[chaliy]

Requested by Михайло · project thread

What changed

Before: fetching a Word, PowerPoint or Excel file returned "Binary content is not supported". ODS files came back as garbled zip bytes, because their media type was not on the binary list.

After: these documents come back as Markdown, converted locally with no network call and no account:

• DOCX: headings (Heading 1-6, Title), nested lists, and tables.
• PPTX: one ## Slide N: title section per slide, in presentation order.
• XLSX, XLSM, XLS, ODS: one ## sheet Markdown table per sheet, with the first row as header. Dates are shown as ISO, and output is capped at 1,000 rows by 50 columns with a sheet_truncated warning.

The processor matches on media type, and on the file extension when a file is served as application/octet-stream or application/zip. The document title comes from docProps/core.xml.

Markdown requests now also send Accept: text/markdown, text/html;q=0.9, … with Markdown first. Sites that serve Markdown on request (Cloudflare's Markdown for Agents) then return it instead of HTML.

Why

Agents regularly hit Office files on the web, and today they get nothing back. This is the local alternative to Cloudflare's hosted toMarkdown converter: documents never leave the process, and it works the same in every embedding.

Before / After

fetchkit fetch https://go.microsoft.com/fwlink/?LinkID=521962 (an XLSX served as application/octet-stream):

extraction_method: "office_spreadsheet"
---
## Sheet1

| Segment | Country | Product | Discount Band | Units Sold | … | Date | Month Number | Month Name | Year |
| --- | --- | --- | --- | --- | … | --- | --- | --- | --- |
| Government | Canada | Carretera | None | 1618.5 | … | 2014-01-01 | 1 | January | 2014 |

calibre's demo.docx produces # Demonstration of DOCX support in calibre, followed by its paragraphs and ## sections. The python-pptx test.pptx produces ## Slide 1: Presentation Title Text followed by Subtitle Text. Before this change, all three returned Binary content is not supported.

Risk

• Medium: parsing untrusted archives. Zip containers are rejected if they have more than 10,000 entries or more than 200 MB of declared uncompressed content, before anything is inflated. Parts are read through a 64 MB cap, and parsing shares the PDF processor's blocking pool and two-document cap. This is recorded as TM-DOS-008 in the threat model.
• Changing the Accept preference may make some servers return Markdown where they used to return HTML. That is the intended effect.
• New dependencies: calamine (MIT), zip 8 (MIT) and quick-xml 0.41 (MIT). These are the versions calamine already uses, so no crate is duplicated.

Checklist

• Unit tests are passed (cargo test --workspace; 9 new processor tests, plus integration tests for DOCX and the Markdown-first Accept header)
• Smoke tests are passed (the real DOCX, XLSX and PPTX files above, via the CLI)
• Documentation is updated (README, CHANGELOG, tool description)
• Knowledge is up to date and not in conflict (fetchers, tool contract, threat model, log; check_okf.py passes)

[chaliy]

The Lint failure isn't caused by this PR's code. CI's stable toolchain is now Rust 1.99, and its new clippy::double_must_use lint flags code that async-trait 0.1.89 generates. That hits every async trait already in the crate (ContentProcessor, Fetcher, FileSaver, HttpTransport), so main fails the same way.

Fix in 1ab5437: bump async-trait to 0.1.92, which no longer generates the duplicate must_use. With that, cargo +stable clippy --workspace --all-targets -- -D warnings (1.99) is clean locally, and cargo test --workspace passes.

---

Generated by Claude Code