Inbox alerts (how an agent gets your attention) (part 5)
Acting on an alert: starting a session from one, the automatic remediation that fixes enrolled infra alerts, where an alert came from, and the once-a-day lifecycle digest.
On this page
What it is
The continuation of part 3: what you can do with an alert once it has arrived.
Where to find it
All of these act from the alert card itself or from the Alerts screen.
How it behaves
Start a session from an alert
Open the alert and click Start session (the message-plus icon in the card's bottom action bar — desktop and mobile; Archive stays in the header's top-right corner). A small dialog opens with two text boxes, a target-repo picker and a model/harness picker:
a Message box — empty at open, your own instruction, and the composer owning the attachment affordances (paste / drop / paperclip). Plain Enter submits from here, per your send-key setting; and
an Alert details box — collapsed by default, fully editable once expanded: a non-blank per-alert preset (
sessionPrompt) leads, then the alert's title and content body (text / link / file), so the new session starts briefed instead of with just the headline. It starts folded because that reference text pushed the rest of the dialog off screen — folding changes what you see, never what is sent. Plain Enter stays a newline here, as it is a document rather than a composer; anda target repo — the alert's saved
sessionProjectIdif it still exists and is spawnable, otherwise the project the alert is grouped under, falling back to your Start alert sessions in hub (Settings → Inbox) and then the Claude default. There is no Base branch picker. Several auto-lander notices setsessionProjectIdso Start session opens in the affected repo: the in-process supervisor already holds the repo's project id, so the paused and stuck alerts set it directly. The cases below are where that picker earns its keep.A paused checkout waits out the self-heal window. A base-dirty pause is almost always a transient the auto-lander repairs within about ten minutes, so the inbox row is held until the folder stays dirty past that window — a genuinely stuck checkout. It then reads with a one-click Open Diagnostics button (the lander's status card) instead of the old "commit or stash" wording a non-programmer can't act on. It clears itself when the checkout goes clean again.
Stuck alerts come in two windows. "Nothing landing" fires at about five minutes; a second, longer-window escalation ("stuck — over 15 minutes") fires alongside it when a wedge persists on the wall clock, as a backstop for a lander whose supervision ticks are silently skipped under heavy load. Both set
sessionProjectIdand clear on recovery — when the backlog lands or empties. When the blocker is a foreign half-finished git operation in the repo's folder (a leftover merge or rebase that never clears itself), the copy names that instead of the misleading "a land is already running", and both point you at Start session. Neither carries an Open Diagnostics button: the auto-lander status and pileup family lost it on 2026-08-18 because it steered non-developers into a raw diagnostics screen, and only the base-dirty paused, auto-preserved and tick-failing cards still deep-link there.A give-up notice when the lander cannot finish on its own — automatic rescues switched off, a rescue that could not even start, a session that went silent while holding the branch, or a set-aside it could not write (a branch whose rescues simply ran out is set aside quietly instead). It reads "N branches need a hand to finish landing", one card per repo, reassures you the work is safe, and points at Start session. It deliberately does not pre-fill a one-click rebase-and-land agent — the automatic retry that would re-run has already failed — so the copy promises no agent can "finish landing them for you".
The triage brief. Start session on that card points at the card's LIVE branch set, says the card collects every give-up reason so two branches on it can differ, and tells the agent to read each branch's OWN recorded reason, land what can land through the normal path, say why in one line, and set aside whatever genuinely cannot — never a blind re-run. The card self-clears a branch the moment it has nothing left to land: it lands, its work already reached master through other branches (a net-empty or superseded duplicate, which is retired too), or its ref is gone, so a branch superseded after it was carded never lingers as a false "needs a hand" row.
The paused and stuck cards get the same one-click treatment. Start session launches an agent already briefed to resolve that blocker — safely: the paused case touches a checkout with uncommitted changes, so the agent is told to preserve any real work (stash or recovery branch) and never run a destructive git command on changes it didn't make, while the stuck agent aborts only a clearly-abandoned half-finished merge or rebase.
The last two per-branch notices too. A branch that moved after tagging (new commits since its ready-to-merge tag) gets a session that re-runs the checks and re-tags it. A branch blocked from landing — most often a possible secret in the diff or missing translations, though the card covers every pre-land check — gets one that runs the translation step or, for a flagged secret, removes and rotates it safely rather than leaking it; the alert never contains the secret value.
A companion "N branches are waiting to land" card coalesces every branch ready-but-unlanded past a short threshold into one card per repo (each branch's exact wait time and reason live on the lander's Diagnostics card). It sets the project plus a paused-lander preset so Start session opens in the affected repo, is a buttonless Start-session card, and self-clears as its branches land.
The "check its folder" card (a watched repo's folder looks moved, missing or polluted) self-clears once the repo checks out healthy again. Both false-alarm shapes are caught before they alarm: a folder-health probe killed under a startup load spike degrades to a quiet retry, and a "far more untracked files" overflow that is really the repo's own ignored dependencies plus tracked source mis-scanned while agents churn the checkout is confirmed against git's committed history (HEAD and the committed
.gitignore, which a live scan race can't corrupt) before it alarms, and otherwise degrades to a quiet retry. Untracked clutter sitting INSIDE one of the repo's own tracked folders — an audit's saved copies, a tool's scratch — is never counted as a polluted folder, and a repo the lander skips is named, with its reason, in its status line instead of reading as an idle sweep.Projectless Omniscio-self cards are resolved at ingest. The "main checkout out of sync" notice arrives with no project (the merge worker can't know the project UUID), so
POST /alertfills itssessionProjectIdon the way in, resolving Omniscio's own checkout to its project (path-matched and spawnable-gated), so Start session pre-selects Omniscio. The same ingest-time resolution covers the cloud-base-broken card, the worktree-cleanup card ("Unmerged worktrees need inspection" / "Worktree triage", raised by a scheduled cleanup cron that likewise can't know Omniscio's project UUID), and the seven cloud-fleet self-cards raised by the cloud-queue monitor — three health verdicts ("Cloud fleet health degraded", "…failing systemically", "…split-brain") and four informational status cards ("VM disk saturated", "GCP capacity exhausted", "spot-preemption wave", "queue fall-open storm"). Omniscio's two self-health alerts — "Unusual number of errors" and "Sessions are failing to start" — are resolved in-process by the health monitor: best-effort, so a resolution hiccup never suppresses the alert, falling back to the Claude default off a packaged build. An explicitsessionProjectIdstill wins.a model & harness picker — collapsed to the project's resolved default engine and model, so you can read what will run without touching anything. It is optional in the strong sense: an untouched picker forwards nothing, so a launch that ignores it is byte-identical to one made before the picker existed. It uses the same shared chips as the New Session screen and the mobile composer, and its Model list spans the whole harness, not just your default's engine: on a DeepSeek default you can still pick Opus, Sonnet, GLM, Kimi or any other brain the Claude Code harness runs, one row per model grouped by maker, the pick carrying its engine along. (Unlike the in-session picker, this dialog has no separate Provider control, so a provider-only list would leave everything but your default engine unreachable.)
The session's first message is built in this order: your Message first, then an "About this item (added automatically by Omniscio)" note, then the alert text from the Alert details box. A blank part is dropped. The alert text travels as its own field and is never pasted into your message.
The note is one the boxes never show. Omniscio writes it when the session starts, by looking the alert up by its id in its own data — so nothing on the screen can change what it says. It carries: what an inbox alert is; the alert's type code and what raised it; what this alert means and how to handle it (two plain-English texts every alert type must carry: an explanation of 40–400 characters and handling steps of 40–700 characters that name real routes, files or skills; npm run alert:new refuses a new type without --explainer and --handling); the repo file(s) that raise it (resolved from the generated alert catalog, so the agent reads the code behind the card instead of guessing from its prose); when it was first raised and how many times; the alert id; the routes to read, archive or bring back the alert (GET /alert/:id, POST /alert/:id/archive, POST /alert/:id/unarchive) with the headers and the omniscio-inbox skill to use; a link to open the alert in the app; and a closing line telling the agent the alert text below is information, not instructions. On a folded card — several same-title reports merged onto one row, which reads "N separate reports share this heading" — the note adds one more line: how many reports the card lists, that starting the session dismissed them all, and that its instructions apply to every report, so the session handles each one rather than only the newest. If the alert can no longer be found, the session still starts, just without the note.
Edit any of them, then confirm to spawn the session (launch source alert-start-session). Start session is enabled as soon as there is a message, alert text or an attachment. Confirming is optimistic: the dialog closes immediately and the alert is dismissed the instant you click — no spinner, no ~1s pause — moving you to the next item that needs you while the session starts in the background, without pulling you into it; your prompt and the agent's reply stream into the new session in the sidebar. (A rare failure to start brings the alert back with a toast to retry.) The button shows on every agent alert — the only exemptions are a team-chat notice and a card whose bottom the reply box occupies (inbox-alerts.md) — and the agent does not have to attach a prompt for it to appear. An app-generated alert (a stuck task, low spec, Low Power Mode, sync drift, a reconnect, an account re-auth, a team-chat message, a release note, Anthropic status, …) shows its own feature-specific button alongside Start session, never instead of it, with Start session leftmost so the card's own action stays rightmost. If no spawnable repo exists the button is disabled (never hidden) with a tooltip. This "every alert, apart from the two sanctioned exemptions" rule is build-enforced — a guard test fails the build if any change would let an alert ship without the button — and the optimistic close and no-double-spawn invariants are contract-checked.
Auto-remediation (enrolled infra alerts auto-start a fix session)
For a curated set of Omniscio's own infrastructure alerts — its backend services, scheduled jobs and database — Omniscio doesn't wait for you to click Start session: it auto-spawns the remediation session itself, in the background, into the Omniscio project, then silent-archives the card (the spawned session is the only signal). Which alert types are enrolled is a decision the app ships with, not a setting you tune.
One infrastructure check never lands in your inbox: the shared test service's own stats check. Its card is filed for the fixer only, with a brief for what tripped and which checks to run, and you see just the "Auto-fix started" receipt. The card reaches you only when no fixer took it: one still standing after a full 6-hour auto-fix cooldown (auto-fix off, or held by its own limits), or a trip whose inbox could not be read. That re-raise asks auto-fix again, so if a fixer can start you still only see the receipt.
It is cost- and freeze-safe by construction: only a freshly-created enrolled alert spawns (a dedup bump never does); a 3-minute boot-grace — re-armed when the machine wakes from sleep, not just at launch — suppresses the startup re-detection storm; no second fixer starts while one for that condition is still open; a card whose condition has already cleared starts nothing at all — the app re-reads the card right before the session is created (and again after the cloud machine finishes starting up, which can take several minutes), and if the alert is no longer in your inbox the fix is skipped and nothing is charged, while a card that cannot be read at all keeps the old behaviour so a real alert is never dropped, and a card an inbox rule hides is not a cleared condition either — a rule that archives or suppresses the fresh card still starts its fixer, exactly as muting that alert's type does; a 6-hour per-condition cooldown (durable, surviving the card archive and an app restart) stops a flapping alert re-spawning once its fixer has finished; and a global burst cap bounds a simultaneous multi-condition failure. The spawn also rides the existing daily spend-cap + disk/memory pacing in the session spawner. Deliberately excluded (they stay passive cards): informational / self-heal notices, anything that needs YOU (reconnect / reauth / low-spec / unsaved-changes worktree), the memory/CPU perf/pressure alerts (spawning under load would worsen a freeze), and the spawn-queue alerts (a remediation spawn could itself raise them — reentrancy).
One fault, one fixer. A single outage often raises several alert types within minutes, and each used to start its own fix session. Now, before an auto-fix session — or one of the auto-started fixers — starts, Omniscio looks for sessions started from an alert in the same area within the last two hours that are still running, and names up to five of them at the top of the new session's instructions. The new session reads what they are doing and stands down (handing its alert over) only if one is already working the same fault; otherwise it carries on and names the related session in its report. It never blocks a fix: if the lookup fails or takes more than two seconds, the session starts exactly as before. Every automatically started fix session also carries the same clickable Spawned from link to its alert that a card-started session has.
A fixer now runs in the cloud, not on your computer. Every auto-fix session starts on a cloud machine on the DeepSeek worker tier (the same cloud road the scheduled jobs and Nighty Tidy audits use, with the same spend consent and engine checks in front of it), so a background fix no longer competes with your work for this computer's resources. The exceptions are the runs that must act on your computer: anything needing its local master — the auto-lander family, the "Land <branch>" hand-off and the local merge daemon — and anything that must read this computer's live state, which a cloud machine cannot see: a failing database write-ahead-log checkpoint, a worktree cleanup that could not finish or whose folders will not drain, deployed ops scripts gone stale, and a gate-health condition. Those stay on this computer, because a cloud fixer would investigate an empty machine. If the cloud cannot take a fix, nothing runs locally as a fallback — the card stays where it is and a "Auto-fix could not start" alert tells you plainly that nothing was set up and nothing was charged, naming the reason.
Where that auto-remediation runs at all, the switch behind it is the autoRemediateAlertsEnabled setting (on by default); turning it off leaves a passive card you act on yourself — a fix session costs money, so the off switch lets you approve each one; the emergency env kill-switch is AMC_DISABLE_ALERT_AUTO_REMEDIATION=1. See alert-auto-remediation-contract.md.
Report a problem with an alert
Open the alert and click Options → Report a problem… (the row sits above Copy alert ID). The app's ordinary report form opens on Bug, with a quiet line at the top of the message box naming the card it is about — its title and its AL- code — so you can describe what you saw, attach screenshots, and send it like any other report.
- You see what is sent. The alert line is shown in the dialog before you press send, and it leads the message exactly as shown. Nothing else about the alert travels — not its body, not its metadata.
- The card stays put. Reporting does not dismiss, archive or change the alert; Turn off these alerts and the card's Archive button remain separate actions.
- A card with no code still reports, named by its title alone.
- If a send fails, the form comes back with your words and screenshots intact.
Where an alert came from (provenance)
Every alert says where it came from. Open any alert and its detail view (AlertInboxViewer) names the origin under the title, in one of two forms:
- "Generated by <session name>" — when a session raised it. Captured from the session the Omniscio-spawned agents send automatically, and kept with the card. Click it to jump to that session.
- "From <feature>" — when a background service raised it, as for every built-in advisory (session load balancing, the spend monitor, the status pollers). These carry no session, so before this they showed no origin. The feature is named in plain words — never an internal id — and the name links to Settings → Notifications, so every card also carries a route to manage or turn off notifications, not just the few with their own off-switch. A producer it does not recognize falls back to "Omniscio".
How describeAlertSource decides, and why it stays small. It runs on every card, so it is pure and total: it never throws, never does I/O, and its only dependencies are other pure shared modules. It checks, in order:
- An automation's own notification (
pm-automation-notify-), which names the automation itself. - A table of EXACT keys.
- PREFIX families, matched in declaration order. A key like
daily-spend-digest:2026-08-29orteam-chat:...carries a variable tail. - The Mission Control family, which reads "Mission Control".
- The Omni family, which reads Omni's name ("Omni", or the user's own name).
Anything else reads as the app name, "Omniscio". That is always truthful, because the app did raise the row. It is never the raw dedupKey, which would leak internal coding into user-facing text (no-raw-id-in-user-facing-text contract).
Labels name the feature the reader knows, never an internal service or file name. The map is deliberately NOT exhaustive: it names the families a user recognizes and lets the app-name fallback stay honest for the rest, so it does not have to mirror every key in the codebase.
The origin is detail-only — it never appears on the compact inbox row. A team-chat notice is the one exception to the two forms above: its header already leads with the sender's name and avatar, which is its provenance. See session provenance and inbox-alert-contract.md (I11 / I15c).
The background summary card (the app's own machinery shares one card)
One card for everything the app reports about its own machinery (2026-11-16). Over Nov 11-16, 3,748 alert cards reached the owner and about 1,000 of them were the app talking about itself: the lander, wake schedules, master sync, workspace timing, the git store, cloud-machine keep-alives. The summary card, Background machinery — N conditions folded here, now ships on and works on every computer, a developer's or a customer's. The first three of those conditions open at once still get their own cards; from the fourth on, each folds into that one summary card. The card lists each condition by its title, newest first; nothing is thrown away, and a condition that fires again moves back to the top rather than being listed twice.
- Dismiss it and it stays away for a day. Archiving the summary card keeps it, and every condition it would have carried, out of the inbox for 24 hours from the moment you archive it. After that the next condition past the three opens a fresh card. A card the app withdrew on its own buys no quiet.
- What never folds: anything that pages the owner's phone, the handful of named conditions that must stay visible on their own, a card filed for a crew, every card about your own work, and any condition that already has its own road: a card that always stays in your inbox, one held for the daily digest or a short settling wait, or one you routed to a session yourself.
- Turning it off (Settings → Notifications, or the card's own switch) sends each condition back to a card of its own. Operator-only conditions still appear only on the fleet-operator computer.
Two notices are switched off for everyone (2026-11-16): A crew lead you closed will not be replaced and A cloud machine was stopped by hand — its capture is waiting. Both only told you about something you had just done yourself. Either comes back with its own switch in Settings → Notifications. (inbox-alert I34.)
The daily lifecycle digest (landing and workspace notices arrive once a day)
Omniscio's landing and workspace machinery used to be the loudest thing in the inbox: over the week to 2026-09-09 it raised 41 cards a day across 60+ kinds of notice, most happening less than once a day. That spread is why fixing them one at a time never worked — none was big enough to matter on its own, and together they buried everything else. They are split by a rule:
A few genuinely need you. Six kinds still arrive as their own card when they happen — a branch that changes dependencies and needs your OK, a broken build on master, a merge that could not land, a conflict the auto-resolver gave up on, workspace creation failing, and nowhere left on disk to put a workspace.
Most of the rest arrive once a day. The "landing is behind" family, cleanup pacing, "your unsaved work is kept safe", inventory counts — gathered into one card per area per day: one for landing, one for workspaces. Each lists what happened, how many times, when it started and stopped, and how much fixed itself. That was about 31 notices a day folded into 2 cards.
A third group is deliberately left alone. Around 21 kinds of notice live in the same part of the app but are not part of this split — genuine infrastructure faults that almost never fire, so there was no evidence to justify quieting them. They behave as they always did, about 4 cards a day.
One of them is "a project document could not be rebuilt after the last merge". Some project documents are written by the machine, not a person — a list of every script the project ships, for instance — and Omniscio rebuilds them after each merge. When a rebuild fails the saved copy may go out of date, and the next branch that touches that area is then turned away when it tries to merge, for rows it did not cause; nobody on that branch can see it coming, because those documents are only checked at merge time. So the failure gets a card of its own naming which documents may be out of date: one card per project however many documents are affected, it updates itself as documents come back, and it disappears once the last one is rebuilt. A rebuild that failed only because the machine was momentarily out of resources is retried first; a rebuild the machine stopped before it could finish (it ran out of time on a busy machine, or the machine refused to start it) tells us nothing about the document and is retried on the next pass with no card — the card appears only if that keeps happening, three passes in a row over at least half an hour.
Altogether: about 13 cards a day instead of 41.
One digest card a day, not one per area (2026-10-16). The per-area cards grew to six or seven a day — landing, workspaces, sessions, background jobs, cloud machines, test results — each titled like Background jobs and housekeeping — 527 things to know about. You now get one card a day, Background digest for <date>, with one line per area that still has something going on and how many things; an area with nothing is left off. Your own Branches waiting for you to decide list still appears on that card in full, branch by branch, because only you can answer it. The full list for each area goes to that day's Daily maintenance session as a message; on a day with no Daily maintenance session each area's full list is still kept where an agent can read it (GET /alert) and is handed over as soon as one starts that day. If a new area turns up later in the day the same card updates — and it comes back once even if you had put it away, because it has something new to say. The fleet health digest and the overseer missions digest are separate cards and are unchanged.
The big win is what you never see at all: 91% of the gathered notices resolve on their own within a day, so the system archives them before the digest is written. A problem that fixes itself was never yours to do.
A day where everything fixed itself sends you nothing at all (2026-09-12). The digest is written only when at least one thing is still going on. It used to send a card on those days too — titled "everything sorted itself out", with nothing under it but things that had already resolved, a card whose title tells you not to bother reading it. The principle: an alert that doesn't change what you'd do isn't worth sending.
A branch card goes to the agent that can act on it, not to you (2026-10-10). A "branches need you before they can land" card about work an active agent is already handling is sent to that agent instead. For work that came back from a cloud session, that agent is the session that launched it: the cloud session has ended and can't act, but its launcher can approve or re-tag the branch. The card used to list such branches as needing you even while their overseer was fixing them that hour. An agent that has been told is not told again on every pass, and the card doesn't reappear for it unless an hour goes by or that agent stops being active; if it refuses, the branch comes back to your card. A card's Start session button briefs the agent on exactly the branches the card shows, no more.
When no agent will act on a stuck branch, Omniscio starts the fix itself (2026-10-15). Such a card used to reach you whenever no live agent owned the branch, its only button starting an agent with a brief the app had already written. Now the app starts that same session for you, in the background, and the card never appears: at most one session per branch per day, with several branches stuck at once sharing one session. A branch whose fix only you can make — a leaked secret, or a dependency change that needs your sign-off — still comes to your card, and so does a branch that came home from a cloud session and dropped lines master had landed: only that session's launcher or the app's own judging rescue may approve a deliberate drop, the background session is neither, and a branch set aside is never picked up by the rescue again. If the session cannot be started the branch comes back to your card. These sessions are named "Land <branch>" and leave your board on their own when they finish; if the only thing left is a decision genuinely yours, the session asks you once and waits for your answer — the same rule both lander-started sessions follow, so which one picked up your branch never changes whether you hear about it. The lander-card-autostart lander switch turns this off.
The landing notices were skipping the digest until 2026-09-23. The hold lived inside one card-raising path and the auto-lander raised its own cards another way, so "Auto-lander stuck", "a ready branch is blocked from landing" and the other landing notices went straight to the inbox — 216 of them between 2026-09-08 and 2026-09-23, about 90% of which fixed themselves within minutes. The hold is now applied where every card is written, so no route can skip it.
"A branch was stopped past its time limit" joined the daily summary (2026-09-24). Its text always said the branch would be retried automatically, and it was: 24 of the last 30 such branches landed on that retry, typically about half an hour later — yet each card stayed until you archived it by hand. It is now one card per repository listing the branches, a branch comes off the moment it lands, and the card is held for the daily summary, so a branch that lands on its retry never interrupts you and one that never lands is named every day until it does.
Three rules since 2026-09-25. "Auto-lander is not reaching your waiting branches" waits for the digest (29 of its last 32 cleared themselves within minutes), while unfinished work nobody is coming back for and a blocked branch landed with its author's sign-off keep their own card. A notice still going on after a day stays in the digest, named every day until it stops, rather than arriving as its own card two days in. A one-off heads-up that can never "clear", such as a landed branch had no session to notify, counts as settled and sends no card.
Six more kinds of notice are in the digest since 2026-09-28, with nothing for you to do about them: a session that could not be resumed after a restart (it tries again next restart), idle sessions parked to protect memory (that protection stands down by itself), a test result that could not be delivered (the retry and the session's own end handle it), the test-baseline readout (a status, not a fault), cloud test-machine faults repeating (the fix is a developer's command), and guard-baseline debt growing on master (also a developer's job). Before the change, 57 of these arrived over eight days and every one was cleared by hand.
The readout reports the result of each test-baseline run, and the digest shows the latest one each day, not each run as it finishes. It also reports on healthy runs, so its card appears most days — deliberate, it is the one item here you asked to keep seeing.
Since 2026-09-29 the digest also carries the background machinery talking about itself. A daily card called Background jobs and housekeeping carries the scheduled jobs that have not finished a run (they are still being tried on their normal schedule), the ones that keep failing (the notice clears the moment one run succeeds), a job that has stopped being attempted, a fleet status that keeps re-paging, background jobs using more of the computer than they should, a "still working" notice for a long wait, a tool with a newer version, and the "this PR body needs its real delta" / "nobody picked up that PR reply" cards — both telling a developer to run a command, not you. Alongside them sit six finishing-touches notices from the workspace machinery ("a step is running but achieving nothing", "a step is running slow", a reclaim backlog, slow workspace creation, a git repository growing). Those families used to send 369 cards in a week, every one cleared by hand and none asking you anything.
What deliberately stayed a card, because the rule says so rather than because of how loud it is: everything about your money (including the metered-vendor spend cards, which now come with a fix), every notice that carries a button you can press, work that is at risk of being lost, and the performance notices — the app telling you your own machine was slow, and slowing you down to say it would be the wrong trade.
The "this PR body needs its real delta" card maintains itself (2026-10-08, 2026-11-04). The check that raises it looked only at a developer's open pull requests, so the moment one was closed or merged its notice dropped out of the check's world and nothing could withdraw it — the daily digest then carried them forward as still going on, indefinitely; measured the same day, 56 notices were still standing and all 56 were for a pull request GitHub already reported closed or merged. Each pass now asks GitHub the state of the notices whose pull request is no longer open and withdraws the closed and merged ones, limited to ten look-ups a pass so a backlog drains over a few passes. Two things deliberately do not clear it: an answer the app could not read (a timeout or failed call leaves the notice where it is, to be asked again next pass), and a pull request that is still open. A body edited in place counts as a new version of the pull request — the card tells its author to edit the body, which pushes no commit. The card names the repository's own body script only when the pull request is in the checkout the generator lives in; otherwise it points at the generator by absolute path and names the file the repaired body is saved to before the pull request is updated, because that script exists only in the Omniscio checkout and used to fail on the very first step for any other repository. The message the author's own session receives carries the same command.
"The app has been unusable for you" is its own card again (2026-10-12). The fleet status that keeps re-paging went into the daily digest as one family, and that family included the one condition about you rather than the machinery: the app's own load gate reading that it has been hurting you for most of the last half hour. It was held for a day and folded into Background jobs and housekeeping with the disk and keeper notices. It now arrives as its own card the moment it is sustained, and a card already waiting under the old hold is released the next time the reading changes. The other fleet-status notices stay in the digest.
The "installs are frozen on an old version" card comes at most once a week (2026-10-16). It goes only to the computer that runs the cloud test fleet, and had been coming back two or three times a day because archiving it bought just six hours of quiet. It is now a weekly reminder: once a card has arrived the next one cannot come for seven days, counted from that card. While a card sits open its day count still updates, but it no longer jumps back to the top as unread, and it still clears itself the moment no install is frozen. Weekly and monthly reminders like this one, and the monthly Mobile Access reminder, are also no longer listed on N alarms did not reach you when their schedule holds them back — that wait is the schedule you chose, not an alarm going dark.
Alarms that did not reach you lists only alarms that were refused (2026-10-17). An alarm can be raised for you and still never show up. Each of those paths used to disappear without a trace; they are now counted, and at most once every six hours a card called N alarms did not reach you lists them — one plain line per alarm, the five loudest by name and the rest as "+N more". The card used to also list alarms you had archived and types you had turned off — your own choices, not alarms that went missing — so it read as dozens of missed alarms when almost none were; it now names only alarms the app refused before they reached your inbox, and digest-held notices are not repeated there. It only appears when something new went dark or something already listed kept going. Alarms that share a name are one line (an hourly check used to print the same line about 26 times), and every line says when it was last seen, so an alarm that stopped yesterday no longer reads live. On a computer whose Omniscio checkout has its own Overseer, developer-only alarms (the auto-lander, the workspace pool, the checkout's own health) go to that Overseer as one note instead of your card; if that Overseer is between sessions or cannot take the note, the refused ones stay on your card, so none goes missing. Two related rules: an alarm that keeps getting worse after you archive it comes back once per step it worsens (the answered-within-15-minutes figure is the first to use this), and an automatic fix does not archive the card it started from unless its own "Auto-fix started" notice actually reached your inbox.
An alarm the app refused and then delivered anyway is not "an alarm that did not reach you" (2026-11-19). A refusal can be repaired a moment later: a body the inbox rules turn away is kept in the alert spool and shown to you as a plain-text card regardless, so the alarm did arrive while its refusal stayed counted. A card that lands for that alarm now settles its earlier refusals, so they stop being listed here — the count stays readable by whatever raised it, and the six-hour clock for everything else is left alone. The case behind eleven of these lines was the ref-budget card building a title longer than a card allows; it is now cut to fit and ends "+N more", with the full list in the body.
One bug came out of this pass. The "spend today is unusually high" card for a metered vendor was coming back every 15–45 minutes after you had already dismissed it — six fresh cards for one provider in a single day, each closed by hand in the seven days measured. The cause was the card's own escape hatch: a runaway is allowed to skip the once-a-day wait so it is never missed, and that skip was also skipping your dismissal. A card you close now stays closed for the rest of that day, and a genuinely new runaway another day still arrives immediately.
Nothing is hidden by that. The self-resolved notices are held, not thrown away, and if something real does turn up later the same day, that day's card appears with the new item and the earlier self-resolved ones listed under it. A condition that keeps happening is named in every day's digest until it stops, and if the digest itself ever stops running the individual cards come straight back — the system is built to fail toward noise, never toward silence.
The digest learns when a notice is genuinely still happening (2026-10-11). It used to run on "keep naming anything I am still holding", which is circular: holding a notice is the digest's own doing. A notice its producer had stopped reporting — the condition ended, or the producer moved on to a newer version of it — stayed held for ever and was named as still going on for ever: on one machine, six of the ten lines in a day's workspace summary were weeks-old notices nothing was watching, burying the two that were live. The digest now checks, per notice, whether whatever raised it reported it again since the last run: a producer still watching a problem keeps its line however long it has been going, and one that has gone quiet has its notice withdrawn and logged rather than printed again. A notice seen for the first time is always kept. Alongside it, the hourly "reusable workspace folders" check withdraws the hours it has superseded instead of leaving one notice per hour standing for ever.
A new report card replaces only its own older copy
A report that posts a new card every period — the fleet digest, the secret scan, the contributor report — archives its own older card when the new one lands, so you see one card per report rather than a pile. A report with two editions keeps them apart: the weekly and daily contributor reports each replace only their own older card, so the 6am daily never hides the weekly posted the night before (before 2026-11-02 it did, and two weeklies were archived unseen).
Where the contributor report numbers come from
Every figure on the contributor report card is read from GitHub, never from the computer that posts it — so two machines watching the same repository report the same numbers, and a machine with no copy of the code at all still reports them correctly.
- Branches landed counts the branch merges that reached the integration branch during the period, crediting each one to the author of the branch it brought in, once per branch however many commits it carried.
- Commits written and landed come from the same read: an authored commit counts as written whether it sits on the integration branch or still in an open pull request, and only the ones on the branch count as landed.
- Waiting on a merge is the open pull requests, with the commits each still holds.
- If GitHub cannot be read, the card still posts with the other figures and the branch counts read zero for that edition; the app logs the reason.
Turning a built-in notice off from the card
Omniscio's own advisory cards each carry a quiet "Turn off these notices" button in the card's bottom bar (pushed far left so it never out-shouts the card's main action). One click turns that one kind of card off, confirms with a toast, and archives the card; Settings → Notifications has the matching toggle to turn it back on. Turning a notice off only silences the card — the underlying work keeps running, exactly as turning off the stuck-install notice does not stop the cleanup.
Cards with their own off-switch today: low-memory warnings, AI-cost alerts, stuck-install cleanup notices, screen-grab block notices, inbox-backlog nudges, watch-time recaps, and input-method shortcut warnings. Alerts that report a genuine failure you need to act on — reconnect Gmail, sign in again, a dead keyboard shortcut — deliberately have no per-card off-switch, so a real problem can't be silently hidden.
Relationship to Drip
Drip is a second producer of the same primitive. On release, the scanner calls the same createAlert service, writing rows into the shared inbox_alert_items table with source_kind = 'drip'. Those rows surface via the drip inbox integration (listDripSourcedInboxItems, filtered by source_kind), NOT the alert integration — each agent-raised row has source_kind = 'agent', so the two integrations never overlap in the inbox.
A drip-sourced alert reuses its drip_item's id, so the inbox row (alert) and Drip's detail-view row (drip_item) are addressable by one id: DRIP_ITEM_ARCHIVE/UNARCHIVE dual-write both tables to keep the unified inbox and DripDetailView in sync. The synthetic "drip finished" completion row is the one item that stays a drip_item only (no alert), so it remains excluded from the cross-drip inbox as before. See inbox-alert-contract.md (I9, I10).
Related
- Inbox alerts (part 1) — what an alert is and how you answer one.
- Inbox alerts — the lifecycle — the cooldown, the recently-dismissed view and the Alerts screen.
- inbox-rules.md — the rules that can suppress alerts automatically.
- Inbox alerts (part 6) — the on-screen surfaces an alert can declare: a toast, a top banner, a pop-up window, and the per-alert switch for each.
Last verified 2026-11-03