[techmahedy]

Why

Two bugs affected anyone using the notifier in production, and the package lacked several things people expect from a notification system (mail, events, a test fake, delivery visibility).

Bug 1: a retry re-sent what had already worked. NotificationDispatcher ran every channel in one job: the database row first, then Slack, Discord and so on. If a later channel threw, the whole job was retried and the database notification was stored again each attempt. With ['database', 'slack'] and Slack down, one notification produced 3 database rows.

Bug 2: notifications() read the signed-in user's notifications. notifications(), unreadNotifications(), readNotifications() and markNotificationsAsRead() filtered by auth()->id(), not by the entity. An admin viewing another user saw their own notifications. In a queue worker or console command there is no signed-in user, and the call threw Auth actor [web] is not defined.

Also fixed: Slack failures were only logged, so the queue never retried them (Discord already threw). ChannelDriver threw a TypeError when constructed without the application.

What changed

Delivery

• One queued job per channel. A failing channel retries on its own.
• DatabaseChannel is now a real channel (the dispatcher no longer special-cases it).
• The dispatcher tries every channel in a legacy multi-channel job and throws the first failure at the end.
• send() returns the notification id (UUID), shared by that send's jobs.

New channels

• mail: structured content (greeting, lines, action button, outro, footer) or your own html/text. Supports cc, bcc, reply-to, attachments, tags, priority. All values are escaped and action links must be plain http(s).
• webhook: JSON body, custom headers, optional HMAC signature in X-Notification-Signature. Addresses must be http(s). Headers containing line breaks are dropped.
• Channels\Support\Http is a small injectable transport, so channels are testable without the network.

Notification API (all optional, defaults keep current behaviour)

• contentFor($channel, $notifiable) routes to toMail(), toSlack() and so on, falling back to content().
• shouldSendVia(), throttle(), priority(), uniqueId(), backoff(), onConnection().
• On the notifiable: wantsNotification() for per-person channel preferences, plus unreadNotificationsCount(), markNotificationAsRead($id), deleteNotification($id), deleteReadNotifications() and notificationDeliveries().

Observability and testing

• Delivery log: notification_deliveries, one row per notification and channel, with status, attempts and last error. It is best effort: when the table is missing it is skipped with a warning and sending is unaffected.
• Events: Notification::sending() (return false to cancel), sent(), failed(). A throwing listener is logged and ignored.
• Notification::fake() with assertSentTo, assertNotSentTo, assertSentTimes, assertSentVia, assertNothingSent, assertCount.
• DatabaseNotification::prune() and the notification:prune command.

Behaviour changes to review

• send() returns a UUID instead of a queue job id. toMany() returns one id per recipient.
• A notification with N channels is N jobs, not one. Jobs already queued in the old form still run, and the delivery log stops them repeating channels.
• The entity-scoped notification methods change meaning for any caller that relied on the signed-in-user behaviour. That behaviour was the bug.
• Slack failures now throw, so a bad webhook now fails and retries.
• ChannelDriver keeps its Application $app property type. Making it nullable would break subclasses that redeclare it, so a channel built without the app leaves the property unset instead.
• NotificationDispatcher gains a fourth constructor argument, $notificationId.

Migration

New migration create_notification_deliveries_table. Recommended, not required. Without it the log, "skip an already-delivered channel on retry" and throttle() are off. Everything else works.

Dependencies on doppar/queue

Priority, uniqueId(), backoff() and onConnection() use the newer queue's job methods. On an older queue they are accepted and ignored. Everything else works with any 4.x.

The queue's queue_jobs table also needs its new columns for the notifier to queue anything on a newer queue. That belongs to the queue package's own upgrade path.

Tests

• New suites cover notifiable scoping, each channel, the mail template and its escaping, the pipeline, the dispatcher (isolation, idempotency, skipping, throttling, missing log table), events, the fake, builders, queue features, a Redis driver suite, the prune command, and the migration on SQLite.
• Redis tests need predis/predis and a reachable server. They skip otherwise.
• Tests that need the newer queue skip against older installs.
• The regressions were checked by breaking each fix on purpose (idempotency, per-channel jobs, entity scoping, Slack throwing, throttling, preferences, event cancellation, webhook signing, mail link validation). Each made at least one test fail.