There was no inbound payment path at all: a button called a fake synchronously
and wrote an order. A real provider does the opposite — it charges, then tells
us, repeatedly, out of order, and sometimes long afterwards.
POST /api/payments/webhook verifies the signature before the body is parsed, so
an unsigned or tampered delivery is refused and recorded without touching an
order. Verified deliveries are stored under the provider's own event id with a
unique constraint, and applied inside the same transaction that marks them
processed: a repeat is a no-op, a crash is retried rather than half-applied.
An approval whose amount disagrees with the reviewed quote does not become an
order. Underpayment would ship artwork nobody paid for, and overpayment means
something a person should look at.
Order creation moved to app/payments.py so the webhook and the local development
checkout share one implementation and cannot drift. That also closes 3.5: the
charge happens inside the transaction that persists the order, rather than
before it.
The adapter contract is create/verify/parse. FakePayment implements it with a
real HMAC scheme so the whole path is exercised now, by tests/payment_test.py:
unsigned, tampered, underpaid, duplicate, re-sent, unknown reference, and
non-approved statuses. Connecting Mercado Pago is one adapter; no service code
changes.
PAYMENT_WEBHOOK_SECRET is optional in production on purpose. Required would
break the next Portainer render, and a guessable default would be worse than
either: with no secret configured the adapter verifies nothing and therefore
accepts nothing, which is the right state until a provider is connected.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
local/ held six unrelated things under a name that stopped being true once it
became the production runtime: the service, the frontend, the tests, the ops
commands, the container definitions and the dependency lock, 65 files with
nothing to tell them apart.
app/ the service: api/ routers, core/ for identity, database, models,
prices and secret loading, and the worker, bootstrap and schema
tests/ the twelve suites, no longer inside the shipped package
ops/ backup, readiness, dependency audit, security summary
infra/ Dockerfiles, gateway templates, ClamAV and storage configuration,
the requirements and their hash lock
web/ the Site, Kanban and portal pages with their scripts
deploy/Dockerfile.api now copies app/ alone, so the tests stop shipping to
production; the local image still carries them, because the suites run inside
the stack's network.
Five kinds of reference had to follow, and each was found by something different
rather than by reading. Imports of the form "from . import db" survived a rewrite
that only matched "from .db import". Tests kept relative imports of modules that
had left the package. A mock.patch target names its module in a string, where no
import rewriting can see it. The browser test resolves a fixture by path. And the
release gate's markers pointed at local/runtime.py and local/worker.py, which is
the decay its new marker test exists to catch — it caught it.
Verified from docker compose down -v: the stack starts, all six integration
suites, both browser suites and the twenty-nine unit tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
