feat(relay): Phase1 waves A2-E — server entry, shared-store data plane, agent runtime, deploy artifacts

RELAY-PHASE1 Wave A2/B/C/D/E (12-agent workflow, all tsc-clean, 314/314 tests pass): - A2: control-plane server.ts entry + boot/redis.ts revocation-bus wiring + start script. - B1: relay-run shared-store EnforceDeps (relay-auth ports over the SAME Postgres+Redis as P3). - B2: registry-backed MtlsVerifier (verifyAgentCert, fail-closed, INV14). - B3: store-backed RouteResolver (subdomain->hostId). - B4: Redis relay:revocations subscriber -> tunnel teardown (INV12). - B5: main-phase1.ts production entry (public bind, real TLS, async-mTLS prefetch bridge) + staging /auth/mint. - B6 (PARTIAL): relay-web operator login + browser DPoP; proof offered via term.dpop.<b64u> subprotocol. - C: agent dist/cli.js build (esbuild) + runTunnel run-loop + CliDeps. - D1: same-origin static serve of relay-web/public from the browser WSS. - E: systemd units + gen-ca/gen-capability-key/issue-tls-cert scripts + deploy/RUNBOOK.md. Adversarial review: all hard invariants PASS. Follow-ups (B7): close DPoP-subprotocol read on browser-server (blocks browser connect); rate-limit /auth/mint (F1); wire activeSessionCount (F2); scrub error logs (F5). Excludes unrelated public/style.css (concurrent iOS job).
2026-07-06 16:13:34 +02:00
parent 95b9cccf07
commit aa1912b962
40 changed files with 4783 additions and 19 deletions
--- a/relay-run/src/wiring/mtls-verifier.ts
+++ b/relay-run/src/wiring/mtls-verifier.ts
@@ -0,0 +1,93 @@
+/**
+ * B2 · registry-backed MtlsVerifier (INV4/INV14) — replaces the Phase-0 stub in `relay-world.ts:149`
+ * that trusted ANY peer cert. Wraps relay-auth's `verifyAgentCert` against the SHARED host registry
+ * (the same `HostRegistryPort` B1 builds over Postgres) and the pinned agent-CA bundle, returning
+ * `{ hostId, accountId }` ONLY for an enrolled, unrevoked host whose leaf chains to our CA and is in
+ * its validity window. Every other outcome is `null` (fail-closed).
+ *
+ * INV3: `hostId`/`accountId` originate ONLY from the authenticated cert material (SPIFFE-ID) matched
+ * against the registry — never from a client-supplied field. INV14: the pubkey is bound to the
+ * registry AFTER CA-chain validation (a cert can chain to our CA yet still be denied if not enrolled).
+ *
+ * ── ASYNC IMPEDANCE (flagged for the orchestrator; adaptation per task B2) ──────────────────────
+ * The term-relay `MtlsVerifier.verifyPeer` (agent-listener.ts:16) is declared SYNC and its caller
+ * (agent-listener.ts:93) does NOT await it (`if (verified === null)`). A registry-backed verifier
+ * CANNOT be sync — `HostRegistryPort.getById` returns a Promise, so `verifyAgentCert` is async. This
+ * factory therefore returns an `AsyncMtlsVerifier` (Promise-returning `verifyPeer`). Wiring it into the
+ * data-plane requires `MtlsVerifier.verifyPeer` to become async END-TO-END: `agent-listener.ts` must
+ * `await` the result, and the `relay-world.ts:149` stub must be replaced. Those files are OUTSIDE B2's
+ * Owns. Until that migration lands, assigning this into the sync slot type-errors at the wiring site
+ * (`Promise<…>` is not assignable to `{…}|null`) — a useful compile-time tripwire, not a silent bug.
+ */
+import {
+  verifyAgentCert,
+  defaultParseX509,
+  type ParseCert,
+  type HostRegistryPort,
+} from 'relay-auth'
+
+const PEM_LINE_WIDTH = 64
+/** Non-global (safe for `.test()`): asserts a bundle actually holds a PEM CERTIFICATE block. */
+const PEM_CERT_RE = /-----BEGIN CERTIFICATE-----[\s\S]+?-----END CERTIFICATE-----/
+
+/**
+ * Async analog of term-relay's sync `MtlsVerifier` (agent-listener.ts). See ASYNC IMPEDANCE above:
+ * a registry lookup is inherently async, so `verifyPeer` returns a Promise.
+ */
+export interface AsyncMtlsVerifier {
+  verifyPeer(peerCertDer: Uint8Array): Promise<{ hostId: string; accountId: string } | null>
+}
+
+export interface MtlsVerifierDeps {
+  /** Pinned agent-CA bundle (PEM: intermediate(s) + root). NEVER the browser LE chain (INV14). */
+  readonly caChainPem: string
+  /** Shared host registry (same port B1 builds over Postgres). accountId/hostId only from here (INV3). */
+  readonly hosts: HostRegistryPort
+  /** Epoch-SECONDS provider (compared against the leaf's notBefore/notAfter). */
+  readonly now: () => number
+  /** Optional observability seam for a registry/parse fault; the verifier still fails closed. */
+  readonly onError?: (e: unknown) => void
+  /** Test seam: cert parser (defaults to production `defaultParseX509`). */
+  readonly parse?: ParseCert
+}
+
+/**
+ * Wrap raw DER bytes (`getPeerCertificate(true).raw`) into a PEM certificate block with 64-column
+ * base64, the shape node's `X509Certificate` (via `verifyAgentCert`) parses.
+ */
+export function derToPem(der: Uint8Array): string {
+  const b64 = Buffer.from(der).toString('base64')
+  const wrapped = b64.match(new RegExp(`.{1,${PEM_LINE_WIDTH}}`, 'g'))?.join('\n') ?? ''
+  return `-----BEGIN CERTIFICATE-----\n${wrapped}\n-----END CERTIFICATE-----\n`
+}
+
+export function createMtlsVerifier(deps: MtlsVerifierDeps): AsyncMtlsVerifier {
+  const { caChainPem, hosts, now } = deps
+  const onError = deps.onError ?? (() => {})
+  const parse = deps.parse ?? defaultParseX509
+
+  // Fail fast at construction on a misconfigured CA bundle (INV14: a real pinned CA is required; an
+  // empty/garbage bundle would otherwise silently deny every agent with an opaque `chain_invalid`).
+  if (typeof caChainPem !== 'string' || !PEM_CERT_RE.test(caChainPem)) {
+    throw new Error('createMtlsVerifier: caChainPem must contain at least one PEM CERTIFICATE block')
+  }
+
+  return {
+    async verifyPeer(peerCertDer) {
+      if (peerCertDer === undefined || peerCertDer === null || peerCertDer.length === 0) {
+        return null // no client cert presented → fail closed
+      }
+      try {
+        const leafPem = derToPem(peerCertDer)
+        const result = await verifyAgentCert(leafPem, caChainPem, now(), hosts, parse)
+        if (!result.ok || result.hostId === undefined || result.accountId === undefined) {
+          return null // expired / chain-invalid / not-enrolled / revoked / account-mismatch → fail closed
+        }
+        return { hostId: result.hostId, accountId: result.accountId }
+      } catch (e: unknown) {
+        onError(e) // registry/DB fault: surface it, never throw out of verifyPeer → fail closed
+        return null
+      }
+    },
+  }
+}
--- a/relay-run/src/wiring/revocation-subscriber.ts
+++ b/relay-run/src/wiring/revocation-subscriber.ts
@@ -0,0 +1,121 @@
+/**
+ * B4 · relay-run revocation subscriber (INV12 data-plane teardown). Bridges the frozen §4.2
+ * `relay:revocations` Redis pub/sub bus onto the running relay node so a revocation tears the live
+ * tunnel(s) down within the INV12 budget.
+ *
+ * Flow: an injected ioredis subscriber-mode client SUBSCRIBEs the one named channel; each message is
+ * validated as a `KillSignal` with the relay-contracts schema (malformed → dropped + counted, never
+ * a teardown and never a bus crash — a poisoned message can neither kill the bus nor fire a spurious
+ * kill); then for every live tunnel THIS node serves the pure relay-auth predicate
+ * `killsScope(signal, hostAccountId, hostId)` decides coverage and, on a match, the whole host is
+ * torn down (revoked ⇒ grace forced to 0 upstream in T11).
+ *
+ * Reuse, don't re-derive: scope math is `killsScope` (relay-auth) and validation is `KillSignalSchema`
+ * (relay-contracts) — this module only wires those pure primitives to a concrete ioredis client and
+ * the relay node. Security: scope selection is blast-radius-bounded (host ⊂ account ⊂ global); a host
+ * this node does not serve is a no-op (no cross-tenant reach). `signal.reason` is metadata only and is
+ * never written as terminal payload or log content (INV10). The owning `accountId` used for
+ * account-scope fan-out comes only from the authenticated tunnel (INV3).
+ */
+import { KillSignalSchema, RELAY_REVOCATIONS_CHANNEL, type KillSignal } from 'relay-contracts'
+import { killsScope } from 'relay-auth'
+
+/** One live tunnel this node currently serves: the host and the account that owns it. Both are
+ *  authenticated material carried on the tunnel (INV3) — never derived from the untrusted signal. */
+export interface ActiveTunnelRef {
+  readonly hostId: string
+  readonly accountId: string
+}
+
+/**
+ * What the subscriber needs from the running relay node: enumerate the tunnels it serves and tear a
+ * whole host's live streams/devices down (INV12). Kept narrow + injectable so the Phase-1 entry (B5)
+ * adapts the real data-plane {node, listener}: `activeTunnels` ← `listener.tunnels()` values,
+ * `closeStream(hostId)` ← `node.closeTunnel(hostId)` (the whole-host revocation lever).
+ */
+export interface RevocableNode {
+  /** Snapshot-friendly view of the tunnels currently attached to this node. */
+  activeTunnels(): Iterable<ActiveTunnelRef>
+  /** Immediately tear down every live stream/device on `hostId` (whole-host revocation, INV12). */
+  closeStream(hostId: string): void
+}
+
+/**
+ * Minimal structural view of an ioredis client in subscriber mode — only the members used here — so
+ * this module carries no hard dependency on ioredis and stays unit-testable with a fake. A real
+ * ioredis `Redis` instance satisfies this shape.
+ */
+export interface RedisSubscriber {
+  subscribe(channel: string): Promise<unknown>
+  unsubscribe(channel: string): Promise<unknown>
+  on(event: 'message', listener: (channel: string, message: string) => void): unknown
+  off(event: 'message', listener: (channel: string, message: string) => void): unknown
+}
+
+export interface RevocationSubscriberDeps {
+  readonly redisSubscriber: RedisSubscriber
+  readonly node: RevocableNode
+  /** Observability (metadata only, INV10): a valid signal was applied to `hostsAffected` hosts here. */
+  readonly onApplied?: (signal: KillSignal, hostsAffected: number) => void
+  /** Malformed-message counter — a dropped signal never fires a teardown (INV12 safety). */
+  readonly onDropped?: () => void
+  /** Boundary error routing for subscribe/unsubscribe failures — never silently swallowed. */
+  readonly onError?: (error: unknown) => void
+}
+
+export interface RevocationSubscription {
+  close(): void
+}
+
+function parseKillSignal(raw: string): KillSignal | null {
+  let json: unknown
+  try {
+    json = JSON.parse(raw)
+  } catch {
+    return null
+  }
+  const result = KillSignalSchema.safeParse(json)
+  return result.success ? result.data : null
+}
+
+export function startRevocationSubscriber(deps: RevocationSubscriberDeps): RevocationSubscription {
+  const { redisSubscriber, node, onApplied, onDropped, onError } = deps
+
+  const onMessage = (channel: string, message: string): void => {
+    if (channel !== RELAY_REVOCATIONS_CHANNEL) return // this client may also carry other channels
+    const signal = parseKillSignal(message)
+    if (signal === null) {
+      onDropped?.() // dropped + counted; bus stays alive, no teardown fired
+      return
+    }
+    // Snapshot first: closeStream tears down (mutates) the node's live-tunnel set as we iterate.
+    const tunnels = [...node.activeTunnels()]
+    let hostsAffected = 0
+    for (const { hostId, accountId } of tunnels) {
+      if (!killsScope(signal, accountId, hostId)) continue // blast-radius bounded; unrelated host = no-op
+      node.closeStream(hostId)
+      hostsAffected += 1
+    }
+    onApplied?.(signal, hostsAffected)
+  }
+
+  redisSubscriber.on('message', onMessage)
+  // Fire the SUBSCRIBE; route a rejected subscribe to onError (boundary I/O — never swallowed).
+  Promise.resolve(redisSubscriber.subscribe(RELAY_REVOCATIONS_CHANNEL)).catch((error: unknown) => {
+    onError?.(error)
+  })
+
+  let closed = false
+  return {
+    close(): void {
+      if (closed) return // idempotent teardown
+      closed = true
+      redisSubscriber.off('message', onMessage)
+      Promise.resolve(redisSubscriber.unsubscribe(RELAY_REVOCATIONS_CHANNEL)).catch(
+        (error: unknown) => {
+          onError?.(error)
+        },
+      )
+    },
+  }
+}
--- a/relay-run/src/wiring/route-resolver.ts
+++ b/relay-run/src/wiring/route-resolver.ts
@@ -0,0 +1,40 @@
+/**
+ * B3 · Store-backed RouteResolver — the PRODUCTION replacement for the one-entry in-RAM
+ * `memoryRouteResolver` (wiring/data-plane.ts:110). It implements term-relay's `RouteResolver`
+ * interface EXACTLY (`resolveSubdomain(subdomain) -> ResolvedHost | null`) by looking the tenant
+ * label up in the control-plane Postgres `hosts` store via `hosts.getBySubdomain(subdomain)`.
+ *
+ * Fail-closed (returns null, → 403 at the T8 upgrade edge) on:
+ *   - unknown subdomain: `getBySubdomain` returns null.
+ *   - revoked host: the row still exists (status is versioned, not deleted — INV8/INV12), so we
+ *     must reject `status === 'revoked'` here; a revoked host must never resolve to a route.
+ *
+ * The resolver is only a HINT for candidate lookup (subdomain-router.ts): the returned `hostId`
+ * is what `authorizeUpgrade` feeds to P5 as `requestedHostId`, and P5 gates the signed token's
+ * `host` against it (INV1). Identity (`accountId`/`hostId`) here comes solely from the CP store —
+ * the ownership source of truth — never from the caller (INV3).
+ */
+import type { HostStore } from 'control-plane/src/store/ports.js'
+import type { RouteResolver, ResolvedHost } from 'term-relay/data-plane/subdomain-router.js'
+
+/**
+ * The single HostStore capability this resolver needs (interface segregation): a subdomain lookup.
+ * A full `createPgStores().hosts` (`HostStore`) satisfies it structurally.
+ */
+export type HostSubdomainLookup = Pick<HostStore, 'getBySubdomain'>
+
+export interface StoreRouteResolverDeps {
+  readonly hosts: HostSubdomainLookup
+}
+
+/** Build a RouteResolver that resolves subdomain → host from the control-plane `hosts` store. */
+export function createStoreRouteResolver({ hosts }: StoreRouteResolverDeps): RouteResolver {
+  return {
+    async resolveSubdomain(subdomain: string): Promise<ResolvedHost | null> {
+      const host = await hosts.getBySubdomain(subdomain)
+      if (host === null) return null // unknown subdomain — fail closed
+      if (host.status === 'revoked') return null // revoked host — fail closed (INV12)
+      return { hostId: host.hostId, accountId: host.accountId, subdomain: host.subdomain }
+    },
+  }
+}
--- a/relay-run/src/wiring/stores-pg.ts
+++ b/relay-run/src/wiring/stores-pg.ts
@@ -0,0 +1,197 @@
+/**
+ * B1 — shared-store `EnforceDeps` for relay-auth's `onUpgrade` pipeline, backed by the SAME
+ * Postgres + Redis as the control-plane (P3). This replaces the Phase-0 in-RAM fakes
+ * (`memory-stores.ts`) so the data-plane and the control-plane read one world of truth
+ * (restart-safe, INV7).
+ *
+ * Design:
+ *  - hosts / sessions / audit -> Postgres via the CP repository adapter (`createPgStores(query)`).
+ *    The CP `HostRecord` IS the relay-auth `HostRecord` (both re-export the frozen relay-contracts
+ *    §4.2 shape), so `getById` is a direct pass-through; sessions/audit remap to the port shapes.
+ *  - revocation / buckets -> Redis. `consumeOnce` is a single-use `SET NX` burn (Finding-4); the
+ *    token bucket is an atomic Lua script (no read-modify-write race).
+ *
+ * SECURITY:
+ *  - INV3: `accountId` is only ever read from authenticated material (the CP registry rows / the
+ *    principal), never from client input. This adapter surfaces stored rows; it never fabricates an
+ *    accountId from a request field.
+ *  - Fail-closed: store/Redis errors are NOT swallowed — they reject and the enforcement pipeline
+ *    denies the upgrade (deny-by-default).
+ *  - INV9/INV10: audit rows carry metadata only (no keys, no terminal payload).
+ */
+import { createPgStores } from 'control-plane/src/store/pg.js'
+import type { QueryFn } from 'control-plane/src/db/pool.js'
+import type {
+  AuditEvent,
+  EnforceDeps,
+  HostRegistryPort,
+  RevocationStore,
+  SessionRegistryPort,
+  TokenBucketStore,
+} from 'relay-auth'
+import { NO_STEPUP_POLICY } from 'relay-auth/src/human/stepup/stepup.js'
+
+// ── Redis surface (ioredis-compatible, injected loosely so tests can pass a mock) ────────────────
+
+/**
+ * The minimal slice of an ioredis client this adapter uses. A real `ioredis` `Redis` instance
+ * structurally satisfies this (its methods are supersets), and a unit-test mock can implement it.
+ */
+export interface RedisLike {
+  exists(key: string): Promise<number>
+  set(key: string, value: string, mode?: 'NX'): Promise<string | null>
+  expireat(key: string, timestamp: number): Promise<number>
+  eval(script: string, numKeys: number, ...args: (string | number)[]): Promise<unknown>
+}
+
+export interface RelayEnforceDepsConfig {
+  readonly query: QueryFn
+  readonly redis: RedisLike
+}
+
+// ── Redis key namespaces (opaque to relay-auth) ──────────────────────────────────────────────────
+
+const REVOKED_PREFIX = 'revoked:' as const
+const USED_PREFIX = 'used:' as const
+const BUCKET_PREFIX = 'bucket:' as const
+
+/** Buffer added to the computed refill window before a fully-refilled bucket key may be reaped. */
+const BUCKET_TTL_BUFFER_SEC = 1
+
+/**
+ * Atomic token bucket (lazy refill). One round-trip, no read-modify-write race:
+ *   KEYS[1]=bucket key · ARGV: refillPerSec, burst(capacity), now(epoch sec), ttl(sec).
+ * Returns 1 when a token was available (and consumed), 0 when throttled.
+ */
+const TOKEN_BUCKET_LUA = `
+local key   = KEYS[1]
+local refill = tonumber(ARGV[1])
+local burst  = tonumber(ARGV[2])
+local now    = tonumber(ARGV[3])
+local ttl    = tonumber(ARGV[4])
+local state  = redis.call('HMGET', key, 't', 'ts')
+local tokens = tonumber(state[1])
+local ts     = tonumber(state[2])
+if tokens == nil then
+  tokens = burst
+  ts = now
+end
+local elapsed = now - ts
+if elapsed > 0 then
+  tokens = math.min(burst, tokens + elapsed * refill)
+end
+local allowed = 0
+if tokens >= 1 then
+  tokens = tokens - 1
+  allowed = 1
+end
+redis.call('HSET', key, 't', tokens, 'ts', now)
+redis.call('EXPIRE', key, ttl)
+return allowed
+`
+
+/** Seconds a bucket needs to fully refill from empty, plus a small reap buffer. */
+function bucketTtlSec(refillPerSec: number, burst: number): number {
+  const refillWindow = refillPerSec > 0 ? Math.ceil(burst / refillPerSec) : Math.ceil(burst)
+  return Math.max(1, refillWindow) + BUCKET_TTL_BUFFER_SEC
+}
+
+// ── Redis-backed ports ───────────────────────────────────────────────────────────────────────────
+
+function redisRevocationStore(redis: RedisLike): RevocationStore {
+  return {
+    async isRevoked(jti) {
+      return (await redis.exists(REVOKED_PREFIX + jti)) > 0
+    },
+    async revokeJti(jti, exp) {
+      await redis.set(REVOKED_PREFIX + jti, '1')
+      await redis.expireat(REVOKED_PREFIX + jti, exp)
+    },
+    async consumeOnce(jti, exp) {
+      // First-use wins: SET NX returns 'OK' only when the key did not exist (Finding-4).
+      const set = await redis.set(USED_PREFIX + jti, '1', 'NX')
+      if (set !== 'OK') return false
+      await redis.expireat(USED_PREFIX + jti, exp)
+      return true
+    },
+  }
+}
+
+function redisTokenBucketStore(redis: RedisLike): TokenBucketStore {
+  return {
+    async take(key, refillPerSec, burst, now) {
+      const ttl = bucketTtlSec(refillPerSec, burst)
+      const allowed = await redis.eval(
+        TOKEN_BUCKET_LUA,
+        1,
+        BUCKET_PREFIX + key,
+        refillPerSec,
+        burst,
+        now,
+        ttl,
+      )
+      return Number(allowed) === 1
+    },
+  }
+}
+
+// ── Postgres-backed ports (over the CP repository adapter) ────────────────────────────────────────
+
+/** Map a relay-auth `AuditEvent` onto the CP `audit_log` row shape (metadata only, INV10). */
+function toAuditRow(e: AuditEvent): {
+  action: string
+  principalId: string
+  accountId: string
+  hostId: string | null
+  ts: string
+  meta: Record<string, string>
+} {
+  const meta: Record<string, string> = {
+    outcome: e.outcome,
+    reason: e.reason,
+    remoteAddrHash: e.remoteAddrHash,
+  }
+  if (e.sessionId !== null) meta.sessionId = e.sessionId
+  if (e.jti !== null) meta.jti = e.jti
+  return {
+    action: e.action,
+    principalId: e.principalId,
+    accountId: e.accountId,
+    hostId: e.hostId,
+    ts: e.ts,
+    meta,
+  }
+}
+
+// ── Assembly ──────────────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Build the relay-auth `EnforceDeps` over the shared Postgres (`query`) + Redis (`redis`).
+ * `stepUpPolicyFor` returns `NO_STEPUP_POLICY` (staging single-operator; Phase 2 -> per-host WebAuthn).
+ */
+export function createRelayEnforceDeps({ query, redis }: RelayEnforceDepsConfig): EnforceDeps {
+  const stores = createPgStores(query)
+
+  const hosts: HostRegistryPort = {
+    // CP HostRecord === relay-auth HostRecord (both = relay-contracts §4.2) — direct pass-through.
+    getById: (hostId) => stores.hosts.get(hostId),
+  }
+
+  const sessions: SessionRegistryPort = {
+    getById: async (sessionId) => {
+      const rec = await stores.sessions.get(sessionId)
+      return rec === null ? null : { hostId: rec.hostId, accountId: rec.accountId }
+    },
+  }
+
+  return {
+    hosts,
+    sessions,
+    revocation: redisRevocationStore(redis),
+    buckets: redisTokenBucketStore(redis),
+    audit: {
+      append: (e) => stores.audit.append(toAuditRow(e)),
+    },
+    stepUpPolicyFor: () => NO_STEPUP_POLICY,
+  }
+}