A parser you can read beats one you trust: hand-rolling .xlsx to keep a CSP promise

A developer who runs a client-side converter for bank statements added .xlsx support and declined to install SheetJS, on the grounds that a large third-party parser weakens the specific privacy claim the site makes ^[4][5]. What makes the decision worth reading is the accounting behind it: he enumerated the subset of the format he actually needed, and it was small ^[6][7].

The site turns bank CSV exports into the file format QuickBooks Desktop accepts, and the guarantee is enforced rather than asserted ^[1]. Every page ships a Content Security Policy with connect-src 'none', so the browser will not permit the page to make any network request at all ^[2]. Open the Network tab during a conversion and it stays empty ^[3]. His objection to the dependency was not bundle size ^[4]. It was that "your file never leaves the browser, and here is the policy that enforces it" degrades into "and also trust this dependency", which he calls a materially weaker claim for a tool handling people's bank statements ^[5].

Then the inventory. An .xlsx is a ZIP of XML: xl/workbook.xml lists the sheets, xl/worksheets/sheet1.xml holds the cells, xl/sharedStrings.xml is a deduplicated string pool that cells reference by index, and xl/styles.xml carries the number formats ^[6]. Reading that needs three capabilities, two of which the browser already provides ^[7]. Unzipping means walking the ZIP central directory, about forty lines ^[8]. Decompression is native, via DecompressionStream('deflate-raw') ^[9]. For the XML he hand-rolled a tag scanner instead of using DOMParser, because his tests run in Node, where DOMParser does not exist, and he wanted one code path rather than two ^[10]. Net result: two of the three pieces written by hand, one of them replacing something the platform supplies ^[1].

Dates cost the most care, because Excel does not store dates. A cell holding 5 January 2024 contains the number 45296, and whether that renders as a date depends on the cell's number format, which lives in a different file inside the archive ^[11]. So the parser reads styles.xml, works out which style indexes correspond to date formats, and checks every numeric cell against that list ^[12]. Neither failure mode raises an error ^[13]. Missing a date format leaves a column of five-digit numbers, which at least looks wrong; treating a currency format as a date turns an amount of 45296 into a day in January 2024, which does not ^[14]. The epoch is 1899-12-30 rather than 1900-01-01, because Excel treats 1900 as a leap year and reserves serial 60 for a 29 February that never existed ^[15]. Older Mac files use a 1904 epoch flagged by date1904="1" on workbookPr, and missing that attribute puts every date off by four years and a day ^[16].

The instructive part is the bug. Twenty-six tests against workbooks the suite generates itself, covering date handling, sparse cells, shared strings and sheet selection, all passing ^[17]. He then dropped a file on the real page and got back one line: Failed to fetch ^[18]. The cause was collecting the decompression output by wrapping the stream in a Response and calling arrayBuffer(), which Chrome counts as a fetch, which connect-src 'none' blocks, producing what reads like a network error in code with no network ^[19]. Draining the reader by hand avoids it ^[20].

Worth watching in your own build: whether the environment your tests run in enforces the constraint your product sells. Node does not enforce CSP, so the green suite could never have caught this ^[21]. A smaller companion detail is corrupt input, where the writer side of the stream rejects alongside the reader, so without a .catch() on both writer.write() and writer.close() an Uncaught (in promise) lands next to the real error ^[22].