* fix(invoices): roll back the header row when a recurring-schedule item replace fails
PATCH /api/invoices/recurring/[id] and the update_recurring_schedule commit
executor wrote the schedule header first, then replaced the items. An item
insert failure restored the items snapshot but left the header update
committed, so a combined edit half-applied: a new day_of_month or
default_dimensions stayed while the line edit was undone.
Both write paths now go through one shared helper,
lib/invoices/apply-recurring-schedule-update.ts, which snapshots the header
before writing it (only for a combined edit, the only case with something to
undo) and compensates it on any items failure. The rollback update is filtered
on the updated_at stamp our own write produced, so a concurrent writer (the
hourly cron, a second edit) wins instead of being clobbered from a stale
snapshot: audit finding C2 in lib/invoices/voucher-matching.ts.
A compensation that itself fails is no longer swallowed. The helper reports
itemsRestored / headerRestored, logs the unrecoverable rows and the intended
restore payload, and both call sites then return the new
INVOICE_RECURRING_UPDATE_PARTIAL registry entry, which tells the user in
Swedish that the schedule may be half-saved and to check fields and items
before retrying. A clean rollback keeps the PG-mapped error so a CHECK
violation still surfaces its specific message.
Also in the rewritten block:
- the items DELETE error is checked, so a failed delete no longer proceeds to
an insert that would duplicate every line;
- the 404 existence check moved above every write, so a PATCH with items for a
missing or cross-tenant id writes nothing;
- the items snapshot uses select('*') with id/created_at stripped on restore
(same idiom as replaceInvoiceItems), so a column added later is carried
through instead of silently dropped;
- NewRecurringScheduleDialog unwraps the nested { error: { message } } envelope
the route returns, which otherwise reached the toast as "[object Object]".
The cron's no-empty-items invariant holds on every failure path: the items are
either untouched, restored, or the failure is reported explicitly.
Fixes#1275
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(invoices): never write when the compensating snapshot is unavailable
Follow-up on the recurring-schedule rollback: the helper still performed two
writes it already knew it could not compensate.
- The header snapshot read now checks its error and a missing row, and the
header UPDATE is skipped entirely when either holds, so no header change is
committed that we already know can never be rolled back.
- An unreadable item snapshot now aborts BEFORE the delete (rolling the header
back) instead of deleting first and reporting itemsRestored: false, so the
cron invariant "a schedule always has items" holds on every failure path.
- That header read now runs whenever items are replaced and is scoped by
company_id, so it doubles as the ownership proof the schedule_id-only item
delete/insert lacks (the commit executor runs with RLS off). Stated in the
JSDoc as well.
- The item snapshot is paginated via fetchAllRows: a schedule with more than
1000 lines could otherwise restore partially while reporting a clean
rollback.
- The executor now returns errorCode INVOICE_RECURRING_UPDATE_PARTIAL,
surfaced as CommitResult.code and persisted as result_data.error_code, so a
staged-op caller can detect the partial state without substring-matching the
Swedish sentence.
- Route: details keys are camelCase throughout, and an item failure is logged
once, with the repair context kept on the partial path only.
Tests: the unreadable-snapshot branches are exercised (including the
previously unused itemsSnapshotError harness hook), and the test that pinned
"header written with no possibility of rollback" now asserts that nothing is
written at all.
Fixes#1275
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Jakob Wennberg
f3bf50d862