Files
zeekay 73b7290e7f square: one webhook route and one verifier; secrets come from the host's KMS
Square keeps an event's created_at across every retry, and the receiver
refused anything older than five minutes, so every retry of an event
first sent while cloud was down (Recreate, ten rollouts a day) was refused
until Square gave up. The verifier now lives in payment/providers/square
and needs only the key and the URL; thirdparty/square is the API client
and no longer registers itself or verifies anything, so every registry
holds one Square type.

Replay is stopped by the event id instead: each verified event is
recorded once for the whole deployment, keyed on the verifying processor
and the event id with an insert-if-absent swap, so a repeat is a
duplicate under any path or org. The path names the one processor that
verifies, and only one that holds its webhook secret may: an unknown name,
or a verifier that needs no secret, is refused rather than tried.
The books an event lands in are the verifying account's, else the
deployment's brand — never a header or session, which on a replay are the
replayer's. An effect that did not land (a store or the ledger did not
answer) leaves the delivery open with a 503, so the provider redelivers;
it used to be acknowledged and never retried. A settlement reads the
payment index before the org's books, because an in-session charge's
receipt is in the payer's org.

POST /v1/commerce/webhooks/:provider is declared once, in setupRoutes, so
standalone and embedded commerce serve the same address; the
/v1/billing/webhooks spelling is gone.

thirdparty/kms read /v1/kms/orgs/<org>/secrets/..., which cloud no longer
serves, and was off in production anyway, so per-org payment credentials
never loaded. It is deleted. secrets.Org reads orgs/<org>/<name> through
the host's reader, the caller re-pointed at that org with zip.ActingAs so
the KMS tenancy check admits it, and payment.Hydrate loads an org's
credentials where the org receives the money: storefront checkout,
storefront subscriptions and wire, and the brand's bank for the billing
wire rail. Hanzo's own billing no longer reads the paying org's
credentials; it charges the deployment's account, as it did. The
standalone binary reads through hanzoai/kms/sdk/go.
2026-10-19 09:12:56 -07:00

98 lines
3.7 KiB
Go

// Copyright (c) 2014-present Hanzo AI, Inc.
// Licensed under MIT OR Apache-2.0. See LICENSE-MIT and LICENSE-APACHE.
// Mount implements the HIP-0106 unified Hanzo Cloud binary contract.
//
// Commerce is a LIGHT ROUTER, NOT in PCI-DSS scope. It handles tokens
// and intent IDs only; it MUST NEVER touch a PAN. Anything that would
// touch a PAN belongs behind the out-of-process Payments + Vault services,
// which the host calls — commerce holds no client to either.
//
// NATIVE zip/ZAP — no gin, no net/http adaptation. Commerce registers its
// handlers directly on the host binary's zip.App: one router, one specificity
// space, one middleware chain. There is no second engine and no request
// crossing an adapter boundary.
//
// A host passes a data dir and a logger. It does not pass itself: commerce is
// mounted BY a host, so naming a host type here would let exactly one host
// mount it and would point the dependency backwards — the host already imports
// commerce in order to mount it.
package commerce
import (
"context"
"fmt"
"log/slog"
"net/http"
"github.com/zap-proto/zip"
)
// Mount registers commerce on app per the HIP-0106 contract.
//
// dataDir is the per-deployment data root; per-tenant SQLite lands under it
// per HIP-0302. Empty falls back to the cloud default tree. log is the host's
// logger — nil means slog.Default(), the same rule Embed uses, so a missing
// log sink never costs a deployment its checkout router.
func Mount(app *zip.App, dataDir string, log Log, secrets SecretReader) error {
if app == nil {
return fmt.Errorf("commerce.Mount: nil zip.App")
}
if log == nil {
log = slog.Default()
}
if dataDir == "" {
dataDir = "/var/lib/cloud/commerce"
}
// App: app is the whole mount. Bootstrap registers commerce's route group
// (/v1/commerce, public and tenant-admin alike) ON THE HOST'S ROUTER and skips
// every standalone-only surface — /healthz, the legacy admin SPA, the
// checkout SPA catch-all, and Listen, since the host owns the listener.
// HTTPAddr stays empty for the same reason.
if _, err := Embed(context.Background(), EmbedConfig{
DataDir: dataDir,
App: app,
Logger: log,
Secrets: secrets,
// RequireIdentity stays env-driven; the gateway in front of the
// cloud binary is the trust boundary per HIP-0026.
RequireIdentity: false,
}); err != nil {
return fmt.Errorf("commerce.Mount: embed: %w", err)
}
// Native zip health endpoint, on a commerce-scoped path so a probe can
// target this subsystem specifically rather than the whole binary. Under
// /v1 like everything else commerce answers: a probe path outside it is a
// second address rule for one subsystem, and hosts already spell it
// /v1/commerce/health.
app.Raw(http.MethodGet, "/v1/commerce/health", func(c *zip.Ctx) error {
return c.JSON(200, map[string]string{
"status": "ok",
"service": "commerce",
})
})
// The /v1 model bundle (api.Route — ~140 routes across /v1/billing,
// /v1/product, /v1/order …) is deliberately NOT registered here.
//
// In the cloud binary those prefixes are cloud's own: /v1/billing is served
// by clients/billing + clients/account. Registering them again on the same
// app is not additive — byte-identical patterns MERGE silently with the
// first registration winning, and two params at equal specificity with
// different names PANIC at registration. Either way the money path would be
// decided by import order.
//
// So the mounted surface is exactly the commerce-scoped one, which is what
// it has always been in practice. cmd/commerced still calls api.Route for
// the standalone binary, where those prefixes have no other owner.
log.Info("commerce mounted",
"prefix", "/v1/commerce",
"data_dir", dataDir,
)
return nil
}