bitspire/docs/adr/003-nfc-access-control-layer.md

# ADR-003: NFC Access-Control Layer (badge-to-enter) + Developer Bypass

**Status:** Accepted — amended 2026-09-20 (see [Amendment](#amendment-2026-09-20-what-shipped) below; the original text follows it unchanged)
**Date:** 2026-07-29
**Context:** batm3 gaining a physical access layer — an NFC card must be presented to unlock the machine before anyone can transact. Reader hardware is not yet on hand; this ADR defines the direction and a non-breaking skeleton that is fully testable without it.

## Amendment (2026-09-20): what shipped

The gate landed in aiolabs/bitspire#86 as **Bolt Card tap-to-enter**, not the npub-QR → UID → serial-reader path planned below. The decisions still stand (opt-in `locked` state, operator-owned authorization, hashed identities, fail-closed, audited); the mechanism differs:

- **Reader.** The batm3 / upboard reader is a USB CCID contactless reader (Feitian KP382, OMNIKEY 5022) driven by `pcscd` + `nfc-pcsc` in the **main process** (`electron/nfc-service.ts`, from #83), not the serial `/dev/ttyNFC` device or Web NFC. Taps reach the renderer over the existing `nfc:card-tapped` IPC and the store routes them by state (`locked` → enter; the cash screens → pay / receive). The renderer-side `AccessReader` abstraction, the camera npub-QR reader and the mock reader were never wired to anything and have been removed; `services/access/` now holds only `authorize`, the Bolt Card parser and the credential types.
- **Credential.** Identity is the card's boltcards `external_id`, parsed **locally** from the tapped `lnurlw` (`AccessScan` kind `boltcard`), not the NFC UID. Only `hashId(external_id, salt)` is compared or logged. The `npub` variant (with its PIN second factor) stays in `authorize()` and its tests for a future non-card credential; `challenge` remains the v2 seam; the planned `uid` variant is gone.
- **Verified entry via `/session`** (superseding #86's soft entry). A tap yields a single-use SUN `p`/`c`, and verifying it at entry would have spent the voucher Complete needed, so #86 made no server call at entry. The fork now exposes `/session/<id>?p=&c=` (`docs/boltcard-session.md`): it spends the SUN **once**, proves a genuine non-replayed card, returns the card wallet's balance + fiat, and hands back the hit-keyed LUD-03 / LUD-06 second steps so Complete still needs no second tap. The session is single-shot — after the first Complete attempt, accepted or declined, the store drops it and the customer re-taps. The balance is shown on the idle menu and cash screens **hidden by default** behind an eye toggle (`CardChip.vue`), priced the way the LNbits wallet page prices it (wallet currency, else instance default, else the ATM's fiat at its own rate).
- **Open enrollment is still not a security boundary.** `/session` proves the card is genuine *to the server the card names*: the session URL is derived from the tapped `lnurlw`'s host, so with `openEnrollment` on a forged NDEF tag pointing at a server that answers `authenticated: true` still unlocks the terminal. Money is unaffected (cash-out never dispenses without `PAYMENT_RECEIVED`; cash-in pays where the holder pointed). Closing the gate means pinning the accepted card-server host(s) and/or a provisioned allow-list — tracked in aiolabs/bitspire#91.
- **Session semantics.** One tap = one session; every transaction terminal state returns to `locked`. Inactivity (60 s on the idle menu) and the absolute cap (10 min) are measured at the DOM layer (`useSessionSecurity`) because an XState `after` cannot observe touches. Both send `END_SESSION`, which the machine accepts **from `idle` only**, so a timer can never abandon stacked bills or an in-flight dispense. An explicit End Session button re-locks immediately.
- **Config.** `ACCESS_CONTROL_ENABLED` / `ACCESS_OPEN_ENROLLMENT` / `ACCESS_DEV_UNLOCK` / `ACCESS_SALT` from env, overridden by `/var/lib/bitspire/access.json` (pushed with `deploy/nixos/provision-access.sh`, no rebuild). `devUnlock` defaults **off**.
- **Audit** is still a console stub; the state.db write promised in decision 7 is tracked in aiolabs/bitspire#90.
- **Phased plan superseded.** PR3 (Web NFC) and PR4 (serial HAL) will not happen — the pcscd reader covers real hardware and there is no laptop dev path beyond the debug paste-box on `LockedView`. PR5's challenge-response idea survives as the `challenge` seam.

---

## Decision

1. **Add a top-level `locked` state to the ATM state machine, and make it the initial state.** It sits *below* the existing initialization gates (`unpaired` / `awaiting-fees` / `maintenance` / `signer-unreachable`, which live in `App.vue`). A healthy, paired machine boots into `locked` and only reveals `idle` (Buy/Sell) after an access grant.

2. **Access control is opt-in via runtime config (`accessControl.enabled`, default `false`).** When disabled, the machine behaves exactly as today (boots straight to `idle`). This makes the whole feature non-breaking for the current test unit and for production ATMs, and lets a half-built access layer never brick a working box. This is the single most important constraint on the design.

3. **The reader lives behind an `AccessReader` abstraction that mirrors the existing `PairingSource` seam.** First implementation is a `MockAccessReader` (dev button / hotkey), so the locked→idle→transaction path is exercisable today with zero hardware. Web NFC and serial-NFC implementations follow.

4. **Credential model is a discriminated union with an explicit upgrade path.** v1 = card UID matched against a hashed allow-list. v2 = challenge-response (card-held key signs a machine nonce), verified against an operator-authorized set. Ship v1; design the types so v2 is additive.

5. **Three-tier developer bypass**, following existing conventions: a build flag (`VITE_SKIP_ACCESS_GATE`), the config disable (`accessControl.enabled=false`), and a runtime operator/dev unlock gesture that dispatches a synthetic grant.

6. **Authorization is owned by the machine operator, not the SaaS operator** — consistent with [ADR-002](./002-remote-access-and-fleet-management.md). The card allow-list is authorized by the operator key (the [#42](https://git.atitlan.io/aiolabs/bitspire/issues/42) allow-list mechanism), local first, operator-synced later. When access control is *enabled* and the reader is absent/broken, the machine **fails closed** (with the operator/dev unlock as the escape hatch); when *disabled*, reader state is irrelevant.

7. **Every grant/deny is audited** to `state.db` (hashed credential + timestamp + role + outcome), with optional later publication as a Nostr event. No PII, consistent with the KYC-free principle.

## Context

### What this is (and what it is not)

This is the **end-user physical access plane**: a person must badge in to use the machine. It is distinct from the three planes in [ADR-002](./002-remote-access-and-fleet-management.md) — it is *not* the operator's SSH/NetBird recovery plane, and *not* the SaaS payment plane. It shares one idea with ADR-002: **the machine operator owns who is authorized**, expressed through the operator key / #42 allow-list.

### Why the codebase is well-shaped for this

Three seams already exist; we extend them rather than invent:

- **The idle→transaction transition is unguarded.** `packages/state-machine/src/machine.ts` starts at `initial: 'idle'` (~L431) and `idle` moves to `cashIn`/`cashOut` via plain `SELECT_CASH_IN` / `SELECT_CASH_OUT` transitions with no guards (~L457-464). Inserting a `locked` predecessor state is a localized change.
- **`PairingSource` is a reader abstraction designed to grow.** Its doc (`apps/machine/src/services/pairing/types.ts`) explicitly anticipates *"an NFC reader or a HAL barcode scanner… a HAL-scanner source can be added the same way without touching the wizard."* `AccessReader` mirrors it: `qr-source.ts` / `nfc-source.ts` → `mock-reader.ts` / `web-nfc-reader.ts` / `serial-reader.ts`.
- **Dev-flag and config conventions are established.** `import.meta.env.VITE_* === 'true'` (e.g. `VITE_MAINTENANCE_MODE`, `VITE_FORCE_MOCK`), plus Electron `get-config` fields that the renderer reads (`electron/main.ts` L280–312: `maintenanceMode`, `branding`). A new `accessControl` config field and a `VITE_SKIP_ACCESS_GATE` flag follow the same shape.

### Hardware reality check (important)

The batm3 already exposes an NFC device, but it is **serial**: `deploy/nixos/hardware/batm3.nix` L154 maps udev serial `A9ZF8ELY` → `/dev/ttyNFC`. The *existing* `pairing/nfc-source.ts` uses **Web NFC** (`NDEFReader`), which drives a phone/laptop NFC radio, **not** a serial reader. So the real batm3 access reader needs a **main-process serial driver** (HAL-style, per [ADR-001](./001-hal-architecture.md)) exposing card events to the renderer over IPC — the Web NFC path is only useful for laptop/phone dev. This ADR keeps that driver as a clearly-scoped later PR so the skeleton doesn't pretend the scaffold "just works" on the panel.

## Architecture

### Boot / render layering

```
Electron get-config ─┐
                     ▼
App.vue init gates (unchanged):
  unpaired? → PairingWizard
  initError (maintenance / awaiting-fees / signer-unreachable)? → maintenance screen
  else ▼
State machine (paired + healthy):
  ┌───────────────────────────────────────────────┐
  │  locked  ──ACCESS_GRANTED──▶  idle             │  ← NEW initial state
  │    ▲                          │  SELECT_CASH_* │
  │    │ re-lock (session end /    ▼               │
  │    │ inactivity / complete)   cashIn / cashOut │
  │    └──────────────────────────┘               │
  └───────────────────────────────────────────────┘
        (when accessControl.enabled === false,
         `locked` immediately `always`-bypasses to `idle`)
```

The access gate is strictly below App.vue's init gates: a machine that is unpaired or in maintenance never reaches `locked`.

### 1. State machine (`packages/state-machine`)

- New top-level state `locked`, `initial: 'locked'`.
- New events on the machine's event union: `ACCESS_GRANTED` (carries an authorized `CardCredential` + resolved role), `ACCESS_DENIED` (carries a reason), `DEV_UNLOCK`.
- `locked` transitions:
  - `always: [{ guard: 'accessBypass', target: 'idle' }]` — instant pass-through when disabled/bypassed (no UI flicker; the view is gated on the same predicate).
  - `on: { ACCESS_GRANTED: { target: 'idle', actions: ['startSession', 'recordAccessGrant'] }, ACCESS_DENIED: { actions: 'recordAccessDeny' }, DEV_UNLOCK: { guard: 'devUnlockAllowed', target: 'idle', actions: 'startDevSession' } }`.
- Re-lock: the existing `complete` auto-return (currently 60s → `idle`) and the inactivity timeouts (`INACTIVITY_TIMEOUT`/`TIMEOUT_MS`, ~L422-429) target `locked` instead of `idle`. Because `locked` `always`-bypasses when disabled, this is one code path for both modes.
- Purity: the state-machine package must not read Vite env. `accessControl.enabled` and the bypass boolean are passed in as **actor input → context** (`context.accessControlEnabled`, `context.accessBypass`); guard `accessBypass` reads context only. New context fields: `accessControlEnabled`, `accessBypass`, `session` (`{ role, grantedAt, credentialIdHash } | null`).
- Guards: `accessBypass`, `devUnlockAllowed`. Actions: `startSession`, `startDevSession`, `recordAccessGrant`, `recordAccessDeny`, and `resetContext` extended to clear `session`.

### 2. Config plumbing

- `apps/machine/src/types/electron.d.ts` — extend `RuntimeConfig` (L5) with:
  ```ts
  accessControl: {
    enabled: boolean            // default false
    devUnlock: boolean          // allow the runtime operator/dev unlock gesture
    // v2: allowListSource, challengeRequired, …
  }
  ```
- `apps/machine/electron/main.ts` — the `get-config` handler (L280) returns `accessControl`, sourced from env for now (`ACCESS_CONTROL_ENABLED === 'true'`, `VITE_SKIP_ACCESS_GATE` → forces `enabled:false`), later from a provisioned file under `/var/lib/bitspire/` alongside branding.
- `apps/machine/.env.example` — document `VITE_SKIP_ACCESS_GATE=true` (browser/dev straight to idle) and `ACCESS_CONTROL_ENABLED`.

### 3. Reader abstraction (`apps/machine/src/services/access/`)

Mirrors `services/pairing/`:

```
services/access/
  types.ts          # AccessReader, CardCredential (union), AccessRole, StopCapture
  mock-reader.ts    # PR1 — fires a card event on demand (dev button / hotkey)
  web-nfc-reader.ts # PR3 — NDEFReader, dev on laptop/phone
  serial-reader.ts  # PR4 — /dev/ttyNFC via main-process HAL + IPC
  authorize.ts      # allow-list check + role resolution (hashed UID v1)
  index.ts          # availableAccessReaders(): AccessReader[]
  __tests__/
```

```ts
export type AccessRole = 'user' | 'operator'

export type CardCredential =
  | { kind: 'uid'; uidHash: string }                                  // v1
  | { kind: 'challenge'; pubkey: string; nonce: string; sig: string } // v2 (seam)

export interface AccessReader {
  readonly kind: 'mock' | 'nfc-web' | 'nfc-serial'
  readonly label: string
  isAvailable(): Promise<boolean>
  start(opts: {
    onCard: (cred: CardCredential) => void
    onError?: (e: unknown) => void
  }): Promise<StopCapture>
}
```

### 4. Renderer wiring

- `apps/machine/src/views/LockedView.vue` (new) — the badge-in screen. Shows brand/logo + "Present your card", a live reader status, and (when `accessControl.devUnlock`) a discreet operator/dev unlock affordance (hidden long-press corner, or a button on the existing debug bar).
- `apps/machine/src/App.vue` — add a `locked` render branch mirroring the `PairingWizard` branch (L179) and the `initError` branch (L183): `<LockedView v-else-if="atmStore.isLocked" />`, then the existing `<router-view>` only when unlocked. Keeps rendering state-driven and matches the current shape.
- `apps/machine/src/stores/atm.ts` —
  - `createActor(machine, { input: { accessControlEnabled, accessBypass } })` at the existing `createActor(machine)` site (L452), seeded from `RuntimeConfig`.
  - `isLocked` computed off the snapshot (peer of `isIdle`, ~L422).
  - `grantAccess(cred, role)` / `denyAccess(reason)` / `devUnlock()` that `send({ type: 'ACCESS_GRANTED' | 'ACCESS_DENIED' | 'DEV_UNLOCK', … })` (peers of `selectCashIn` at L1382, using the existing `send` at L1378).
  - On init, when `accessControl.enabled`, subscribe to `availableAccessReaders()[0]`; on `onCard`, run `authorize()` → `grantAccess`/`denyAccess`. When disabled, do nothing (machine `always`-bypasses).

### 5. Audit

- Add `recordAccessEvent({ credentialIdHash, role, outcome, at })` alongside the existing state.db handlers (`state:record-transaction` etc. in `electron/main.ts`, exposed via `preload.ts`). v1 writes locally; a later PR can mirror to a replaceable Nostr event.

### Session semantics (decided)

**One badge = one transaction-scoped session.** A grant unlocks `idle`, the user runs a single transaction (Buy or Sell), and the machine re-locks on `complete`, on inactivity, or on an explicit "Done". The `session` context field is deliberately shaped as a general access session (`{ role, grantedAt, credentialIdHash }`), not a transaction handle, because this terminal may later handle **non-transaction functions** — so "unlock the terminal" and "authorize a transaction" stay separate concepts.

**Step-up authorization (future seam, not in PR1).** The badge tap grants *terminal access*; a specific sensitive action can independently *request re-authorization* — e.g. "tap your phone" or "enter a PIN" — without conflating the two. This is why the credential model is a union and the machine carries a `session` rather than a boolean "unlocked": a later `REQUIRE_STEPUP` event can gate an individual action against a fresh credential/PIN while the terminal session stays open. PR1 ships only the entry gate; step-up is a documented extension.

## Alternatives considered

- **Gate between `idle` and the transaction** (idle visible, tap requires a card). Rejected: the requirement is "gain access to the exchange" — the whole machine should be locked, not just the transact button. A `locked` predecessor matches the mental model and gives a clean re-lock boundary.
- **Web NFC only** (reuse `nfc-source.ts` as-is). Rejected: the batm3 reader is serial (`ttyNFC`); Web NFC can't drive it. Web NFC stays a dev-only convenience.
- **Fail-open by default** (no card → allow). Rejected for an access-control feature; but note the *disabled* default sidesteps this — access is simply off until an operator turns it on, at which point it fails **closed**.
- **OS/kiosk-level lock** (lock the desktop, not the app). Rejected: too coarse, no per-transaction audit, no role model, and it fights the existing state-driven UI.
- **UID allow-list as the permanent model.** Rejected as an endpoint (UIDs clone trivially) but accepted as v1 behind a union type, so challenge-response is additive.

## Security considerations

- **UID cloning** — card UIDs are not secret and are cloneable; v1 is "better than nothing" and is explicitly labeled upgradeable. The v2 challenge-response path (card signs a machine nonce) is the real security boundary; design the credential union and the `authorize()` seam for it now.
- **Data at rest** — store only a salted hash of the credential id; never raw UIDs or any PII (KYC-free). Never log a card secret or nsec (repo security priority #1).
- **Fail-closed when enabled** — reader absent/broken + `enabled` ⇒ locked, escape hatch = operator/dev unlock. Reader problems on a *disabled* machine are inert.
- **Operator ownership** — authorization derives from the operator key / #42 allow-list, not the SaaS operator (ADR-002 boundary). Local allow-list first; operator-published (NIP-51-style) sync later.
- **Dev bypass blast radius** — `VITE_SKIP_ACCESS_GATE` is build-time and never set in a production image; `devUnlock` is gated by `accessControl.devUnlock` (off in a locked-down deployment) and every dev unlock is audited with role `operator`/`dev`.

## Resolved decisions (2026-07-29)

1. **Session model** — ✅ one badge = one **transaction-scoped session**, with the `session` context modeled generally (terminal access, not a transaction handle) to allow non-transaction functions and per-action **step-up auth** (tap phone / PIN) later. See *Session semantics* above.
2. **v1 credential** — ✅ **UID allow-list first** (hashed), behind a `CardCredential` union so challenge-response is additive (PR5).
3. **Allow-list home** — ✅ **local first** (`state.db` / provisioned `access.json`, peer of `branding/`); operator-Nostr sync is a later PR.

Still open (cosmetic, decide during PR2):

4. **Dev unlock affordance** — hidden long-press corner vs a labeled button on the existing debug bar.

---

## Implementation plan (phased PRs)

Each PR is independently mergeable. **PR1 changes nothing observable while `accessControl.enabled=false` (the default).**

### PR1 — Non-breaking skeleton (state + config + mock reader + dev bypass + audit stub)
**Goal:** the locked→idle→transaction path is exercisable on the batm3 today, and the flag-off machine is byte-for-byte behavior-identical.
- `packages/state-machine`: add `locked` state (`initial`), `ACCESS_GRANTED`/`ACCESS_DENIED`/`DEV_UNLOCK` events, `accessBypass`/`devUnlockAllowed` guards, `startSession`/`recordAccess*` actions, context fields + actor `input`. Re-point `complete`/inactivity re-lock targets to `locked`.
- `apps/machine/src/types/electron.d.ts`: `RuntimeConfig.accessControl`.
- `apps/machine/electron/main.ts`: `get-config` returns `accessControl` (env-sourced); `.env.example` documents `VITE_SKIP_ACCESS_GATE` + `ACCESS_CONTROL_ENABLED`.
- `apps/machine/src/services/access/`: `types.ts`, `mock-reader.ts`, `authorize.ts` (UID allow-list, hashed), `index.ts`.
- `apps/machine/src/stores/atm.ts`: actor `input`, `isLocked`, `grantAccess`/`denyAccess`/`devUnlock`, reader subscription (only when enabled).
- `apps/machine/src/views/LockedView.vue` + `App.vue` `locked` render branch.
- Audit stub: `recordAccessEvent` handler + preload exposure (local write only).
- **Tests:** state-machine `locked → idle` on grant; `always`-bypass when disabled; re-lock from `complete`; `authorize()` allow/deny; `devUnlock` gated by config. 
- **Flag state on merge:** `enabled=false`. CI green = no behavior change.

### PR2 — Locked UX polish
- Reader status/animation, brand-aware LockedView, denied-flash + reason, inactivity copy, operator/dev unlock affordance per open-question #4. Pure renderer.

### PR3 — Web NFC reader (dev)
- `web-nfc-reader.ts` (`NDEFReader`), registered in `availableAccessReaders()` behind availability check. Lets a laptop/phone drive the gate for demos/dev. No hardware dependency.

### PR4 — Serial NFC HAL driver (real batm3 hardware)
- Main-process serial driver for `/dev/ttyNFC` (HAL-style per ADR-001), IPC channel `access:watch-card` + `preload.ts` exposure; `serial-reader.ts` renderer client. Requires the physical reader to validate. Document the reader's protocol/baud in `docs/device-configuration.md`.

### PR5 — Challenge-response credential + operator allow-list sync
- Extend `CardCredential` with the `challenge` variant; `authorize.ts` verifies a signature over a machine nonce against the operator-authorized set; allow-list synced from an operator-published event (#42 mechanism). This is the real security upgrade; v1 UID path stays as a fallback/dev mode.

### Cross-cutting
- **Docs:** update `docs/machine-installation.md` (enabling access control, enrolling cards) and `deploy/nixos/README.md` (the `accessControl` config + `/dev/ttyNFC`) as PR4/PR5 land.
- **Provisioning:** a later change can add an `access.json` under `/var/lib/bitspire/` (peer of `branding/`) with the allow-list + `enabled`, plus a `provision-access.sh` mirroring `provision-branding.sh`.