Agent schedules

The schedule tool lets an agent decide when it should run again. This differs from cron: an operator writes cron jobs in config.yaml, while an agent creates a schedule during a turn and gives its future self the note it will need when the wake arrives.

schedule(action="once", when="10 minutes", note="check whether the deploy finished")
schedule(action="repeat", every="weekdays at 9am", note="review overnight alerts")
schedule(action="list")
schedule(action="cancel", id="a3f1")

Every accepted booking returns the absolute time TAI resolved. That echo lets the agent catch a mistaken date or timezone while it can still correct the schedule.

One-shots and recurrence

One-shot when values accept elapsed durations, ISO-like date and time values, clock times, and day words such as tomorrow 9am. Recurring every values can be elapsed intervals (every 30 minutes) or wall-clock schedules (weekdays at 9am). TAI preserves that distinction across daylight-saving changes.

Schedules persist in SQLite. The runner claims a due row before dispatch, so a slow agent cannot start the same wake twice underneath itself. When the service was offline, an overdue schedule is still present on restart; a recurring schedule advances to the next future occurrence rather than replaying every missed interval.

time:
  provider: system
  timezone: America/Los_Angeles

schedules:
  enabled: true
  tickSeconds: 30
  maxPerAgent: 20
  minIntervalMinutes: 15
  maxHorizonDays: 365
  maxDeferrals: 3

agents:
  operator:
    tools: [schedule, tasks, recall]

schedules.enabled: false also removes the tool. The normal autonomous pause stops scheduled wakes. One-shots remain due until the pause lifts; a skipped recurring occurrence advances normally.

There is no operator-facing schedule screen or CLI command yet. Agents can list and cancel their own entries; operators can inspect the agent_schedules table. The full behavior is documented in docs/schedules.md.