo.CString directly (regression check in goexlib-repro/).
• Known separate issue (pre-existing, not from this fix): in LO 7.3 headless, GetClipboard never returns data (getClipboardTry=false in all runs) even right after a successful SetClipboard round-trip; clipboard read-back is not verifiable on this LO version.