Draft: synchronized X launch thread¶
Do not post. This is final local copy for owner publication. Publish only after the article and evidence links pass logged-out preflight.
Launch window¶
Publish Post 1 within the same ten-minute window as the Hacker News submission. Use the package-drift bar chart as the only attachment to Post 1. Post the rest as a native thread, then add the Hacker News link as a reply after the submission is live.
Thread¶
Post 1
An AI agent edits one cell. What else changes inside the workbook?
I re-saved five valid XLSX files through LibreOffice bundled with ChatGPT Desktop's Codex runtime. It changed 8 to 151 unrelated OOXML parts. WolfXL changed 0.
We published the receipts.
Post 2
The test was intentionally boring: open, request no content change, save to a new file, then compare every OOXML package part.
Fixtures covered external links, validation and protection, custom XML, a pivot chart with slicers, and a PowerPivot model.
Post 3
Changed does not mean corrupted. Office export filters can legitimately normalize XML and relationships.
The finding is mutation blast radius: a no-op save touched 8, 19, 14, 31, and 151 unrelated parts. That is state an agent should inspect before replacing the source.
Post 4
Same-source macOS benchmark, 3 samples per row: WolfXL was 2.2x to 3.9x faster on the paired tasks, and its measured install was 37.5 MB vs 423.0 MB.
The red row stays in: isolated footprint RSS favored LibreOffice, 34.5 MiB vs 53.1 MiB.
Post 5
WolfXL's agent path is a guarded transaction:
- declare operations and ranges
- patch an isolated copy
- verify the requested change and unrelated drift
- publish atomically
Unsupported work fails visibly or uses an explicitly named fallback.
Post 6
Full method, exact-source receipts, counter-results, and reproduction links:
https://wolfgangschoenberger.com/blog/when-an-ai-agent-edits-a-spreadsheet
If you test your own workbook, use a public or synthetic file. Please do not upload confidential workbooks.
Reply after Hacker News is live
The Hacker News discussion is here: [HN URL]
I would especially value adversarial fixtures and feedback on which workbook features should be hard refusal boundaries for an agent.
Asset¶
Use the final package-drift chart exported for this launch. Do not use the older speed or footprint cards because they contain source-mismatched values.
Recommended alt text:
Package drift after a no-op XLSX open and save. Across five fixtures, WolfXL changed zero unrelated OOXML parts. Bundled LibreOffice changed 8, 19, 14, 31, and 151 unrelated parts.
Pre-post checks¶
- The portfolio article is live and loads while logged out.
- The three
github.com/SynthGL/wolfxl/blob/main/docs/trust/receipt links load. - Pull request 626, or its exact replacement commit, is on the public default branch.
- Every selected number matches the 2026-09-03 source-matched receipts.
- The package-drift image uses the same five values and makes no corruption claim.
- No private path, hostname, session identifier, workbook, screenshot, token, or mutable temporary result appears.
- The owner posts from the intended account.