# Run notification audit fixes `notify_mode=auto` retains existing behavior: stage policies plus run-level waiting/failure defaults. `on` also alerts terminal success, even with no stage notification hook. `off` suppresses stage, run-level system and browser tray notifications. Needs you remains a status/action surface, not a delivery channel. Summaries and pushed events carry the mode; the browser seeds and folds it. `FinishRun` and `MarkWaiting` enqueue the notification intent in the same SQLite transaction as the run/step state change. Enqueue failure rolls back the state; after commit, the worker can recover the intent without any runner callback. The store derives the mode, name, stage and sequence from stored rows. A failed FinishRun does not emit a terminal-success event or run alert. Terminal stage hooks run only after FinishRun commits. Waiting/terminal hooks and the worker share an atomic sending lease to avoid racing publications. Declared hooks get a one-minute initial opportunity to publish their artifacts; if a restart loses the callback, the generic outbox alert becomes eligible after that delay. `run_notifications` is an additive, idempotent SQLite migration. Each run/sequence/event/channel has one durable outbox row. Its status is `pending`, `sending`, `sent`, `failed`, or `covered` (a stage already published this event). The worker polls once per second, uses an atomic 60-second sending lease and allows three total attempts with 5/10-second retry delays. Pending and expired leases survive restart. A final expired sending lease becomes failed with an unknown-acknowledgement diagnostic. Existing retention deletes associated rows. `get_run` exposes run-level `notifications` receipts (sequence, event, channel, status, attempts, diagnostic, message_id), separate from step receipts. Successful stage publications cover the corresponding waiting/failure alert; a successful stage targeting `#success` also covers final success. Failed stage publication does not swallow the run-level fallback/retry. Stage artifact publications themselves are not made retryable by this change. Publisher acknowledgement is **not device reception**. A crash after publish but before the acknowledgement is saved can duplicate a notification. State and intent commit together, closing the transition-to-enqueue crash window; this is still bounded retry, not guaranteed delivery. Browser delivery remains page-session best effort and requires permission plus HTTPS/localhost; successful runs do not become Needs you entries. The older L5 parity description is attention-only for auto; on now adds success. Tests use temporary databases and fake HTTP publishers only, covering modes, no-hook success, stage dedupe, failed publisher retry/exhaustion, restart/lease recovery, audit persistence, and browser off/on. No live notifications, deployed database/configuration, service restart, binary installation or OS reception verification was part of this work. No new UI captures were taken.