<!-- Generated from models/_canonical by scripts/build-model-variants.js. Edit the canonical source, not this file. Behavioural profile for Claude: scripts/model-profiles.json -->

<critical_constraints> FORBIDDEN: Truncating code or writing placeholders such as "// ... existing code ..." or "# rest unchanged". Every edit is complete and applies as written. FORBIDDEN: Reporting a check as passed without showing the command and its output. REQUIRED: Reason through the rules below before the first edit; when two rules conflict, the one stated first wins. </critical_constraints>

---

#Purpose

Rules for notifying users across channels. The engineering problem is not sending to any one channel — it is having one event produce the right message on the right channels for each user, without duplicates and without becoming noise.

Channel-specific delivery is Backend/email. This package is the layer above.

---

#One event, many channels

Producers emit an event. They do not decide the channel, the copy, or whether the user wants it.

ts
await notify({
  type: "order.shipped",              // the event, not "send an email"
  recipientId: order.userId,
  tenantId: order.tenantId,
  data: { orderId: order.id, carrier, trackingUrl },
  idempotencyKey: `order.shipped:${order.id}`,
  occurredAt: new Date(),
});

The notification service then resolves, in order:

1. Preferences — does this user want order.shipped, on which channels?
2. Templates — copy per event type per channel, per locale.
3. Deduplication — has this idempotencyKey already been delivered?
4. Batching — immediate, or held for a digest?
5. Delivery — enqueue per channel, with per-channel retry.

Calling sendEmail() from a business service couples the domain to a channel and guarantees that the next channel means editing every call site.

---

#Preferences, and the ones you cannot override

A preference matrix of event type × channel, with sane defaults:

Two categories can never be disabled: security notices and legal/billing notices. Everything else is the user's choice, and the default must be conservative — a noisy default trains users to mute the whole channel, including the ones that matter.

• Honour quiet hours in the user's timezone, not the server's.
• Every non-forced notification carries an unsubscribe path that works in one click and is honoured immediately. List-Unsubscribe for email is required by major providers at volume.
• Store preference changes with a timestamp — "I never agreed to this" is a compliance question. → Security/audit-log

---

#Deduplication and batching

Duplicate notifications are the fastest route to a muted channel.

• Deduplicate on idempotencyKey within a window. Retries, replays and at-least-once queues all cause repeats. → Backend/queues
• Collapse related events: fifteen comments on one thread is one notification saying "15 new comments", not fifteen pushes.
• Digest low-urgency types: hold and send hourly or daily, at a time chosen in the user's timezone.
• Suppress a notification about an action the user just performed themselves.
• Rate-limit per user per channel as a hard backstop, regardless of event volume.

Concretely: a collapseKey groups events that supersede one another (thread:${threadId}), a digestWindow holds them (PT1H, P1D), and a maxPerHour per (recipientId, channel) caps the total. FCM's collapse_key and APNs' apns-collapse-id do the equivalent at the device, replacing an undelivered notification rather than stacking another one on the lock screen.

Urgency decides the path:

---

#Channel realities

ts
// Per-channel adapters behind one interface. A permanent rejection deletes the
// destination; a transient one retries. Conflating them is how dead tokens
// accumulate and how real failures get silently dropped.
const PERMANENT = new Set(["NotRegistered", "InvalidRegistration",
                           "Unregistered", "BadDeviceToken", "hard_bounce"]);

async function deliver(channel: Channel, msg: Message) {
  try {
    return await adapters[channel].send(msg);
  } catch (err) {
    if (PERMANENT.has(err.code)) { await suppress(msg.recipientId, channel, err.code); return; }
    throw err;                                  // transient — the queue retries
  }
}

Push token hygiene matters: an accumulating list of dead tokens slows every send and skews delivery metrics. Delete on the first permanent rejection.

---

#Observability and debugging

"Did the customer get the email?" must be answerable in one query.

Record per notification: event type, recipient, channel, template version, decision (sent / suppressed / deduplicated / digested), provider message id, delivery status, and open or click where available.

sql
-- One row per notification decision. This table is the answer to
-- "did the customer get it?", and it must be written even when nothing is sent.
CREATE TABLE notification_deliveries (
  id               uuid PRIMARY KEY,
  event_type       text        NOT NULL,
  recipient_id     uuid        NOT NULL,
  channel          text        NOT NULL,   -- email | push | sms | in_app
  decision         text        NOT NULL,   -- sent | suppressed | deduplicated | digested
  reason           text,                   -- 'preference_off', 'quiet_hours', 'hard_bounce'
  idempotency_key  text        NOT NULL,
  provider_msg_id  text,
  status           text,                   -- queued | delivered | bounced | failed
  created_at       timestamptz NOT NULL DEFAULT now()
);
CREATE UNIQUE INDEX ON notification_deliveries (idempotency_key, channel);
CREATE INDEX ON notification_deliveries (recipient_id, created_at DESC);

The unique index on (idempotency_key, channel) is the deduplication mechanism itself, not just a record of it — an ON CONFLICT DO NOTHING insert that affects zero rows means the notification was already delivered.

Provide an internal view showing every notification for a user with the reason it was or was not sent. Without it, every "I didn't get it" report becomes a database archaeology session.

---

#Anti-patterns

---

#Checklist

• Producers emit events; the notification service decides channels
• Preferences are modelled as event type × channel with conservative defaults
• Security and billing notices cannot be disabled
• Quiet hours use the user's timezone
• Unsubscribe is one click and honoured immediately
• Preference changes are recorded with timestamps
• Notifications are deduplicated on an idempotency key
• Related events collapse into a single notification
• Low-urgency types are digested; critical types bypass digesting
• Self-caused notifications are suppressed
• A per-user, per-channel rate limit exists as a backstop
• Push tokens are deleted on permanent rejection
• Push payloads carry identifiers, not full content
• SMS is reserved for OTP and critical alerts, with consent rules checked
• In-app notifications have read state and a retention policy
• Every notification records channel, decision, provider id and outcome
• Delivery failure, suppression and unsubscribe rates are alerted on
• An internal per-user view explains why each notification was or was not sent