[nofuturekid]

What changed

A "Needs attention" card at the top of the Overview lists what needs the user's attention, most urgent first. One row per item, coloured with the existing Warn / Danger / Info tones and also labelled in words (Critical / Warning / Info), so it does not rely on colour alone. Tapping a row opens the Array tab, or the notifications sheet for notification rows. The card is hidden when the list is empty and while the server is offline (an Error on the info, array, docker or VMs stream, the same rule as the top bar's Offline pill).

What it flags:

• Array not started (warning), degraded (critical), or rebuilding while the API reports no parity operation (warning).
• A parity operation: running (info), paused (warning), last check failed (critical), last check completed with errors (critical).
• Disks with numErrors above zero (critical), one row per disk.
• Disks at or above their critical (critical) or warning (warning) temperature, using the per-disk threshold, else the global display threshold.
• Unread alert (critical) and warning (warning) notifications, as counts.

Architecture

The rules are one pure function in the data layer, attentionItems(AttentionSnapshot): List<AttentionItem> in data/model/Attention.kt, with no Compose or Android dependency, so background alerts (#321) and a widget (#322) can reuse it. Each AttentionItem has a severity, a title, a detail, a target (Array or Notifications) and a stable key that names the condition and not the moment: progress, error counts and temperatures stay out of the key, while warm to critical is deliberately a new key.

Only data the app already loads is used: the array stream, the notifications stream and the display thresholds. No new polling and no new query.

Decisions

• Unknown is not a problem. An array that is not loaded, a spun-down disk (isSpinning is the truth source), a missing threshold, ArrayState.Offline (the mapper's "unknown") and an unreported parity error count raise nothing.
• Temperature needs a threshold the server reported (per disk or global, each coalesced independently, values of 0 ignored). The Array tab also falls back to fixed 42/50 degrees; this card does not, because a guessed limit would raise alerts the server never configured. This differs from the Array tab only when the server reports no threshold at all.
• Notifications are a sheet, not a tab, so the target is AttentionTarget.Notifications; the Overview opens the sheet.
• A cancelled parity check never raises an item.
• Small mapper extension: ArrayInfo.parityLastRun (outcome and the error count if the API gives one), filled from array.parityCheckStatus, which the existing GetArray query already asks for. ParityCheck and the existing parity mapping are unchanged. Neither queries.graphql nor UnraidRepository.kt is touched.

Deliberately left out

• Containers that are not running but have auto-start on. The flag stays set when a user stops a container on purpose and one-shot containers exit by design, so the loaded data cannot tell "should be running" from "was stopped". When in doubt it stays out.
• A parity error count as a source of truth: the API does not fill errors today, so "completed with errors" can only appear once a server reports it; a FAILED status is flagged now.
• Notifications outside the app, a widget, background alerts, version bump, .github/.

Checks

• ./scripts/local-ci.sh: passed (generateApolloSources, lintDirectDebug, assembleDirectDebug, assembleRelease, testDirectDebugUnitTest).
• New AttentionItemsTest (32 tests: each rule raised and not raised on unknown data, stable keys, ordering, healthy server empty) and four new mapper tests in ArrayMappingTest; 0 failures, 0 skipped.
• Not checked on a device: the card's look and the tap behaviour (maintainer acceptance).

Closes #319

🤖 Generated with Claude Code