Omniscio

Tasks (markdown outliner, in development) (part 12)

The AI reads a full per-task record: how it orders the day's plate with a written reason for each pick, how it classifies tasks by urgency and importance, the day-log you can deliberately close out, snooze history, planning personalisation, task ideas caught from your sessions, delegation records, the waiting-on dependency clock, meeting alerts, and calendar fusion. Each section below states what actually ships and marks the parts that are designed but not yet built, so a reader is not told a capability exists before it does.

On this page

What it is

This is part 12 of the Tasks (markdown outliner, in development) page. It carries the next stretch of the material on that page, moved here because a single page is capped at 40,000 characters.

Where to find it

Reach this part through Tasks (markdown outliner, in development) — it lists every part and explains where the feature lives in the product. Everything below is reached from the same place.

How it behaves

Everything below is the behaviour, detail and edge cases that belong to this stretch of the Tasks (markdown outliner, in development) page.

Closing out the day — opening, confirming and the evening reminder

Press Close out my day, the moon button in the Today header, to open the close-out. It answers one question: is tomorrow safe? When nothing is at risk it says Tomorrow's safe; otherwise it says how many items need your eyes and lists them, each with a fix button (below). It also sums up today — what you finished, what rolled over and what is still open — and previews tomorrow.

Press Confirm & rest to record that you looked at the day. Press Remind me later to put the close-out off for 90 minutes; it tells you when it will come back. Closing the window without either cancels any pending reminder for that evening.

From 6 PM the close-out opens by itself once a day while you are on Today with something on your Today list, unless you already confirmed today. When a Remind me later runs out, it opens once more. To stop it opening by itself, turn off Evening wrap-up in Customize. The moon button always works.

When Caught is on, the close-out also lists up to five waiting Caught suggestions, each with Keep (or Check off) and Dismiss. Dismiss shows a "Suggestion dismissed." note with an Undo button for a few seconds, and Ctrl+Z (Cmd+Z on a Mac) brings it back too, as long as the close-out is open and you are not typing in a text box. The dismiss goes on the same Tasks undo list as every other Tasks change (dismissCaughtWithUndo in tasks-caught-slice.ts), so Ctrl+Z undoes the most recent Tasks change; the close-out listens for the key itself (useCloseOutUndoKey in TaskV2CloseOut.tsx) because focus can drop to the page once the dismissed row disappears. If bringing it back fails, it says "Couldn't bring back that suggestion." and the undo stays available to try again.

Evening wrap-up note. Once a day, at or after your Evening wrap-up hour (6 PM unless you change it), Tasks puts one Evening wrap-up note in your inbox: how today went and what tomorrow looks like. Its Open Evening wrap-up button opens Tasks on the close-out, the same review Close out my day opens; with Tasks turned off, it tells you so and offers the way to Lab. It is skipped on a day you already confirmed with Confirm & rest, and never sent twice for the same day. The note keeps up with the day: while it is in your inbox, each check rewrites it with that day's current counts, and it retires itself once the day is over or the day is confirmed, so a note from yesterday never waits in the inbox telling you about yesterday. It is on by default; turn it off with Evening wrap-up note in Settings > Tasks (Daily check-in group), where the hour is set too. The service checks every 15 minutes (tasks-v2-closeout-digest-service.ts, switch tasksV2CloseoutDigestEnabled, hour tasksV2CloseoutHour); an install that saved the old "off" value before the switch had a screen is moved to on once (migrate-tasks-v2-closeout-digest-default-on-for-existing-users.ts).

If your last confirmed close-out was yesterday or earlier, the close-out shows a short note naming that date. After five days without one, or if you have never confirmed one, the note says it needs your eyes.

Closing out the day — one-click fixes for tomorrow

When you close out your day, Tasks checks whether tomorrow is safe. It looks for tasks that are due today, overdue, important but not planned, due soon but not planned, or waiting on someone with no follow-up set. Each risk it finds gets a button that starts a helper session to deal with it, so you can fix it without leaving the close-out:

Risk Button What it starts
Due today Focus the task's own Claude agent, on a light model
Waiting on someone, no follow-up set Follow up a DeepSeek helper
Due soon, not planned Plan a DeepSeek helper
Overdue, or important and not planned — and a big task Break it down a helper session that splits the task into clear steps (DeepSeek or Claude, as below)
Overdue, small task Reschedule a DeepSeek helper
Important and not planned, small task Plan a DeepSeek helper

The Follow up, Plan, Reschedule and Break it down helpers run on DeepSeek only when DeepSeek session spawning is switched on and a DeepSeek key or supply is ready on this computer; otherwise they start a Claude helper instead, so a button never fails for a provider you have not set up. Focus always starts the task's own agent.

A task counts as big when it is estimated at more than 60 minutes, or its importance is set to High. Every button starts a session that uses AI. Once any of these sessions has started, its button stays off for the rest of that close-out, so a second click can't start another paid session. The reschedule, reminder, plan and focus buttons also show a "Session started." note with an Open session button, because those sessions start in the background. If a session failed to start, you can click its button again.

The helper gets the task's title as information, not as instructions. The title is cleaned with the same safety filter the other Tasks helpers use and shown in a quoted block the title itself cannot close, under a line telling the AI it is data. So a task kept from someone else's email cannot use its wording to steer the helper, for example into setting up a recurring job.

These buttons work only inside the app. There is no command-line route that starts one of these helper sessions. A script or an agent can read the risks with GET /tasks-v2/closeout, but it cannot start the fix for any of them.

About you — what Tasks has learned about you

Open Tasks → ⋯ More → About you to see, and correct, what Tasks knows about how you work. It has two halves:

  • Your planning profile. How many items you like per day, how detailed your plans should be, your standing priorities, a free note, and the four facts Plan my day learns about you: preferred work hours, meeting buffer, task type preferences and energy patterns. Every plan reads all eight, so you can see and correct each one here. When a plan places tasks around your calendar, it leaves the meeting buffer you set between meetings and tasks (15 minutes if you have not set one). Clear wipes the profile, so the next plan starts fresh. Accepting a plan never replaces words you wrote in the text fields. When Plan my day learns something new, it adds it on a new line below your text, and skips it if it is already there or if your text plus the new line would be too long for that field. If a field is empty, the new text is saved as is. The two number fields (items per day and meeting buffer) take the value you reviewed.
  • What it has learned. Once a day Tasks writes short notes about how you work, such as how much you usually finish, tasks you keep carrying over, or kinds of task you keep snoozing. You can edit a note (it then counts as yours, and the daily pass never changes it again), dismiss one, or reset them all. A reset also deletes the day-by-day history the notes are learned from. Notes you wrote or edited survive a reset. Tasks' own notes are refreshed by each daily pass: one it no longer finds true is removed by the next pass. Notes you wrote or edited are never removed by the daily pass.
  • Where the notes are kept. The notes are part of About You, the same place Omni keeps what it learns about you, and each signed-in account has its own. They are stored only on this computer. You also see them in Settings → Mission Control → About You, marked From Tasks, and erasing one there removes it here too. With Global Memory off only Tasks reads them; with it on, your other sessions can use them as well. About you shows the newest 200 notes; Tasks adds at most five new ones a day, and your own notes never count against that five.

Opening the panel is free. Only the daily pass that writes the notes uses AI. When Tasks plans your day, it marks each note as either something you told it or something it guessed from your activity. A note you edited counts as something you told it. Every note is sent in full: Tasks does not cut your notes short or send only the first few.

Omni can suggest turning delegation on

On an existing install Do it for me stays as it was, so it is often off. If Omni is on and you clearly rely on AI, Omni's Suggestions list offers to switch it on for you. Nothing changes until you say yes. A no keeps it quiet for 90 days, and so does switching it off yourself. See Omni for how Omni decides.

Deleting a task stops its agents

When you delegate a task, a background agent works on it, and that agent uses paid AI. If you delete the task, or file it away, while the agent is still working, Tasks stops the agent. The run is marked failed with the reason "Stopped because its task was deleted or filed away." Tasks checks for this whenever your tasks change, when the app starts, and every five minutes. If you undo the delete, the task comes back with that run still marked failed, so you can delegate it again.

Deleting the task also closes any chat you launched from it (Launch agent or g), exactly as if you had closed it yourself — filing it away leaves the chat running. Undoing the delete brings the task back but does not restart the chat.

A delegated agent can read, but cannot act without asking

A delegated agent is held to the read-only permission level for its whole run, whatever your global permission level is. It works inside its own scratch folder, where it reads and searches freely; reading a file anywhere else, or using any web tool, waits for your approval. Anything else it reaches for, such as running a command, writing a file, searching the web or using a connected app like mail or chat, waits for your approval instead of going through on its own. It is not given the app's full control key. Like every agent, though, it does get its own limited key, and that key can delete a single task. Using it means running a command, and every command waits for your approval — so reading a command before you approve it is what keeps "draft only, never send" true even if the agent tries to send something. A task deleted this way waits in Recently deleted for 30 days, where you can restore it.

Accepting a draft saves it on the task

When you accept a delegated draft, Tasks saves it as a note on that task (labeled AI draft, AI research or AI summary), and the toast says so. Open the task's context to read it again, and any agent launched on the task later receives it. In the review window, Delegation history lists the task's past runs; open a row to read that run's draft, or the reason it produced none. It works the same on a paired phone, and it only ever shows runs on tasks in the account you are signed in to.

Every task delete is also written to the app's audit record. The record keeps the task's id and how many items the delete removed, never the task's text. It is kept for two years. This covers deletes in the app, from the command line, and bulk deletes.

Delegations count toward the launch limit

Each Do it for me delegation starts a background agent, so it uses one of the 200 task-agent launches you get each day — the same limit as Launch agent and inbox triage (see Per-task agent chat). The app also starts at most 30 delegations a minute; past that it turns the delegation down and asks you to try again shortly.

Delegation status on a task row

A task row carries the delegation's state as a chip:

  • AI would: … appears on a task Tasks thinks it could hand off; it lets you pick what kind of help it offers.
  • Drafting pulses with how long the background agent has been working.
  • Needs you: open means the agent is waiting on you; clicking it opens its session in Tasks' own Sessions pane, so you never leave Tasks.
  • Draft ready — review opens the review.
  • Draft failed — retry starts the same kind of draft again straight away; that is a new background run that uses paid AI.

When a delegation counts as done. When the agent's turn finishes cleanly, its last message you can see becomes the draft you review (a hidden reply the agent gave to one of the app's own background checks is skipped, and a session whose messages were moved to the archive is still read) and the run shows Draft ready about ten seconds after the session finishes. That short wait is deliberate: agents sometimes carry on working just after a turn looks finished, and the wait stops half-finished text becoming the draft. A background check every five minutes catches anything that was missed. If you replied after the agent's answer and its last message is only a short closing line (less than half as long as that answer), the draft keeps the answer with the closing line below it, so a back-and-forth never leaves you with just the sign-off. This holds when the agent marked that closing line as its final message too: the answer and the closing line are both kept.

When a delegation needs you. If the agent stops to ask a question, to ask permission to use a tool, or to get a plan approved, the task shows Needs you: open instead of Drafting, and clicking it opens the session in the Tasks Sessions pane (in the Mission Control Tasks panel, it opens Tasks with that chat showing). The chip goes back to Drafting once you answer, and the run carries on to a draft. The app reads the wait from the session, so the delegation list on the command line carries the same fact as awaitingUser: true. The run keeps waiting however long that takes. If the agent's message was really the finished work and only looked like a question, archive the session: its last message becomes the draft. Replying instead makes the agent's next message the draft, though a short closing line after your reply keeps the earlier answer above it. Archiving a session stopped on a permission or a plan marks the run failed, because nothing finished. If the run is stopped before it finishes — the Stop button, an error, or the session ending mid-turn — it is marked failed with "The AI session was stopped before it finished", never passed off as a draft. A paused run is left alone however long it waits. A run waiting on a usage limit, or another hold that clears on its own, is left alone for up to three days from when you delegated it, then marked failed so you can delegate it again. Any other run that makes no progress for two hours is timed out. A run that finished while you were signed in to a different account is never timed out for that: it stays waiting and its draft is picked up when you sign back in to the account that started it.

Waiting on someone — follow-up reminders

You can mark a task as waiting on a person and set when to follow up. When that time passes, a reminder titled "task — follow up with person" appears in the Tasks section of your Inbox. It stays until you deal with it, and clears itself when the task is finished or no longer waiting. Tasks also tells the day planner who you are waiting on and since when, so it can plan around it.

Not built yet: Tasks doesn't track who holds you up most often, or how long a person usually takes to reply. It only remembers names you used before, to suggest them as you type.

Meeting heads-up

You can mark a task as a meeting and give it a time. Tasks then gives you a heads-up before it starts. The default is 10 minutes before; change it under Meeting alarm → Heads-up (minutes). Loudness sets how it rings: Full-screen, Toast (the default) or Silent. Toast and Full-screen play your default alarm sound once (the Alarms default sound, at your notification volume); Silent plays nothing. Full-screen also brings the Omniscio window to the front once, unless the Alarms "bring to front" setting is off. While Focus Mode is on, a meeting still shows its ring but makes no sound and leaves the window where it is. (The event clock puts soundId on the TASKS_V2_EVENT_FIRED push and raises the window in tasks-v2-event-service.ts; the renderer plays it through the same playAlarmSoundOnce helper the alarm banner uses.)

If you keep Tasks open in its own window, the heads-up rings there too, so a meeting still reaches you while the main window sits behind it or is minimized.

On a paired phone, a heads-up that is still ringing shows when you open or come back to the app, even if it started while the phone was away. The phone only ever shows meetings from the account you are signed in to.

If the app was closed or your computer was asleep, a heads-up that is more than 30 minutes late is skipped without ringing. That way, reopening the app never replays a pile of meetings that already happened. For a snoozed heads-up, the 30 minutes count from when the snooze ends, so Snooze 10m still rings even long after the first heads-up.

A heads-up always leaves a card in your Inbox, whatever its loudness. Every heads-up, including a Silent one and one skipped for being more than 30 minutes late, also puts a "Deadline reached" card for the meeting in the Tasks section of your Inbox, and that card goes to your phone and, if you turned them on, by email and text like any task reminder (see Reminders by email, text and phone). After the app has been closed, each meeting you missed sends one such card.

A repeating meeting moves on by itself. You do not have to check a repeating meeting off. An hour after it starts, Tasks makes its next copy and files the one that just happened away in Archived (it is not marked done), so the meeting shows once, at its next date. Restore it from Archived if you want to check it off.

When a reminder card is held, its email and text wait too. If the inbox holds a task reminder for a few minutes before showing it, the reminder's email, text and phone push wait with it and arrive together when the card shows. If the card is dismissed or silenced while it is held, none of them are sent. This waiting has limits: if the card is still held more than 30 minutes after its hold ends, the email, text and phone push for it are not sent at all and only the Inbox card appears. A restart does not drop them: held cards are picked back up when the app starts again and send as soon as they show. At most 200 held reminders wait at once; beyond that the oldest is dropped and arrives as the Inbox card only, without email, text or phone push.

The day planner sees only each meeting's name and time. It plans your tasks around your meetings, but it doesn't know how early you want warning, so it does not set aside prep time for you.

Not built yet

  • Snooze nudges. Tasks counts how often you snooze a task. The alarm-clock icon on a task that came back from a snooze shows that count, and the day planner sees which tasks you snooze a lot. There is no automatic nudge ("you've pushed this off three times — want to break it down?"), but Learn about your habits on Today names the task you snooze most and lets you Break it down or Drop it yourself.
  • Learning from your estimates. You can give a task a time estimate, and the day planner sees it. But Tasks does not yet compare your estimates with how long things really took, and it does not size your day around that.
  • Learning from delegation. Tasks records how each delegated task turned out, but it does not yet use that to improve which tasks it suggests delegating.

Import tasks from other apps (CSV/JSON)

Users can bring existing tasks from Todoist, Trello, Asana, TickTick, and similar apps by importing a CSV, TSV, TXT, or JSON file. The import opens from the empty-state button in the outliner, in the "All Projects" master view and on an empty Today, and from Import tasks… in the Tasks ⋯ More menu. All four appear only in the desktop app: the file picker belongs to the computer running Omniscio, so the phone and browser views hide them rather than open a file box on that computer.

Right under it, Import notes from Obsidian… (desktop app only) brings Obsidian notes into the Vault rather than into Tasks: it opens the same one-press Obsidian import the Vault and Settings use, and if the Vault is off it first asks to switch it on and names the folder the notes will go in. See KMS Obsidian Import.

Wizard flow: a dialog with four steps — (1) choose a file with the system file picker, (2) map columns to task fields (suggested from the headers and adjustable; the preview stays disabled until a column maps to the task title), (3) preview and confirm, and (4) Import completed with errors when some rows could not be written. Nothing is written until the user confirms the preview. The preview shows how many rows will import, which have no title (skipped), which are duplicates, and which new projects (lists) will be created. The failed step lists every row that did not import with its reason, and Try again retries only the rows that can still succeed. If the app loses touch with the import before it answers, it says so and asks you to check your list before importing the same file again.

Preview warnings: two warnings can appear beside the preview. When every date in the file reads the same either day-first or month-first, it says so and shows how the dates were read, so they can be checked in the preview before anything is written. When a cell holds a value the import cannot read, it shows the value and says the row will import without it.

Duplicates: on by default and toggleable at both steps. A row is a duplicate when its text (trimmed, case-insensitive) already exists in the list it would land in, or appeared earlier in the same file; a row bound for a project that does not exist yet is never matched against another list. Unticking "Skip N duplicate tasks" on the preview imports them as well, and the confirm button's count follows the choice. The preview and the import run the same planning pass, so the preview counts what the import does.

Supported fields: task title (text), description, due date, project/list, done/completed status, completion time, importance/priority, and tags. A completion-time column (Completed Time, Completed At, Done Date, Todoist's COMPLETED_DATE, also with a qualifier such as Completed Date (UTC)) is suggested as the completion time and never as the done flag or the due date, so a finished task keeps the day it was really finished rather than the day it was imported. A row with a completion time and no readable done value imports as done; a status column that says open still wins. Dates parse ISO 8601, US (MM/DD/YYYY), European (DD.MM.YYYY), and named-month formats, and a date whose day or month does not exist (30 February, say) is left empty rather than shifted into the next month, whichever format carried it. The done flag recognizes true/1/yes/done/complete/completed/finished/x/✓. Importance maps 1–4 and words (low/medium/high/critical) to a 1–3 scale. Tags split on commas, semicolons, or pipes; in a JSON file whose tags value is a list of objects such as { "name": "release" }, the import reads each object's name as a tag (a plain string is its own tag) and a value that names nothing adds no tag, so a structured tags value never becomes a tag made of its raw JSON. Three exports are recognized by their shape: a TickTick export (headers include Title and List Name) skips the preamble lines above its header row, reads Priority as 1 low, 3 medium, 5 high (0 none), prefers List Name for the project and Due Date for the due date, and treats Status 2 as done; a Todoist export (headers include TYPE, CONTENT and PRIORITY) imports only task rows and skips section and note rows as "Not a task" (an issue-tracker file whose TYPE means bug or feature, and carries no CONTENT/PRIORITY pair, still imports), and reads Priority on Todoist's own scale — 1 high, 2 medium, 3 low, 4 none — so the values land the right way up on the dial instead of being read as a low-to-high scale.

Undo: the success toast offers one Undo for the whole run, and the same entry sits on the app's undo stack (Ctrl+Z), like other Tasks actions, for the rest of the app session. It soft-deletes every task the run created and removes a list only if this run created it and nothing else is in it — a list the user already had, or has since added a task to, is kept. If an Undo fails partway, the tasks it already removed leave the open board straight away, the error is shown, and the same Undo goes back on the stack, so pressing it again (or Ctrl+Z) finishes the job rather than leaving a half-undone import. The run is rebuilt from the run id saved on each imported task, so an undo of an older run still works after many later imports — including removing the empty list it created — and the CLI undo route still reverses a run whose id the caller kept across a restart.

Limits: 5,000 rows, 10 MB file size, 20,000 characters per cell. A file over any limit is refused with a plain message, never cut down to the limit. UTF-16 files are rejected with a re-export prompt; JSON must be an array of objects. One import runs at a time.

A field the importer cannot read is reported, not dropped. A due-date cell, a priority word, an urgency word, a completion time or a completed-marker the mapper does not recognise is named on the result (droppedFields: the rows and which field each lost) and in the wizard's confirmation, so an import of "12 tasks" can no longer hide a discarded deadline. An empty cell is not a drop — only a cell that held something unreadable is, and a priority cell that names the source's own "none" value (TickTick's 0, Todoist's 4) is a value it read, not a drop.

The overview, the other parts, and everything else worth reading next all sit on Tasks (markdown outliner, in development).

Last verified 2026-11-19