Support one-time scheduled tasks

[manvankaliber]

Problem

I migrated scheduled automations from Codex to T3 Code. Codex tasks include runs for a specific future date, but T3 Code currently offers interval and fixed daily or weekly schedules only. The workaround is a daily task that checks the date and exits on every other day, which creates an unnecessary agent run each day.

Proposed solution

Add a one-time schedule that runs at a chosen date and time in a selected time zone, then marks itself complete after it fires. It should be durable across restarts and editable or cancellable before it runs. This request is for one-time triggers, not a new annual recurrence type; annual workflows can arrange their next one-time run after completing.

Migration context

In my Codex migration, nine automations use annual dates and one is a one-time run. The daily date-check workaround creates repeated no-op runs for the annual tasks and the one-time task until its target date.

[wrgrant]

Having launch condition scripts that run before the job could also help with this. But I definitely like the idea of one time scheduled jobs.

[manvankaliber]

Point is that a job runs once, not that another job has launch condition scripts launching other jobs. Sometimes I just want to set a reminder in the future and have it summarize stuff relevant at that time before sending me the reminder or whatever

[robin-liquidium]

+1. We hit this while moving our automations from systemd timers to T3 scheduled tasks. We have three date-specific one-shots this month: a held-back commit push that has to happen on a set date, a loan due-date check, and a governance vote reminder. They stay on systemd for now because the workaround doesn't hold up:

• A weekday-limited fixed_time with a "delete this task first, stop if today isn't X" prompt fires on every matching weekday until the date. One of ours is 8 days out, so it would wake its thread once for nothing.
• If the host is down more than 10 minutes at fire time (MISSED_FIXED_TIME_GRACE_MS), the run is skipped and the next one is a week later, on the wrong date. For a one-time task that means it's silently missed.

So on top of the proposal: a one-time run that was missed while the server was down or asleep should fire once when it comes back (like systemd's Persistent=true), not be dropped. And being able to create one over MCP (e.g. schedule_task with { type: "once", at: "2026-10-25T23:00:00Z" }) would let agents arm their own follow-ups.