Reminders

Ask Iva to remind you, and she will — at the right minute, in your timezone, in your chat. A reminder is not a note she might recall later: it is a row in data/, picked up by a dispatcher that wakes every minute. Restart the server, update Iva, roll the version back — the row is still there, and anything that came due while the machine was down fires on the first tick after it comes back.

How to ask

Plain words in the chat, the way you would tell a person:

• «remind me in 30 minutes to call the clinic»
• «remind me at 14:30 to send the invoice»
• «on 14 September at 09:00 — the contract deadline»
• «every weekday at 09:00 remind me about standup»
• «what reminders do I have?» — she lists them
• «cancel the standup one» — she finds it and removes it

Two kinds: a one-time reminder fires once and is done, a repeating one takes its next time from the calendar and waits for it. You never write the date arithmetic yourself and neither does she: you say when, Iva turns it into an exact moment and tells you back the time she stored.

Everything is read in your timezone — the one Iva is configured with, not the server’s. A date that does not exist on your clock (the hour skipped by a daylight-saving change, 31 February) is refused with that reason instead of being rounded to something nearby. So is a time in the past, and anything more than a year out.

What happens at the time

Your reminder text is an instruction Iva gives her future self. At the due minute she wakes in a fresh session that knows nothing about the conversation you asked in, does what the text says — with her tools, so «find the news and send it» really goes and finds them — and the code sends the final text of that turn back to the chat and topic where you asked. One turn, one message.

That is why the text is written to stand on its own: «remind me about the standup» and «collect yesterday’s tasks and list what is still open» both work, but «do that thing we discussed» will not — by then the discussion is gone.

If the turn cannot run at all (the model provider is down, the agent breaks, the answer comes back empty), the code sends your text back to you exactly as you dictated it, and the row says what broke. So a reminder always arrives, with or without the model.

Nothing repeats. A row fires once: the moment it fires it is marked, and a repeating row moves on to its next time. There is no retry ladder, no window that removes a reminder hours later, and no warning per failure — the fact of the firing stays with the row: fired_at, delivered and the reason in error.

Where to see the list

/menu → ⏰ shows the nearest reminders, what failed to go out, and one line about the dispatcher: when it last ticked (its pulse is the file data/reminders.tick, touched every minute). A fresh pulse means the machinery is alive; a stale one, or none at all, is your signal that reminders are stored but nothing will fire until Iva is running again.

You can also just ask in the chat: she lists the same rows with their ids, the next time, and the last firing with the delivery fact. Rows stay in the list for a day after they fire — that is where you see whether the text went out. Ask her to cancel one and she removes it by id.

When something breaks

• Nothing arrived. The row shows delivered: false and the reason in error. Fix the chat settings or the token, then ask again — a one-time reminder has already fired, so put a new one.
• The agent turn could not run. Your text still arrives, verbatim; the reason is in error and in the journal (journalctl --user -u iva.service | grep reminders).
• The dispatcher is not ticking. The list warns that the reminder is stored but will not fire, and iva doctor says the same by the pulse file. It also lists every reminder that fired in the last day and did not go out.
• no activity for 180000ms during an external-job check. The limit counts silence in the Eve event stream, including a blocking tool that produces no events; it is not a three-minute limit on the whole turn. A check should read a bounded progress/status snapshot and return it, even if the job is still running. It should not wait for process exit, run tail -f or keep checking in a loop. The background-check skill describes this procedure. Code sends the final status; the model must not send a report separately and then keep waiting. A later check belongs in a new request from the owner, not a new reminder created inside the firing turn.

Scheduling and delivery

Use the remind tool for schedules. Delivery goes through the turn response; use iva post for an external chat.

Limits

• Back to where you asked. A reminder returns to the chat and forum topic it was created in; another person as the recipient is not supported. Reminders created before 0.4.4 go to the owner chat from the settings.
• No more often than every ten minutes. A repeating schedule tighter than that is refused: that is a monitoring job, not a reminder.
• Your timezone, one of them. Everything is computed in the timezone Iva is configured with.
• A reminder is one turn, not a project. It fires, Iva does what the text says in that single turn and answers once (or, if repeating, moves to the next occurrence). The turn has no time limit: it runs as long as the work needs, and only two things end it — /stop sent from the same chat, or three minutes of silence from the model. The reminder uses /stop in its originating chat. Private chat drafts use the native ■ control when the Telegram channel owns the turn. She sets no new reminders in it: that turn may list and remove them, not add.