Один чек часто смешивает несколько категорий трат (гипермаркет: часть — продукты,
часть — бытовая химия). Раньше такую транзакцию приходилось огрублять до одной
категории или дробить вручную на несколько записей. Теперь у транзакции есть
второй режим: 2+ доли (TransactionSplit) с суммами по разным категориям, которые
обязаны в точности складываться в сумму транзакции.
- Новая модель TransactionSplit + миграция.
- split_rule_violation() проверяет инварианты на создании и редактировании.
- category_rollup.py — общий helper, которым теперь пользуются Dashboard, отчёт
по категориям и бюджеты, учитывая доли сплитов наравне с обычными транзакциями.
- Фильтр по категории в списке транзакций находит и сплит-транзакции.
- Сплиты попадают в полный бэкап/восстановление БД.
- 13 новых backend-тестов.
- Фронтенд: переключатель «Разбить на несколько категорий» в форме транзакции
с повторяемыми строками и индикатором остатка суммы; строка транзакции в
списке показывает все категории сплита.
TransactionRead inherited TransactionBase's model_validator, so the rules a
new row must satisfy were also applied on the way out, to rows already in
the database. A row can stop satisfying them through no fault of its own:
transactions.transfer_account_id is ON DELETE SET NULL, so deleting an
account leaves every transfer that pointed at it without a destination.
The result was that deleting an account could 500 the entire transactions
list — permanently, and with the list gone there was no way in the UI to
find and delete the offending row.
The fields move to TransactionFields, which reading uses; TransactionBase
keeps the invariants for the write path. The frontend already tolerates a
missing destination (TransactionsTable renders the transfer suffix only when
transfer_account_id is set, and the edit modal makes you pick a destination
before saving), so such a row is now visible, editable and deletable
instead of fatal.
The transfer rules — destination required, destination different from the
source account, no category, and no destination on a non-transfer — were
enforced by TransactionBase's validator on create, but the PATCH route only
re-checked category-vs-type.
So `PATCH /api/transactions/{id} {"type": "transfer"}` on an ordinary
expense committed a transfer row with transfer_account_id = NULL, and only
then failed serializing the response (TransactionRead inherits the same
validator). The caller got a 500 and would reasonably assume nothing was
saved — but from that point on every `GET /api/transactions` failed the same
validation, so the transactions list, the app's main screen, returned 500
permanently, with no way left in the UI to delete the offending row.
The rules now live in one place, transfer_rule_violation(), used by both the
create schema and the update route, so the two paths can't drift apart
again. The update path validates the row as it would look after the patch
rather than only the fields sent, since switching type alone invalidates
fields that were left untouched.
The frontend already sends a complete payload on edit (explicit nulls for
category_id / transfer_account_id) and does these same checks client-side,
so this only closes the hole for direct API calls — no UI change needed.
Claude-Session: https://claude.ai/code/session_01UmuMDFrwdvQkce885VaaTh
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Fourth of the four ideas picked from the Maybe/Sure feature audit, and the
last one queued up. A 3-step wizard (account -> column mapping -> preview)
at /transactions/import, parsed entirely client-side (new lib/csv.ts —
delimiter auto-detect, 5 date formats, thousands/decimal-separator-aware
amounts) and submitted through a new all-or-nothing POST /transactions/bulk
endpoint, mirroring backup restore's validate-everything-before-touching-
anything approach.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Transactions/recurring templates entered with mixed casing ("траты на
продукты") are now capitalized on create/update via a Pydantic validator,
so the fix applies to every client, not just the web UI. Adds a one-off
backfill script for existing rows.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Lets the user find a past purchase by keyword instead of paging through
months manually — search hits description/merchant/notes and ignores the
month filter, with a "jump to month" action on each result to view it in
its normal monthly context.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
backend/tests/: 33 pytest tests against a real Postgres database
(aurum_test, created/migrated/dropped per run) exercised through the actual
HTTP API — not sqlite, not mocks, and the real `aurum` database is never
opened (app lifespan is skipped, get_session is overridden per test).
Covers transaction category-kind validation, year/month/date-range
filtering and pagination, dashboard summary arithmetic, budget status, cash
flow, and the /transactions/years picker.
e2e/: 6 Playwright specs driving a real browser against a fully separate
throwaway docker-compose stack (run.sh: separate project name, database,
port, and volume, torn down with `down -v` on exit) — regression coverage
for the Recent Transactions month bug, the Card grid-overflow bug, the
frozen-year bug, the Net Worth default range, and the Dashboard->
Transactions deep link, all fixed earlier this session.
Also synced the backend's APP_VERSION constant (stale at 0.15.2, embedded
in exported backups) with the frontend version.
