---
url: https://docs.repoos.org/notifications.md
---
# Notices and notifications

The bell in the top bar collects things that happened without you watching:
release runs, and the outcome of every **Move to done** (close-out).

## The notice bell

Every row is one event, with a short title, a one-line detail, and a relative
time. Hover or focus a row for the exact local time; click it to jump to the
Releases page or open the task drawer. Dismiss a single row with **×**, or
**Mark all read** to clear the badge. Dismissed and read state is remembered
across reloads.

The bell also lists **Tasks needing you** below the notices — review
hand-offs, tasks flagged `needs input`, and branch drift (`needs merge`). A task
that is already `done` is never listed, even if it carries a stale flag left by
an earlier metadata-only state change.

### Provider, spend, preview, and remote-runner notices

| Notice | Meaning |
| --- | --- |
| Provider error | A recent agent run failed with a credit, auth, rate-limit, or model-unavailable error. |
| Silent agent run | An engineer is still running but has stopped producing output for longer than the stall threshold. |
| Spend alert | Provider-reported board spend reached your threshold (Settings → Notifications). |
| Awaiting visual check | A UI-area task is in `review` — open the preview and verify it in a browser. |
| Ran locally | Remote validation was enabled but a close-out used the full local gate instead. |

### Machine-readable feed

Scripts and agents can read the same items as **`GET /api/attention`**
(`ok`, `generatedAt`, `items[]` with stable `id`, `kind`, `severity`, `taskId`,
`message`, `detail`, `link`, and `at`). The server also emits an
**`attention.updated`** SSE event when the feed changes so clients can poll once
per signal instead of reassembling close-out outcomes and task flags by hand.

### Close-out notices

A Move to done that **finishes**, **fails**, or **times out** records a durable
outcome on the server, so the notice is there even if you closed the tab while
the run happened — and it shows the same time on every client. Timeout is kept
separate from failure because the advice differs:

| Notice | Meaning |
| --- | --- |
| Move to done: #id merged into main | The task's changes were merged into main. |
| Move to done: #id failed | A real gate failure; the task stays in `review`. |
| Move to done: #id timed out after Nm | The run passed `closeOut.timeoutMs`; a retryable failure. Retry, or raise the budget in Settings. |

A retry that later succeeds produces a new notice for the new run; it never
resurrects a notice you dismissed. **Stop MTD** produces no notice at all — a
cancel is not an outcome.

### Release notices

Release notes ready, release succeeded, and release failed appear in the same
feed and link to the Releases page.

## Notification settings

Settings → **Notifications** controls the bell sound and browser push. Both are
**off by default**, and every event type has its own toggle:

* Task attention — review ready, paused, stuck, needs attention.
* Releases — notes ready, release succeeded, release failed.
* Close-out — merged into main, failed, timed out.
* Provider and spend — provider/credit errors, silent runs, spend threshold,
  awaiting visual check, remote fallback to local.

A per-type toggle only suppresses the sound and push for that type; the bell
still shows the notice. Browser push needs the browser's permission and a
secure context (localhost or HTTPS on a real host) — see
[Troubleshooting](/troubleshooting) if it doesn't appear.