Offline banner (is it my network?)
A slim amber bar across the top of the app when your computer loses its internet connection, so a failing session or a stalled sync reads as your network rather than Omniscio. Always on with no setting, it clears itself on reconnect — and while it is up your sessions hold quietly instead of burning retries.
On this page
What it is
What it is — A thin banner across the top of the app that
A thin banner across the top of the app that appears when your computer loses its internet connection, so you know a failing session or a stalled sync is your network — not Omniscio. It is always on; there is no setting to configure.
This is the counterpart the Anthropic outage alerts monitor defers to: that monitor watches Anthropic's status page and never alerts about the status page being unreachable, because an unreachable status page just means your connection is down — which is the offline banner's job to show.
Where to find it
The top of the app window. The banner is automatic and always on — there is no setting, no menu and no control surface for it.
How it behaves
What you see
- When you go offline, a slim amber bar slides in at the top of the window with a Wi-Fi-off icon and the message "You are offline. Some features are unavailable."
- The bar clears itself the moment your connection returns.
- A brief blip does not flash the banner: it appears only after you have been offline continuously for ~5 seconds, so a one-second drop while switching networks never shows it.
How it works (under the hood)
- A main-process connectivity service (connectivity-service.ts) polls Electron's
net.isOnline()on a self-rescheduling timer — every 10 s while online, tightened to every 2 s while offline so a restored network is detected within ~2 s rather than lagging a full interval. - On a transition it emits a single
NETWORK_STATUS_CHANGEDpush ({ online }); the renderer'suseAppNetworkStatushook updates the connectivity store, anduseOfflineBannerGateapplies the ~5 s debounce before the banner inAppBannersrenders. - The poll loop is fully fail-safe: a throw in a listener or in
net.isOnline()can never stop the reschedule, so the offline/online signal — and every isOnline()-gated poller downstream — keeps recovering.
While you're offline, your sessions wait instead of burning retries
The same connectivity signal also pauses the automatic session-retry engines while you're offline, so a dropped connection no longer makes Omniscio hammer the Claude API with retries that can't possibly succeed:
- A session that hits an API error while offline holds quietly — it stays green (not "Needs you") with a one-line "No internet connection — will retry automatically when you're back online. No action needed." note — instead of burning through its ~5 silent paid retries, and it sends a single "continue" the moment you reconnect.
- The background stuck-session recovery sweep stops re-driving failed sessions while offline: no wasted respawns, and nothing is pushed toward "given up" for a transient outage. It resumes on its own within ~90 s of your connection returning.
- Both read the same online/offline verdict above —
isOnline(), which folds Electron's link signal together with a shared reachability probe (viashouldDeferRetryWhileOffline()in connectivity-service.ts) — and revert instantly withAMC_DISABLE_OFFLINE_RETRY_GATE=1. This covers a genuine loss of internet, and also a link that is up while name resolution is dead; a link that is up and resolving but cannot reach the API (an Anthropic outage) is still handled by the normal bounded retries + the Anthropic outage alerts.
When the outage lasts, Omniscio tells you, tries your Wi-Fi, and picks up afterwards
A blip is only the banner. An outage that lasts is more: the internet watchdog (internet-watchdog.ts) steps in.
- It waits before it calls it an outage. The connection has to stay gone for about two minutes, and a known non-AI website (Windows' own connection-test page) has to fail too. A short drop, or the AI provider alone being down, declares nothing.
- It tells you once. One inbox card, "This computer is offline", says since when and what it tried. It goes to your phone through the same push path as other urgent cards. A computer with no internet cannot reach a phone, so the page only arrives if some route out still exists. You can dismiss it like any card.
- It tries your saved Wi-Fi (Windows only). If a Wi-Fi network you saved as Connect automatically is not connected, Omniscio asks Windows to join it. It tries once when the outage starts, then after 1, 2, 4, 8 and 15 minutes, so 6 times at most. Each attempt is logged. It never changes a setting, never turns an adapter on and never touches a network you did not set to connect automatically. The card never names the network.
- It updates the same card when you are back. After the connection has held for about 90 seconds, the card changes to "This computer is back online" with how long it was out. A card you already dismissed stays dismissed.
- It resumes what the outage interrupted, once each. A scheduled wake that never reached its session during the outage fires once, however many times it missed. A session whose last turn died on a network error, and that nothing has touched since, is asked to continue once. A session that stopped for any other reason, is mid-turn, is archived or is paused is left alone.
- Kill switch
AMC_DISABLE_INTERNET_WATCHDOG=1(on by default). The quiet session hold above keeps working either way. - Known limit: the outage is remembered in memory, so an app restart during an outage starts a new
one. The Wi-Fi step reads Windows' English
netshoutput; on another display language it finds no network and only logs that.
When your Claude Code CLI goes missing, the same engines wait
The one condition that looks like an outage but is not about the network: your computer has no working Claude Code CLI — usually because an upgrade is mid-install. Nothing can spawn while it lasts, so every automatic retry is doomed in exactly the same way. Omniscio now holds that fact once for the whole machine, instead of letting each session conclude on its own that it is broken:
- The background stuck-session recovery sweep stands down completely while the CLI is missing — no respawns are spent, and nothing is pushed toward "given up" for a condition that is not the session's fault.
- A scheduled wake whose delivery is refused because of it is recorded as skipped, not a permanent failure, and stays armed — so a heartbeat is not torn down by an upgrade that takes a few minutes.
- It clears itself: the app re-checks about once a minute, and the sweep resumes on its own the moment a working CLI is back — no restart and no nudge. A Retry or re-arm you ask for yourself is never gated.
- It is deliberately hard to trip: a merely busy machine is never mistaken for a missing CLI, because
the check reads the files on disk, which do not care how loaded the box is. Kill switch
AMC_DISABLE_CLI_OUTAGE_GATE=1.
Related
Related — Anthropic outage alerts
- Anthropic outage alerts — the sibling monitor that watches Anthropic's status page and hands the "is it my network?" question to this banner.
Last verified 2026-10-22