Refund system (Round 3 fixes + follow-up + alignment):
- Serialize cancellation refunds against the manual handler via
per-payment advisory locks taken before the prior-refunds read
(pg_advisory_xact_lock, ascending, same crussell:refund: key space)
- Aggregate pending cancellation refunds into ONE Square refund per
charge (stable charge-level -square-agg key); atomic group UPDATE
keeps crash-retry amounts identical for Square key-dedup
- Persist paymentID-square-amount idempotency keys on cancellation
refunds; scheduler reads the stored key (legacy fallback for old rows)
- Add sweep-pending-square-refunds cron (*/5, concurrency 1) with
refund_attempts cap; sweep retries stale manual pending refunds with
each row's own stored idempotency key
- Reconcile at Square (GET /v2/refunds ListPaymentRefunds) before every
terminal failed transition: tri-state result leaves rows pending on
reconcile error instead of false-failing; PAYMENT_ALREADY_REFUNDED
resolves to completed
- Move over-refund guard inside the lock, counting completed + pending
(excluding failed); ErrRefundDeclined distinguishes definitive vs
ambiguous outcomes
- forgiveFees now executes a real full refund (forceFullRefund override)
with admin_forgiven_fees reason threaded to Square
- Surface failed card refunds in the admin notification centre
(refund_failed enum, RETURNING-id pre-pass inserts, NOT EXISTS dedup)
- Dedup double-cancel refund inserts via ON CONFLICT (idempotency_key)
DO NOTHING without consuming refundRemaining
Frontend:
- Remove all raw-PAN card entry: zero card_number/card_cvc/new_card_token
in request bodies; gate new-card entry behind CardEntryUnavailable
notice + newCardDisabled prop across all 8 flows
- Delete hand-rolled CardInput.svelte; keep CardSelection saved-card UI
and CardEntryUnavailable fallback
- Update cancellation-policy page to in-person cash pickup wording
Tests:
- Rewrite the two amount-blind dedup tests to assert real money movement
(single call, aggregated amount, shared refund ID)
- Add coverage: manual refund vs cancellation serialization (concurrent
goroutines), reconcile error vs no-match branches, stale manual retry,
forgive-fees real refund row + reason, double-cancel dedup, mock refund
key dedup, ListPaymentRefunds filtering
- Fix time-dependent booking flakes with fixtures.NextWorkingDayAt
- 25/25 packages pass; -race clean on payments/square/db/jobs/bookings
P0 — float truncation: applied math.Round to all remaining int64(x*100)
sites (till penceAmount, refund over-refund guard, GetAlreadyRefundedAmount,
payment summary conversions). A £1.14 till sale previously charged 113p.
P0 — raw PAN stopped at the API edge:
- Deleted CardNumber/CardExpMonth/CardExpYear/CardCVC from TillSaleRequest
and CardNumber/Expiry/CVC from CreatePaymentMethodRequest. Both now accept
card_token (Square nonce) and return 400 when absent. PAN+CVV no longer
transit the application server (PCI-DSS SAQ-A scope).
- Deleted CreateCardOnFileRaw from the SquareClient interface and all
implementations (MockClient, ProdClient, devProdClient).
- Added idempotency_key column to refunds table (UNIQUE).
P0 — RefundPayment hardened: advisory lock on payment ID (prevents two
concurrent refunds passing the over-refund guard), pending-refund-record-
then-Square pattern (scheduler reprocesses on failure), same-key dedup.
P1 — till sale pending-retry now re-attempts the Square charge instead of
returning the stale 'pending' status (gift card was already funded in the
committed tx — silent money loss otherwise). Sale row reused, not duplicated.
P1 — idempotency key caching in frontend: BuyGiftCard and
UserPaymentModal/BookingFlow now cache the key per amount+card, regenerated
on change and cleared on success — matches the tip-flow pattern so a
lost-response retry dedups instead of double-charging.
P1 — CreateTerminalPayment cash/giftcard INSERTs now persist idempotency_key.
Key is unique per payment (booking+type+amount would wrongly dedup two
legitimate identical payments, e.g. two £50 cash receipts).
P1 — gift-card codes no longer logged (spendable credential; value+recipient
only).
Tests: till pending-retry re-attempt, refund same-key dedup, mock CreatePayment
idempotency dedup, CreatePaymentMethod nonce happy path + raw-PAN rejection,
till online_square card_token required/valid.
N1 (HIGH) — BuyGiftCard concurrent same-key retry could double-issue gift
cards (2× value for 1 charge). Added pg_advisory_lock on the idempotency key
(mirroring the tip pattern) acquired before the idempotency check, so
concurrent same-key retries serialize and only one executes gift-card
creation.
N2 — Amount-equality guards in both reuse branches (CreateTipPayment and
BuyGiftCard). A same-key retry with a different amount now returns 400
instead of silently mutating the pending record's books/VAT/refund caps.
N3 — test coverage:
- TestBuyGiftCard_RetryPending_ReattemptsCharge: pending record + same-key
retry re-attempts, reuses the record (count=1), completes, and issues the
gift card exactly once.
- TestCreateCheckoutHTTP_DeviceOptionsWireShape: httptest.Server asserts
device_id is under checkout.device_options (not top-level). Extracted
createCheckoutHTTPWithClient for injectable base URL.
- MockClient.CreatePayment now dedups on idempotency key (paymentByKey map),
matching real Square behaviour.
N4 — Corrected the savepoint comments in handlers.go and giftcards.go: the
savepoint only exists in the test harness; in production db.Conn.Begin is a
plain tx and the status UPDATE runs on a separate pooled connection. Commit
is a harmless no-op in prod but required in tests.
Bonus bug fixed: CheckIdempotencyByKey scanned NULL booking_id/gift_card_id
(gift-card purchases) into plain string, failing with 'cannot scan NULL'.
Now uses sql.NullString.
Docs: Technical Manual.md:53 and Feature Catalog.md (2.1, 2.5) corrected —
no longer claim Web Payments SDK is live; new-card entry is documented as
pending P11, saved-card flow works via ccof tokens, dev mock rejects raw PANs.
CRITICAL — same-amount tip retry silently never charged:
- CreateTipPayment idempotency check now only short-circuits when the
existing record is 'completed'. A 'pending' record (previous Square call
failed) is REUSED and the charge re-attempted with the same key (Square
dedups safely), instead of returning the stale pending record as 200 with
a success toast and no charge.
- Same fix in BuyGiftCard: pending records trigger a re-attempt, not a
false-success response. Unique idempotency_key constraint means the
pending record must be reused, not re-inserted.
- Fixes the savepoint/rollback interaction: the nested tx (savepoint) is
now committed in the reuse path so the deferred rollback doesn't undo the
later status UPDATE on the same connection.
- Regression test: TestTipPayment_RetryPending_ReattemptsCharge verifies a
pending record + same-key retry re-attempts and completes, reusing the
record (count stays 1).
MAJOR — terminal checkout wire contract:
- device_id now sent as checkout.device_options.device_id (Square's required
shape), not a top-level field which Square rejects with 400.
- 'checkout pending' detection now uses typed sentinel ErrCheckoutPending
with errors.Is in both handlers, matching mock and real HTTP client.
MAJOR — exp_month/exp_year omitted from card creation payload when unset
(now *int with omitempty) — Square would 400 on 0/0; expiry comes from the
tokenized source.
Docs:
- README payments/infrastructure sections corrected (Web Payments SDK claim
replaced with accurate P11-backlog note; dev mock parity described)
- Future Work P11 updated to reflect raw-PAN rejection is now enforced in
both mock and prod (new-card flows are a documented dead end)
- Added plans/p11-square-web-payments-sdk.md: full implementation plan +
handoff prompt for the agent picking up P11 (Web Payments SDK nonces)
Previously the dev MockClient was more permissive than production:
- MockClient.CreateCardOnFileRaw processed raw PANs and stored mock cards,
while ProdClient and devProdClient both block raw PANs. A dev testing the
raw-card flow saw it succeed, masking a production failure.
- MockClient.CreateCardOnFile accepted raw PANs as source_id via an
isAllDigits branch. Real Square only accepts cnon:xxx/ccof:xxx tokens.
Now the mock behaves identically to production:
- CreateCardOnFileRaw returns the same PCI error as ProdClient
- CreateCardOnFile validates source_id is token-like (cnon:/ccof:) and
rejects raw PANs
- Removed dead isAllDigits helper
Tests updated to assert the parity behavior:
- TestDevClient_CreateCardOnFileRaw_Rejected_ProdParity (table-driven,
replaces 5 brand-specific raw-PAN tests)
- TestDevClient_CreateCardOnFile_RejectsRawPAN (replaces RawNumber)
- TestCreatePaymentMethod_HappyPath / SecondCardNotDefault now expect
500 instead of 200, documenting the prod block
- square_dev.go: return a copy of CheckoutResult to the caller
instead of sharing the map-stored pointer, preventing the goroutine
from racing with the caller on Status field reads
- scheduler_test.go: remove t.Parallel from TestInitHostname_* tests
which modify a shared package-level hostname variable
New test files cover previously untested paths across DAV, validators,
S3, Square, mw, bookings, user, and payments packages.
Includes mock fix: HoldCheckouts flag on MockClient allows tests to
pause auto-complete goroutine for testing PENDING checkout states.
Coverage: 50.4% → 65.0% (+14.6pp)
Move m.mu.RUnlock to defer in GetCheckout and remove duplicate RLock/RUnlock around m.completed lookup. This prevents a deadlock when CheckoutPending error is returned but the RLock was already released.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
popertots
