orca/docs/reference/mobile-push-contract.md

# Mobile push: contract and build spec

Tracking issue: stablyai/orca#8129. Design page: `/tmp/orca-mobile-push/orca-mobile-push.html`.
This document is the single contract every lane builds against. Do not deviate without updating it.

## Summary

A small Orca-hosted push gateway (`cloud/apps/push`) holds the APNs key and FCM credentials and sends
to phones. The desktop host registers each paired phone's native push token with the gateway and asks
the gateway to push each eligible desktop notification. Native APNs/FCM delivery is the only ordinary
mobile OS-banner path. The notification socket is retained only for live dismissals and reconnect tray
reconciliation; ordinary notification frames never create banners. Reconciliation compares the current
native tray with durable host dismissal history and does not replay alerts. Desktop notification
categories are authoritative. Legacy category, replay-notification, and summary fields remain on the
wire only for mixed-version compatibility. No ack gate, no generic mode, no staging gateway, one auth
path for signed-in and accountless hosts.

## Identities

- **Host public key**: the desktop's existing X25519 E2EE public key (`src/main/runtime/e2ee-keypair.ts`),
  32 bytes, base64. The phone already stores it per host as `publicKeyB64`.
- **hostFingerprint**: `sha256(hostPublicKey)` base64url, first 16 chars. Identical derivation to
  `deriveRelayHostId` in `src/main/runtime/relay/relay-http-client.ts`. Both desktop and phone can compute it.
- **deviceId**: the desktop's `DeviceEntry.deviceId` for the paired phone. Opaque UUID.
- **registrationId**: gateway-assigned opaque id for one (hostFingerprint, deviceId) pair.

## Gateway HTTP API

Base URL: `https://push.onorca.dev` (dev override via env). JSON bodies, `Content-Type: application/json`.
All schemas are zod, `.strict()`, exported from `cloud/packages/push-contract`.

### Host authentication: challenge, proof, session

The host keypair is X25519 (box), so it cannot sign. Reuse the relay's challenge shape.

`POST /v1/host/challenge`

```json
{ "v": 1, "hostPublicKeyB64": "<32 bytes b64>" }
```

→ 200

```json
{ "challengeId": "<opaque>", "gatewayEphemeralPublicKeyB64": "<32 b64>", "nonceB64": "<24 b64>",
  "ciphertextB64": "<b64>", "expiresAt": <epoch ms> }
```

- Gateway generates an ephemeral box keypair per challenge, a 24-byte nonce, and a 32-byte secret.
- `plaintext = "orca-push-host-challenge/v1\0" || u32be(len(transcript)) || transcript || secret(32)`
- `ciphertext = nacl.box(plaintext, nonce, hostPublicKey, gatewayEphemeralSecretKey)`
- Transcript is the relay's length-prefixed field encoding (`field(name, value)` =
  u32be(len(name)) || name || u32be(len(value)) || value), fields in this exact order:
  `protocol="orca-push-host-proof/v1"`, `version=0x01`, `gatewayOrigin`, `gatewayEphemeralPublicKey`,
  `challengeNonce`, `challengeId`, `issuedAt` (u64be ms), `expiresAt` (u64be ms), `hostFingerprint`,
  `hostPublicKey`.
- Challenge TTL 10 s, and 10 s is the whole window the gateway honours. The 30 s clock skew tolerance
  is the host's alone: it validates a timestamp the gateway chose, so it needs the allowance and the
  gateway does not. A gateway that subtracted the tolerance from its own check would run a 40 s TTL.
  Store challenge (id, secret hash, host fingerprint, host public key, expiry) in DB so any Cloud Run
  instance can verify. Expired rows are pruned 30 s late so a slow proof reads as expired rather than
  as an unknown challenge.
- Issuing a challenge writes no `push_hosts` row. It is unauthenticated, so a `push_hosts` row would be
  a free permanent write for any caller. The row is upserted in `POST /v1/host/session` once the proof
  verifies, from the public key the challenge row carries.

`POST /v1/host/session`

```json
{ "v": 1, "challengeId": "<opaque>", "proofB64": "<32 b64>" }
```

- Host opens the box with its secret key, validates every transcript field (same checks as
  `validateTranscript` in `src/main/runtime/relay/relay-host-proof.ts`, adapted to the push fields),
  and returns `proof = HMAC-SHA256(secret, "orca-push-host-proof/v1\0ack\0" || transcript)`.
- Gateway verifies with `timingSafeEqual`, consumes the challenge (single use), and returns

```json
{ "sessionToken": "<opaque 32 b64url>", "expiresAt": <epoch ms>, "hostFingerprint": "<16 chars>" }
```

- Session TTL 24 h. Stored hashed (sha256) in DB. Bearer on every other call:
  `Authorization: Bearer <sessionToken>`. 401 with `{ "error": "session_expired" }` on expiry; host
  re-runs the challenge.

### Device registration

`POST /v1/devices` (Bearer)

```json
{ "v": 1, "deviceId": "<uuid>", "platform": "ios" | "android", "token": "<native token>",
  "apnsEnvironment": "sandbox" | "production",   // ios only, required for ios
  "filter": { "sources": ["agent-task-complete", "terminal-bell", "plugin"],
              "agentStates": ["needs-input", "finished"] } }
```

→ 200 `{ "registrationId": "<opaque>" }`. Upsert keyed by (hostFingerprint, deviceId); a new token
replaces the old. `deviceId` is caller-chosen, so a host is capped at 64 registrations: the 65th
distinct `deviceId` → 409 `{ "error": "too_many_devices" }`. Re-registering a `deviceId` the host
already owns is always accepted, and deleting a registration frees its slot. `GET /v1/devices` is
bounded at 1024 rows to match its response schema, which the per-host cap keeps well out of reach.
The gateway `filter` contains only the legacy `sources` and `agentStates` arrays. It remains stored and
accepted for mixed-version schema compatibility but is not an updated delivery-policy input. The host
persists and enforces desktop eligibility plus the phone-specific away-only, sound, and expiry settings.
iOS tokens are variable-length, hex-encoded byte strings; Android tokens are FCM registration strings.

`DELETE /v1/devices/:registrationId` (Bearer) → 204. Only the owning host may delete.

`GET /v1/devices` (Bearer) → `{ "devices": [{ registrationId, deviceId, platform, dead: boolean }] }`.

### Send

`POST /v1/send` (Bearer)

```json
{ "v": 1,
  "registrationIds": ["<id>", "..."],
  "notification": {
    "notificationId": "<max 2048 chars, may be absent for terminal-bell>",
    "notificationSeq": <int>, "notificationEpoch": "<uuid>",
    "source": "agent-task-complete" | "terminal-bell" | "plugin",
    "agentState": "needs-input" | "finished" | null,
    "title": "<max 80 chars>", "body": "<max 180 chars>",
    "worktreeId": "<max 2048 chars|absent>" } }
```

→ 200

```json
{ "results": [{ "registrationId": "<id>", "status": "queued" | "dead" | "rate_limited" | "error" }] }
```

- `queued` means the logical event, recipient, and pending delivery payload have committed to SQL. A
  worker resumes pending work after restarts; provider acceptance is not proof of visible delivery.
- Each host gets 300 logical alerts and, independently, 300 dismissals per rolling 15 minutes. Fanout
  across phones counts the event once. There is no daily per-phone quota. Over quota returns
  `rate_limited` per result with HTTP 200; requests above 20 registration IDs return HTTP 400.
- Optional notification `kind` is `alert` (default) or `dismiss`; optional `expiresAt` is an absolute
  Unix-millisecond deadline capped at five minutes from first acceptance. These fields require the
  updated gateway; deploy the gateway before the updated desktop sender.
- Notification JSON remains limited to 3000 UTF-8 bytes. Duplicate event identity with changed content
  is rejected. Retries retain the original deadline and do not spend additional quota.
- Event and recipient identities/outcomes remain for 24 hours; payloads are cleared at completion or
  expiry by minute-level cleanup. Quota reservation and enqueue share a per-host SQL transaction.

### Request limits and unauthenticated abuse

- Every POST is capped at 16 KiB by a streaming body limit, not by `Content-Length` alone: a chunked
  body declares no length. Over the cap → 413 `{ "error": "request_too_large" }`.
- `POST /v1/host/challenge` and `POST /v1/host/session` are the only unauthenticated routes. They share
  one token bucket per client IP, 30 requests per minute, refilling continuously. Over the bucket → 429
  `{ "error": "rate_limited" }`. The client IP is the **last** `x-forwarded-for` hop, not the first:
  Cloud Run appends the connecting peer, so everything left of that value is caller-supplied and can be
  a fresh forgery on every request, which would hand a flood a new bucket each time.
  `ORCA_PUSH_TRUSTED_PROXY_HOPS` (default 0) says how many appenders sit between the platform and the
  client, so a future load balancer sets it to 1. A header with fewer hops than that depth is not
  trusted at all. Falls back to `x-real-ip` and then to a single shared bucket. The bucket is per
  instance and in memory, so the effective cap scales with the instance count; it exists to blunt a
  flood, not to meter.
- Authenticated routes use a per-host bucket of 600 requests/minute per instance, so normal usage by
  one desktop does not consume another desktop's budget behind the same office IP. Invalid bearers
  spend a separate 30/minute IP budget. Session lookups have bounded concurrent and waiting admission.
- The gateway cannot prove that a host owns the token it registers: any host with a session may
  register any well-formed token and send text to it, within its own quota. The phone drops such a push
  in the foreground because the fingerprint resolves to no paired host, and never routes a tap on it,
  but the OS banner shows while the app is backgrounded. Reaching it needs the victim's native token,
  which the gateway never returns and which only the phone and its host ever see.

### Individual delivery (gateway)

Each accepted event immediately creates one queued delivery per eligible registration. Bursts retain
their original title, body, routing fields, and replacement identity; the gateway does not generate
summary text or summary membership. APNs groups alerts visually by the host thread identifier. Android
uses a distinct notification tag for each event and relies on platform behavior rather than a custom
summary notification. Quota accounting still counts logical alerts, not deliveries or recipients.

Four worker lanes per instance claim deliveries with expiring, renewed SQL leases. Retry state
is persistent, with exponential backoff and provider minimum delays. Retry-After is never shortened
to fit event lifetime: expire instead. Device validity is checked before each attempt. Dismissals
cancel the matching pending alert and use silent provider messages. Mobile OS
background execution remains best effort, particularly after force-quit on iOS.

The queue table retains its historical `push_delivery_batches` name to avoid a data migration. New
rows keep a one-element JSON array only as a rolling-deploy storage envelope so an older worker can
read them, but new code models and sends one notification. New rows are immediately due, which keeps
an overlapping older gateway from appending to them. A new worker atomically splits a pre-deployment
multi-event row into individual rows, preserving its fixed expiry, attempt count, and existing event
recipient records; legacy rows were capped at 32 events, bounding that transaction. A legacy row
already leased by an older revision completes under that revision's behavior, while immediately due
new rows cannot be appended to by its admission path.

Individual presentation is guaranteed only after every older worker revision has retired. During
the overlap, an older worker can still send a pre-existing row as a summary and can assign its former
host-wide collapse identity to a new singleton identity-less bell. Do not add fabricated IDs or new
wire fields to conceal old-binary behavior. Post-retirement acceptance must verify that two alerts for
one host retain independent provider replacement identities and can be dismissed independently.

Shutdown stops admission and work acquisition, waits for active work within the platform grace, and
leaves unfinished deliveries recoverable after their leases expire. A provider acceptance followed by a
crash before SQL completion can still cause a repeated send; collapse identity mitigates this without
promising exactly-once delivery.

### Provider payloads

APNs (HTTP/2, `api.push.apple.com` or `api.sandbox.push.apple.com` by `apnsEnvironment`; JWT auth
from key id + team id + `.p8`, token cached and refreshed every 50 min):

- headers: `apns-topic: com.stably.orca.mobile`, `apns-push-type: alert`, `apns-priority: 10`,
  `apns-expiration: fixed event deadline (at most five minutes)`, `apns-collapse-id: <sha256(host + notification identity)>`
  (identity-less events use their epoch and sequence)
- body: `{"aps":{"alert":{"title","body"},"sound":"default","thread-id":"<hostFingerprint>"},
"orca":{ hostFingerprint, worktreeId, notificationId, notificationSeq, notificationEpoch, source,
agentState }}`
- Dismissals use `apns-push-type: background`, priority `5`, no collapse header, and
  `aps: {"content-available": 1}` with `orca.kind: "dismiss"`. They carry no alert or sound.
- Dead token: 410, or 400 with `BadDeviceToken`/`Unregistered`/`DeviceTokenNotForTopic`.

FCM (V1 `projects/onorca-cloud/messages:send`, bearer from the runtime service account via the GCE
metadata server or `GOOGLE_APPLICATION_CREDENTIALS` locally):

- `{"message":{"token","notification":{"title","body"},"android":{"priority":"HIGH","ttl":"<remaining event lifetime, at most 300s>",
"collapse_key":"<sha256(collapseId) hex 32>","notification":{"channel_id":"orca-desktop","tag":"<collapseId>"}},
"data":{ all orca fields as strings }}}`
- Dismissals are data-only (`kind: "dismiss"`); omit both `message.notification` and
  `android.notification`. No visible alert or sound is requested.
- FCM notification messages are inherently collapsible while offline, and FCM supports only a small
  number of concurrent collapse keys per device. Per-event `collapse_key` and Android `tag` preserve
  individual replacement identity while a message is retained, but excess offline pending messages
  may be discarded and every alert is not guaranteed to appear. Android automatic grouping remains
  platform-owned and is unverified on physical devices. Socket reconnect never recovers missed OS
  banners; the tray is not an event log.
- Dead token: `UNREGISTERED`, or `INVALID_ARGUMENT` whose message names the token.

### Gateway storage (Postgres in prod, SQLite in tests, same pattern as `cloud/apps/relay/src/database.ts`)

- `push_hosts(host_fingerprint pk, host_public_key, created_at, last_seen_at)`, written only on a
  verified proof and pruned after 1 h of no contact when no `push_devices` row still names the host.
  Nothing reads it, and any keypair mints a host for free, so it is not allowed to accumulate.
- `push_sessions` holds one row per host, enforced by a unique index and transaction lock. Minting a
  session deletes the host's earlier one, since a desktop holds a single session and only re-proves once it is gone.
- `push_challenges(challenge_id pk, host_fingerprint, host_public_key, secret_hash, transcript,
expires_at, consumed_at)`
- `push_sessions(token_hash pk, host_fingerprint, expires_at, created_at)`
- `push_devices(registration_id pk, host_fingerprint, device_id, platform, token, apns_environment,
filter_json, dead_at, created_at, updated_at, unique(host_fingerprint, device_id))`
- `push_events` holds logical event identity, content fingerprint, quota timestamp, and expiry.
- `push_event_recipients` records accepted event/phone pairs for idempotent fanout.
- `push_delivery_batches` holds individual payload envelopes, retry deadlines, and renewable worker
  leases.
- `push_dismissed_events` fences older alerts from replaying after dismissal.
- Queue identities and dismissal fences are retained for 24 hours; completed payloads are cleared.

Logging: aggregate counters only. Never log tokens, titles, bodies, or raw fingerprints (log the first
4 chars of a fingerprint at most).

### Gateway env

`PORT`, `ORCA_PUSH_PUBLIC_URL`, `ORCA_PUSH_DATABASE_URL` (absent → SQLite under `ORCA_PUSH_DATA_DIR`),
`ORCA_PUSH_APNS_KEY` (PEM text), `ORCA_PUSH_APNS_KEY_ID`, `ORCA_PUSH_APPLE_TEAM_ID`,
`ORCA_PUSH_APNS_TOPIC` (default `com.stably.orca.mobile`), `ORCA_PUSH_FCM_PROJECT_ID` (default
`onorca-cloud`), `ORCA_PUSH_TRUSTED_PROXY_HOPS` (default 0,
proxies appending to `x-forwarded-for` after the client).
Secret Manager names (already exist in `onorca-cloud`): `orca-cloud-push-apns-key`,
`orca-cloud-push-apns-key-id`, `orca-cloud-push-apple-team-id`. Runtime SA:
`orca-cloud-push@onorca-cloud.iam.gserviceaccount.com` (already has FCM admin + secret accessor).

## Desktop (`src/main`, `src/shared`)

- Capability `NOTIFICATIONS_REMOTE_PUSH_RUNTIME_CAPABILITY = 'notifications.remote-push.v1'` in
  `src/shared/protocol-version.ts`, advertised statically.
- RPC `notifications.registerPush` params `{ platform, token, apnsEnvironment?, filter }`. The mobile
  filter includes away-only, expiry, and sound settings plus the legacy category fields; the host
  persists that full policy, while its gateway `POST /v1/devices` forwards only `sources` and
  `agentStates` for compatibility. `deviceId` comes from `ctx.pairedDeviceId`. The RPC returns
  `{ registered: true, registrationId } | { registered: false, reason: 'gateway_unreachable' |
'gateway_rejected' | 'not_mobile' | 'registration_storage_failed' | 'throttled' }`. A device may
  register at most 10 times per minute (`throttled` beyond that, its earlier registration untouched):
  each call is a gateway write plus a synchronous registry write on the main thread, and a paired
  phone could otherwise loop it. The unregister RPC is not throttled, since with nothing registered it
  is a lookup and with something registered it can only run once per successful register. The params
  schema is strict, so a caller-supplied `deviceId` is an error, not a key silently dropped. Persists
  `pushRegistration: { registrationId, platform, filter, registeredAt }` on `DeviceEntry` in
  `device-registry.ts` (new
  optional field, tolerated by old registries). When the gateway accepted the token but the host could
  not store it — the device left mobile scope mid-call (`not_mobile`) or the registry write threw
  (`registration_storage_failed`) — the host queues the gateway delete in the unregister outbox rather
  than leaking a registration nothing will ever push to. Registration, unregister, and outbox deletes
  are serialized per device; re-registration first settles earlier cleanup. Authentication failure
  never drops a durable delete. Stale send responses only clear the exact local registration observed,
  while provider dead-token updates match the token/platform/environment that was sent. Phones must
  treat any `registered: false` as "retry later", so an unknown reason string is safe to add.
- RPC `notifications.unregisterPush` params null → `{ unregistered: boolean }`. Removes the field and
  enqueues a gateway delete in a durable outbox (`src/main/runtime/push/push-unregister-outbox.ts`,
  modelled on `relay-revoke-outbox.ts`). Unpair/revoke (`revokeMobileDevice`) enqueues the same. The
  drain re-reads the queue as it goes, so a delete queued mid-drain lands in the same pass, and a pass
  that leaves retryable items schedules an unref'd backoff retry (30 s, doubling, capped at 10 min)
  instead of waiting for the next launch.
- Both RPCs added to `runtime-rpc-mobile-method-allowlist.ts`.
- Push client `src/main/runtime/push/push-gateway-client.ts`: challenge/proof/session with token cache,
  register, delete, send. Node `fetch`. Gateway URL from `profile-cloud-auth-config.ts`
  (`pushGatewayUrl`, default `https://push.onorca.dev`, env override `ORCA_PUSH_GATEWAY_URL`).
- Host proof answering: new `src/main/runtime/push/push-host-proof.ts`, a copy of the relay's
  `answerRelayHostChallenge` with the push transcript fields. Shared code with the relay proof is
  welcome if it stays a pure refactor.
- Dispatch hook: in `RuntimeMobileNotificationController.dispatch`, after the socket fan-out, call
  `pushDispatcher.enqueue(eventWithSeq)`. The dispatcher requires desktop category eligibility, applies
  each device's phone-specific away-only and sound preferences, skips `dismiss` events, maps `agentState`
  to `needs-input | finished` (blocked/waiting → needs-input, else finished),
  batches matching registrationIds into `POST /v1/send` requests of at most 20 registrations each (the
  gateway's per-request cap; extra devices get their own request rather than being dropped), and drops
  unchanged registrations the gateway reports `dead`. Failure categories are counted without payload
  values and logged at most once per minute (with a final flush on shutdown). Fire-and-forget with
  one retry after 2 s per request; never throws into dispatch.
- Add `agentState` to `MobileNotificationDispatchEvent` and set it in `src/main/ipc/notifications.ts`
  from `args.agentState`. Fix `buildAgentTaskCompleteNotificationOptions` so `working|running|busy`
  never yields "finished" (title says "working" and the dispatcher treats it as not-final, i.e. no push).
- Headless serve: no renderer means no `notifications:dispatch`. Document in
  `docs/reference/headless-linux-server.md`; do not fix here.

## Mobile (`mobile/`)

- Commit `google-services.json` (from `/tmp/orca-mobile-push/google-services.json`) at `mobile/` and set
  `"android": { "googleServicesFile": "./google-services.json" }` in `app.json`. Add `"expo-notifications"`
  to `plugins` so prebuild writes the `aps-environment` entitlement.
- Token: `Notifications.getDevicePushTokenAsync()`; `data` is the APNs hex or FCM string. iOS
  `apnsEnvironment`: `__DEV__ ? 'sandbox' : 'production'` (dev-client builds are debug, TestFlight and
  App Store are release). Listen with `addPushTokenListener` and re-register on change.
- Settings (`mobile/app/notifications.tsx`): one default-off **Enable notifications** switch
  controls native push registration. Hint: “Get agent alerts even when the app is closed.
  Delivered through Orca’s push service and Apple or Google.” Desktop category controls are
  authoritative and are not duplicated as phone overrides. Phone sound and viewing controls remain
  independent. **Only when away from desktop** defaults on (180 seconds of OS input idle,
  or locked). Unknown/headless presence does not suppress; it is never inferred from remote CPU
  activity. The detailed payload disclosure remains in the notification documentation.
- `notifications.delivery-policy.v1` advertises the away and mobile-inactivity lease policy.
  Filter flags are optional and ignored by older hosts; the UI identifies paired hosts requiring
  an update. New hosts preserve old phones’ existing policy when these flags are absent.
- Registrations with `expireAfterInactivity` receive a persisted seven-day `expiresAt` on the
  paired desktop. Delivery and transport retries exclude expired registrations. Only foreground
  mobile registration renews it: on connection, foreground return, and every 15 minutes while
  active. Background sockets, desktop use and notification delivery never renew a phone lease.
  Opening mobile and reconnecting restores delivery without replaying expired push sends.
- Registration: on switch-on (after OS permission), and on every host reaching `connected` while the
  switch is on, call `notifications.registerPush` on that host if it advertises the capability. On
  switch-off call `notifications.unregisterPush` on every connected host and remember to retry on hosts
  that were offline. On host removal, best-effort unregister before deleting credentials.
- Receive: `addNotificationReceivedListener` (foreground) validates that the host is paired, master
  consent is enabled, the destination is not currently viewed, and the event is not fenced by a
  persisted dismissal. A bounded process-local identity claim suppresses concurrent duplicates by
  host, epoch, sequence, and optional notification ID. Background and killed delivery remains owned by
  the OS and is best effort.
- Tap: `data.orca.hostFingerprint` → hostId by computing the same sha256/base64url/16 derivation over each
  stored host's `publicKeyB64`; then existing `getNotificationNavigationTarget` + `useOpenNotificationRoute`.
- Reopen: subscribe to socket notifications for live dismissals, but ignore ordinary
  notification frames for banner presentation. Reconnect reconciliation sends identities currently in
  the native tray and applies returned dismissal decisions; it does not replay notifications or create
  banners. Live dismiss events also remove matching presented notifications.
- Old host without the capability: nothing changes.

## Infra (`cloud/infra/terraform`, `.github/workflows`)

- Cloud Run service `orca-cloud-push`, region `us-central1`, project from the environment tfvars, runtime
  SA `orca-cloud-push@<project>.iam.gserviceaccount.com` (exists in prod; declare and import), the three
  secrets mounted as env (exist; declare and import), Cloud SQL connector to the dedicated HA PostgreSQL instance with its
  own database `orca_push`, min instances 1, max 2 in production, concurrency 80, ingress all, unauthenticated invoke.
- IAM: `roles/firebasecloudmessaging.admin` and `roles/serviceusage.serviceUsageConsumer` on the runtime
  SA (exist in prod; declare and import). Secret accessor per secret.
- Hostname `push.onorca.dev`. The DNS zone lives in the apps root in `stablyai/orca-cloud`; add the
  Cloud Run domain mapping here and leave a TODO comment naming the record the other repo must add.
- Workflow `.github/workflows/cloud-push-deploy.yml`: gated on `vars.ORCA_CLOUD_OPERATIONS_ENABLED`,
  Workload Identity like `cloud-relay-*`, builds a reviewed full `source_sha`, deploys with `--no-traffic`, probes the new
  revision's `/ready` and a validate-only FCM send, then shifts 100% traffic. Uses
  `.github/actions/cloud-sql-rollout-lease` around the schema step.
- Add the new root files to `cloud/dev/contracts` and `cloud/dev/fixtures` partitions so
  `terraform-root-partition.test.mjs` and `Cloud Verify` pass.

## Non-goals for this release

Ack gate, generic-alert mode, staging gateway, iOS Notification Service Extension, Android data-only
alert messages, Live Activities, account-based quota tiers.

### Device delivery preferences

The desktop advertises `notifications.delivery-preferences.v1`. Completion detection remains active
when desktop notifications are off; semantic validity checks still precede delivery. IPC publishes
`desktopAllowed: false` when the desktop master or source/category switch rejects an event. That
desktop category decision is authoritative for both desktop and phone alerts. Desktop focus and
native authorization remain desktop-only presentation gates and do not change mobile eligibility.

Updated mobile subscribes with `includeDesktopSuppressed: true` for mixed-version compatibility and
to receive dismissals, but it never turns ordinary socket notification frames into OS banners. The
optional subscribe/replay fields, replayed `notifications`, `followDesktop`, `sources`, and
`agentStates` remain accepted and populated only for mixed-version compatibility. Updated hosts ignore
the mobile category fields and gate provider alerts on `desktopAllowed`; old hosts and clients retain
their previous behavior without a wire break.

The mobile registration always sends `followDesktop: true` and complete legacy category arrays.
There are no active phone category overrides or category defaults. Optional `emittedAt` and replay
notification response fields remain compatibility surface rather than a second delivery policy or
banner path.

`filter.sound` is also host-local. False groups that device's requests separately and adds
optional `notification.sound: false` to gateway sends. The gateway omits APNs `aps.sound` and
uses Android's `orca-desktop-silent` channel. Missing sound preserves existing audible delivery.
Deploy the updated gateway before distributing hosts that send the optional sound field: older
gateways strictly reject unknown notification fields. No token or database migration is needed.

The phone's master switch disables native push registration. Sound and viewing preferences belong to
the receiving phone. The phone suppresses a foreground banner for its
currently viewed host/workspace only while active; it never assumes desktop focus means the
phone is viewing that workspace. Once registered, the host applies its persisted phone settings while
the phone is disconnected; preference changes synchronize when it reconnects. No live APNs/FCM
delivery is implied by simulator notification injection.

APNs/FCM is the sole ordinary OS-banner path whether the app is foregrounded, backgrounded, or
killed. Socket notification events do not wait, schedule a local fallback, or recover a missed native
alert. Hosts without push registration therefore have no mobile OS-banner fallback.

Native notification readers accept Expo's iOS `request.trigger.payload` as well as
`request.content.data`. APNs custom fields can exist only in the former; foreground identity claims,
dismissal, reconciliation, and tap routing all use the same reader. Legacy summary members remain
readable only for notifications already delivered during a mixed-version transition.

### Dismissal recovery and desktop presence

Automatic acknowledgement of a visible agent pane requires confirmed desktop presence from the
same three-minute native idle check used for push delivery. A focused window alone is insufficient.
Trusted input in the renderer confirms the user has returned. Unknown presence preserves unread
attention; explicit mark-read actions remain available, including in browser clients.

The runtime persists notification identities and dismissal sequence fences in its own user-data
directory before fanout. History is bounded to 4,096 records retained for seven days. It stores no
notification text or push tokens. On reconnect, mobile sends up to 256 `deliveredPushes` identities
from the current native tray in each `notifications.getMissedSince` request; updated hosts return
optional `dismissedPushes` for confirmed handled identities. Mobile processes only those dismissal
decisions. Unknown IDs, newer sequences, and different epochs are preserved. Older hosts ignore the
optional request field, and older clients ignore the additional response fields. Legacy replayed
`notifications` and epoch fields remain wire-compatible but do not drive current mobile banners.

On iOS, a local Expo module handles silent dismissals directly through the native notification
center, independent of JavaScript initialization. Native and JavaScript dismissal paths use the
same host/epoch/sequence fences; native dismissal fences retain up to 512 entries for 24 hours. Older
native shells and Android retain the JavaScript implementation. A native callback test proves
processing only when invoked: iOS background push delivery remains best effort, including while
suspended or force-quit. Narrow legacy decoding keeps already delivered summary notifications from
being mistaken for individual events during the transition. Reconciliation inspects up to 2,048
represented identities in pages of 256. It has no stored replay watermark: every connection compares
the current tray with host dismissal history and never replays an alert.