open-standards/openpay
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Transaction Science fc5081ae3f
All checks were successful
standards workspaces build + test / which workspaces this change touches (pull_request) Successful in 10s
standards workspaces build + test / macOS energy backends (pull_request) Has been skipped
repository checks / lockfiles, tests in CI, workflow hygiene, spec and section references, public text (pull_request) Successful in 1m21s
standards workspaces build + test / openpay (openpay, openpay, linux-arm64) (pull_request) Successful in 6m0s
op-connect: an unrun PEP screen is not a clearance
`ConnectScreener` takes its PEP and adverse-media screeners as
`Option<Screener>`, documented so that "operators who only have a
sanctions feed can still use this crate (PEP / AM are then no-ops
returning empty hit-vectors)". Both screens are gated on
`if let Some(..)`, so with `pep: None` neither `pep_hits` nor
`any_review` is ever touched, and the fall-through returned
`Decision::Clear` — whose own doc comment read "No hits anywhere; safe
to proceed."

The module doc states that at each BeneficialOwners submission the
platform *must* "screen every person against a PEP watchlist ... per
FATF Recommendation 12 and EU AMLD5 Art. 20". So the crate emitted
"safe to proceed" for a screen it is required to run and did not run.
`ScreeningResult` recorded no field saying which screens had run, so
the persisted row was indistinguishable from one where a real PEP feed
was queried and returned nothing: a sitting head of state onboarded
under a sanctions-only configuration produced the same record as a
genuine three-screen clearance, and the gap was invisible on review.
The crate's own test `clear_when_lists_empty` built exactly that
configuration and asserted `Clear`, pinning the behaviour.

Unknown is now a distinct value rather than the best possible one:

- `ScreenCoverage { pep, adverse_media }` records which optional
  screens ran, and travels with the hits. Sanctions is absent from it
  because it cannot be absent from the run — `new` takes it by value,
  so a bool could never be false. `#[serde(default)]` makes a stored
  row that does not state its coverage default to "neither ran": a
  record that does not say cannot assert it.
- `Decision::Incomplete` reports a mandated screen that did not run.
  Precedence is `Block` > `Review` > `Incomplete` > `Clear`; a hit
  needing adjudication outranks a missing screen because the gap is
  reported independently in `coverage`.
- `Decision::may_proceed()` is true only for `Clear`, so a caller
  cannot read `Incomplete` as proceedable by forgetting a match arm.

The mandated/optional line is drawn from the crate's own doc: items
1-3 are "must", item 4 is "Optionally", so a missing adverse-media
screen is recorded but still clears, and a missing PEP screen does not.

`clear_when_lists_empty` is replaced by a discriminating pair with
identical empty hit-vectors and opposite decisions — unconfigured PEP
yields `Incomplete`, a configured-but-empty PEP index yields `Clear` —
which is the distinction the old single test could not express.

Co-Authored-By: Claude Opus 5 <[email protected]>
2026-10-17 04:13:10 -04:00
..
.github/workflows Initial import — three open standards under one shelf 2026-05-23 07:58:09 -04:00
compliance ci: no-vanity-metrics scanner + pre-commit hook + CHARTER policy 2026-06-02 10:39:43 -04:00
crates op-connect: an unrun PEP screen is not a clearance 2026-10-17 04:13:10 -04:00
docs ci: no-vanity-metrics scanner + pre-commit hook + CHARTER policy 2026-06-02 10:39:43 -04:00
examples Repo-wide toolchain refresh + jouleclaw policy/sandbox updates 2026-06-07 17:04:45 -04:00
vocab vocab v1 SHACL shapes across sister standards (wave-23 #4/#5) 2026-06-02 18:55:56 -04:00
.gitignore Initial import — three open standards under one shelf 2026-05-23 07:58:09 -04:00
Cargo.lock chore(deps): update dependencies for Rust 1.98, pinning the crypto generations 2026-08-26 12:19:50 -04:00
Cargo.toml msrv: raise the rust-version floor to 1.97 across the workspaces 2026-07-23 00:17:08 -04:00
CONTRIBUTING.md Initial import — three open standards under one shelf 2026-05-23 07:58:09 -04:00
LICENSE Initial import — three open standards under one shelf 2026-05-23 07:58:09 -04:00
README.md docs: strip vanity metrics from arl + openpay + wai READMEs 2026-06-02 07:53:21 -04:00

OpenPay

An open, reference payment-acceptance stack in Rust.

A single safety-critical core compiled to every platform — iOS, Android, browser, Linux — that runs the full payment lifecycle: tap-to-pay, vault, on-device fraud, multi-rail routing, double-entry ledger, reconciliation, settlement, refunds, disputes. Apache-2.0. No per-transaction fee to OpenPay.

OpenPay is for merchants, banks, fintechs, and public-sector deployers who want to accept payments without paying a tax on every transaction to a hosted SaaS, and without ceding control of the routing, fraud, and ledger logic that decides whether their business runs.


Why this exists

The modern payment stack is built around assumptions that no longer hold:

  • Card networks aren't the only rail anymore. FedNow, PIX, UPI, SEPA Instant clear in seconds, at near-zero cost, with no chargeback window. A merchant that can route across rails — card to A2A, primary PSP to backup PSP — captures revenue that a card-only stack drops on the floor.
  • Fraud has to be decided before the rail. A2A settlements are irrevocable in <20s. There is no "refund the fraud later" with FedNow or PIX. The fraud decision has to run on the merchant device, pre-routing, in milliseconds — not in a cloud-hosted post-hoc batch job.
  • The ledger is the system of record. Most stacks treat the ledger as a derived view of PSP webhooks. OpenPay treats it as a first-class append-only graph with bi-temporal history, so you can ask "what was this account's balance on May 4th, as we knew it on May 5th?" as a single query, not a six-week archaeology project.
  • Software shouldn't tax money movement. Stripe Terminal, Adyen, Square — all charge per-transaction fees on top of interchange. OpenPay is Apache-2.0 software. You pay the rails. You don't pay OpenPay.

What's in the box

OpenPay is a multi-crate Rust workspace plus reference examples, organized into five tiers. Each tier depends only on the tiers below it.

┌──────────────────────────────────────────────────────────────────────┐
│  TIER 5 — FFI + CLI                                                  │
│  op-ffi-swift   op-ffi-jni   op-wasm   op-cli                        │
│  iOS / macOS    Android      browser   operator command line         │
├──────────────────────────────────────────────────────────────────────┤
│  TIER 4 — Deployment                                                 │
│  op-server  op-refund  op-dispute  op-settlement  op-subscriptions   │
│  HTTP API   refunds    disputes    batch payouts  recurring billing  │
│  op-driver-sdk  op-fx                                                │
│  driver authors FX conversion + quote providers                      │
├──────────────────────────────────────────────────────────────────────┤
│  TIER 3 — Orchestration & ledger                                     │
│  op-orchestrator   op-ledger   op-reconciliation   op-graph          │
│  routing + retry   double-      ledger ↔ bank      Minigraf-backed   │
│  + idempotency     entry        statement diff     stores (Datalog,  │
│  + 3DS resume      append-only                     bi-temporal)      │
├──────────────────────────────────────────────────────────────────────┤
│  TIER 2 — Rails & risk                                               │
│  op-rails-card   op-rails-a2a   op-rails-crypto   op-fraud   op-webhook │
│  Hyperswitch     FedNow / PIX   USDC on Base /    on-device  Stripe-shaped │
│  driver          / SEPA Instant Ethereum / etc.   scoring    outbound │
├──────────────────────────────────────────────────────────────────────┤
│  TIER 1 — Foundation                                                 │
│  op-core   op-iso20022   op-emv   op-vault                           │
│  typed     PACS/PAIN/    EMV TLV  token-only                         │
│  money,    CAMT          parser   storage                            │
│  typestate                                                           │
│  payments                                                            │
└──────────────────────────────────────────────────────────────────────┘
Crate Responsibility
op-core Domain types: Money, Currency, Payment<S> typestate machine, PaymentMethod, RailKind. No f64 money. PAN never appears in default builds.
op-iso20022 Thin facade over the open-payments-iso20022 crates. PACS/PAIN/CAMT message helpers and Camt053Source for statement parsing.
op-emv BER-TLV codec for parsing EMV tap blobs from iOS Secure Enclave / Android StrongBox.
op-vault Token-only storage. Vault trait + reference InMemoryVault (AES-256-GCM-SIV).
op-rails-card CardAcquirer trait + Hyperswitch driver + Stripe-compatible patterns.
op-rails-a2a A2aAcquirer trait + drivers for FedNow, PIX (Bacen SPI), SEPA Instant (RT1/TIPS). Direct ISO 20022, no PSP intermediary.
op-rails-crypto Stablecoin rail. CryptoGateway trait + EvmJsonRpcGateway (feature evm) for USDC on Base / Ethereum / Polygon / Arbitrum, with a hand-rolled ERC-20 calldata encoder. LocalKeyEvmSigner (k256 + RLP, EIP-155) ships behind the same feature for hot-wallet operators; production deployments wire Fireblocks / KMS / multisig behind the EvmSigner trait.
op-fraud On-device scoring. Scorer trait with three backends: heuristic, ONNX (via ort), and pure-Rust Burn.
op-webhook Outbound event delivery with Stripe-compatible signing + auto-disable on chronic failure. Real HTTPS delivery via ReqwestTransport (feature reqwest-transport); MockTransport for tests.
op-orchestrator Routing across rails, idempotency, soft-failure fallback, 3DS / SCA resume. The "main" library API.
op-ledger Double-entry append-only ledger. Transaction, Entry, Account, Balance.
op-reconciliation Ledger vs. bank-statement diffing. Deterministic UUID v5 matching.
op-graph Minigraf-backed implementations of every domain store (LedgerStore, WebhookStore, ReconciliationStore, RefundStore, DisputeStore, SettlementStore, SubscriptionStore, IdempotencyStore). Bi-temporal history, Datalog queries, audit report builder — all persisted to one .graph file.
op-refund Refund state machine: Requested → Submitted → Approved → Settled.
op-dispute Chargeback / dispute workflow with evidence attachment.
op-settlement Batch settlement, holdback computation, and NACHA payout-file generation.
op-subscriptions Recurring billing: plans, intervals (calendar-aware), trial periods, dunning policy, proration math.
op-fx Foreign-exchange primitives: Quote, QuoteProvider trait, integer-exact convert with banker's rounding, StaticQuoteProvider + CachedQuoteProvider ref impls.
op-server Axum HTTP server with env-driven main. POST /v1/intents + /v1/intents/resume + /v1/refunds + /v1/disputes + /v1/settlement/batches + /v1/subscriptions + /v1/fx/quote + /v1/audit/report. API-key auth + token-bucket rate-limit middleware. Auto-registers a usdc-base rail when OP_BASE_RPC_URL + OP_USDC_BASE_PRIVATE_KEY are set.
op-cli op command for operator ergonomics. 14 subcommands across health, refund, dispute, batch, subscription, FX, webhooks, audit.
op-driver-sdk For driver authors. DeterministicCardAcquirer / DeterministicA2aGateway / DeterministicCryptoGateway mocks + conformance::run_card / run_a2a / run_crypto harnesses that catch contract violations in custom drivers.
op-ffi-swift Swift bridge via swift-bridge + C ABI. Generates OpenPay.swift + C header.
op-ffi-jni Android bridge via JNI. Opaque handles in Kotlin; PAN never crosses the boundary.
op-wasm Browser / Node.js bridge via wasm-bindgen.

Status

v0.1.0. This is a reference stack: the architecture is complete and the code compiles, runs, and tests cleanly across all platforms, but no part of it has been certified by a card scheme, audited by a PCI QSA, or run in regulated production. Use it as a starting point, not as one.

Workspace phases 0 through 31 are implemented end-to-end — foundation, rails (card / A2A / crypto), orchestrator with 3DS resume, ledger with bi-temporal time-travel, settlement + payout file generation, subscriptions, FX, webhook delivery, HTTP API, driver SDK with conformance harness, single-file Minigraf persistence across every domain store, and an operator bring-up sprint that ships a real EVM signer (LocalKeyEvmSigner via k256), env-driven config, an op CLI, and a demo-merchant example you can run on a laptop against Base mainnet.

cargo test --workspace is 1124 passing, 0 failing; feature-gated paths add 28 EVM + 111 reqwest-transport tests. cargo clippy --workspace --all-targets is zero warnings.

See docs/ for the per-phase progress documents that capture design intent and rationale.


Quick start

Library: take a card payment

use op_core::{Currency, Money, Payment, PaymentMethod, RailKind, VaultRef};
use op_rails_card::{
    acquirer::{AuthRequest, ThreeDsMode},
    hyperswitch::HyperswitchClient,
    CardAcquirer,
};
use uuid::Uuid;

let acquirer = HyperswitchClient::new(HyperswitchClient::SANDBOX, api_key);

let amount = Money::from_minor(10_00, Currency::USD);            // $10.00
let method = PaymentMethod::Vault(VaultRef::new("tok_test_visa"));
let payment = Payment::new(amount, method.clone(), RailKind::Card);

let auth = acquirer.authorize(&AuthRequest {
    amount,
    method,
    auto_capture: false,
    idempotency_key: Uuid::now_v7().simple().to_string(),
    three_ds: Some(ThreeDsMode::Skip),
    metadata: None,
})?;

// The typestate machine enforces lifecycle correctness at compile time.
// You cannot refund a Payment<Created>; you must Capture first.
let authorized = payment.authorize(auth.psp_payment_id.clone());
let captured   = authorized.capture(amount)?;
let refunded   = captured.refund(amount)?;

Full source: examples/headless-server/src/main.rs.

Library: route across rails with the orchestrator

use op_fraud::HeuristicScorer;
use op_orchestrator::{Orchestrator, PolicyRouter, PaymentIntent, IdempotencyKey};
use op_core::{Currency, Money, PaymentMethod, VaultRef};

let orchestrator = Orchestrator::new()
    .with_scorer(Box::new(HeuristicScorer::new()))
    .with_router(Box::new(PolicyRouter::new(
        vec!["hyperswitch".into(), "stripe".into()],   // card driver priority
        vec!["fednow".into()],                          // A2A driver priority
    )));

// ... register CardAdapters and A2aAdapters ...

let intent = PaymentIntent::new(
    IdempotencyKey::new("ORDER-001"),
    Money::from_minor(12_99, Currency::USD),
    PaymentMethod::Vault(VaultRef::new("tok_visa_4242")),
);

let outcome = orchestrator.run(&intent)?;
//  outcome.terminal_status       → Approved / RequiresCustomerAction / Declined
//  outcome.attempts              → full retry / fallback trail
//  outcome.psp_payment_id, .uetr → settlement references

A replay with the same IdempotencyKey returns the cached outcome without touching the rail — no double charges. A soft failure on the primary PSP transparently falls back to the backup; a hard decline does not. Full source with five end-to-end scenarios: examples/kiosk-linux/src/main.rs.

Server

cargo run -p op-server
# POST /v1/intents     { amount_minor, currency, method, ... }
# POST /v1/refunds     { original_payment_id, amount, ... }
# POST /v1/disputes    { original_payment_id, reason, ... }
# POST /v1/settlement/batches
# GET  /v1/audit/report?tx_count=100

op-server ships with in-memory stores so it runs out of the box. Operators replace the stores with their persistent backends (e.g. op-graph for Minigraf, or their own LedgerStore impl) by wiring a different AppState into the exported router(AppState) function.

Mobile and web

The FFI crates produce native bindings:

  • iOS / macOS: cargo build -p op-ffi-swift → generates OpenPay.swift + C header.
  • Android: cargo build -p op-ffi-jni --target aarch64-linux-android → loadable .so with Java_dev_openpay_* symbols.
  • Browser: wasm-pack build crates/op-wasm → JS module.

Card data is held behind opaque handles on every FFI boundary. The PAN never crosses into Swift, Kotlin, or JavaScript.


Project layout

openpay/
├── Cargo.toml          # workspace: 26 member crates + 3 examples
├── crates/             # the OpenPay stack
├── examples/
│   ├── headless-server/   # single-payment lifecycle via op-rails-card
│   ├── kiosk-linux/       # full orchestrator with multi-rail fallback
│   └── demo-merchant/     # generates a fresh wallet, listens for USDC on Base
├── docs/               # per-phase progress documents (NN-name-progress.md)
│   └── deploy/         # Caddyfile, systemd unit, env sample, launch.sh
└── tests/              # workspace-level integration tests

Design principles

These aren't conventions; they're enforced by the type system or by the build.

  • No f64 money. Money is integer minor units paired with Currency. There is no floating-point money type in op-core, and you can't add Money(USD) to Money(EUR) — the operation does not compile.
  • Typestate payment lifecycle. Payment<Created>, Payment<Authorized>, Payment<Captured>, Payment<Refunded> are distinct types. Refunding before capture is a compile error, not a runtime check.
  • No PCI scope in default builds. The PaymentMethod enum has no raw-PAN variant unless you turn on the pci-scope feature. Orchestrators, fraud scorers, and FFI bridges only ever see VaultRef (opaque token) or Emv(SecureBlob).
  • Append-only ledger. Entries are never updated or deleted. Corrections are themselves entries. Bi-temporal history (via op-graph over Minigraf) is free, not bolted on.
  • One core, every platform. op-core, op-fraud, op-vault are no_std-friendly and compile to iOS, Android, WASM, and Linux from the same source. The HTTP server lives in its own crate so the device builds never pull in tokio or axum.
  • Drivers are external and verifiable. Operators write their own CardAcquirer / A2aAcquirer implementations and run op_driver_sdk::conformance::run_card(&driver)? to catch contract violations before deployment.

Toolchain

  • Rust 1.95 (edition 2024). The MSRV is pinned in Cargo.toml and enforced by CI.
  • The release profile uses lto = "fat", codegen-units = 1, and symbol stripping. The device-side builds are designed to be small.

Documentation

The docs/ directory holds a NN-name-progress.md file for each implementation phase. They are not user manuals — they capture the design choices, the alternatives considered, and the rationale that the code can't express on its own. If you want to know why the ledger is graph-backed, why the orchestrator owns idempotency, or why we swapped IndraDB for Minigraf mid-build, start there.


Contributing

OpenPay is early. The contributions that help most right now:

  • Driver implementations for additional PSPs and A2A rails. The op-driver-sdk conformance harness will tell you whether your driver behaves.
  • Backend implementations of LedgerStore, WebhookStore, ReconciliationStore against your persistence layer of choice.
  • Platform glue for the FFI bridges — Swift package, Android AAR, NPM package.
  • Test vectors for ISO 20022 messages, EMV tags, and reconciliation scenarios.

Discussions and PRs welcome.


License

Apache-2.0. See LICENSE.