ai-bookkeeper reads bank exports from many banks, but most of its tests use files we wrote ourselves. So we keep importing other projects' sample files. This time: the fixtures in Paisa (an Indian personal-finance tracker) and Maybe.
Four CSVs, one problem. Maybe's Mint-style and plain transaction files read right. Paisa's ICICI card statement read right too, with a preamble skipped and a warning that every amount was positive. Paisa's Paytm wallet statement read all 74 rows with correct dates and amounts, but every description was empty. Its header has Activity, Source/Destination and Comment, and none matched a description name. Blank descriptions mean no categorization, so the books would have come out uncategorized without any error message.
The fix reads Activity ("Paid for order", "Added to wallet") as the description and appends Comment when present. It applies only when the header also has Wallet Txn ID, so other files are unaffected. A regression test covers it, using made-up rows in that layout.
Still open: the ICICI file's dates (01/03/2023) are ambiguous between 1 March and 3 January; the importer assumes month first and does not say it guessed. Those fixtures are three rows, so there is nothing to settle it. Real statements from Indian banks are the next thing to try.