diff --git a/prisma/migrations/0028_transaction_source_identity/migration.sql b/prisma/migrations/0028_transaction_source_identity/migration.sql
new file mode 100644
index 0000000..2644f7b
--- /dev/null
+++ b/prisma/migrations/0028_transaction_source_identity/migration.sql
@@ -0,0 +1,54 @@
+-- A feed that re-sends yesterday's rows needs an identity the importer can recognise.
+--
+-- The CSV import that exists today is built for a one-off: `batchInsertCSVTransactions`
+-- assigns `row_index` as MAX(row_index) + 1 over the owner's manual rows, so the same
+-- file imported twice produces two sets of rows with different indexes. That defeats
+-- `uq_transaction_identity` (statement_id, transaction_date, description, amount,
+-- row_index) by construction — the constraint cannot fire, because the fifth column is
+-- guaranteed fresh on every run. Nothing else stops it either.
+--
+-- For a hand-driven bank CSV that is tolerable; the operator sees the file once. For a
+-- recurring aggregator export it is not, because the windows overlap *by design*: a
+-- 12-month export pulled weekly re-states ~51 weeks of rows it has already sent. The
+-- failure would be silent and cumulative, and it is not one we can clean up afterwards —
+-- every child of `transactions` is ON DELETE CASCADE, so a duplicate must be superseded
+-- rather than deleted (migration 0023), and ~$42k of re-imported statement rows already
+-- exist as evidence of how this goes.
+--
+-- So the provider's own key travels with the row. Frollo issues a stable per-transaction
+-- `id`, which is exactly what the wallet-capture lane lacked and could never synthesise
+-- from a notification. The same two columns serve any future feed that has one.
+
+ALTER TABLE transactions
+  ADD COLUMN IF NOT EXISTS source TEXT;
+
+COMMENT ON COLUMN transactions.source IS
+  'Feed this row was ingested from (e.g. ''frollo''). NULL = entered by hand, imported from a statement, or predates the column.';
+
+ALTER TABLE transactions
+  ADD COLUMN IF NOT EXISTS source_ref TEXT;
+
+COMMENT ON COLUMN transactions.source_ref IS
+  'The provider''s own identifier for this transaction, verbatim. Idempotency key for re-imports; never generated locally.';
+
+-- Partial, so the millions of rows with no source do not have to be unique on (NULL,
+-- NULL). Enforced in the database rather than in the importer: an ON CONFLICT DO NOTHING
+-- that silently depends on application-side dedupe is one refactor away from not.
+CREATE UNIQUE INDEX IF NOT EXISTS uq_transaction_source_ref
+    ON transactions (source, source_ref)
+ WHERE source IS NOT NULL AND source_ref IS NOT NULL;
+
+-- A source row is only half-identified without knowing which account it came from --
+-- two accounts at the same institution can legitimately carry the same provider id
+-- namespace. Kept as free text rather than a foreign key: finance-app has no account
+-- entity, and inventing one to hold a label from an external system would be the tail
+-- wagging the dog.
+ALTER TABLE transactions
+  ADD COLUMN IF NOT EXISTS source_account TEXT;
+
+COMMENT ON COLUMN transactions.source_account IS
+  'Account label as the source system names it, for provenance and for scoping an import to particular accounts. Not an entity reference.';
+
+CREATE INDEX IF NOT EXISTS idx_transactions_source
+    ON transactions (source, transaction_date)
+ WHERE source IS NOT NULL;
diff --git a/prisma/schema.prisma b/prisma/schema.prisma
index 77f1c7d..4653e44 100644
--- a/prisma/schema.prisma
+++ b/prisma/schema.prisma
@@ -187,6 +187,9 @@ model transactions {
   superseded_by_id        Int?
   principal_amount        Decimal?      @db.Decimal(12, 2)
   interest_amount         Decimal?      @db.Decimal(12, 2)
+  source                  String?       // feed this row came from, e.g. "frollo"; NULL = statement/manual (migration 0028)
+  source_ref              String?       // the provider's own transaction id — idempotency key for re-imports
+  source_account          String?       // account label as the source names it
   statement               statements?   @relation(fields: [statement_id], references: [id], onDelete: Cascade)
   reconciled_with         transactions?  @relation("reconciled", fields: [reconciled_with_id], references: [id], onDelete: SetNull)
   reconciled_by           transactions[] @relation("reconciled")