Schedule Dates

The dates block names the dates a scheduled run loads, once, derived from its schedule slot. Every run of a slot gets the same dates: the scheduler's run, a run started for the slot with --scheduled-at, a retry, and a backfill run.

schedule: "CRON_TZ=Asia/Tokyo 0 2 * * *"
calendar: jp-banking

dates:
  target_day:     -1d
  target_compact: { from: target_day, format: date_compact }
  month_start:    [-1d, month_start]
  last_open_day:  -1bd
  window_start:   { from: interval_start, format: rfc3339 }
  window_days:    { each: 1d }
  loaded_at:      { format: "%Y/%m/%d %H:%M", timezone: UTC }

steps:
  - id: load
    run: ./load.sh --date ${dates.target_day} --since ${dates.window_start}

For the run fired at 2026-11-10 02:00 in Tokyo, target_day is 2026-11-09, target_compact is 20261109, month_start is 2026-11-01, and window_days is ["2026-11-09"].

Entries

An entry is a shift written as a string or a list, or an object:

Field	Meaning	Default
from	scheduled_at, interval_start, interval_end, next_scheduled_at, or another entry	scheduled_at
each	A step through the window, such as 1d. The entry becomes a list.	none
shift	One operation, or a list applied left to right	none
format	A named format or a strftime pattern	date
timezone	The zone used for shifting and formatting	see below

target_day: -1d means { shift: -1d }. An entry that starts from another entry starts from its instant after its shift, takes its zone unless it sets one, and does not take its format.

Shifts

Operation	Meaning
+N or -N with min or h	Elapsed minutes or hours
+N or -N with d, w, mo, y	Calendar days, weeks, months, or years in the zone
+N or -N with bd	Business days of the slot's calendar
hour_start, day_start, week_start, month_start, year_start	The start of that unit; weeks start on Monday
bd_or_next, bd_or_prev	The same day when it is a business day, else the next or previous one

• A day shift keeps the wall-clock time across a daylight saving change, and a month shift clamps to the last day of a shorter month: March 31 minus one month is the last day of February.
• m is rejected, so a month never silently becomes a minute: write min or mo.
• [month_start, +1mo, -1d, bd_or_prev] is the last business day of the month.

Formats

date (2026-11-09), date_compact (20261109), datetime (2026-11-09 02:00:00), rfc3339 (2026-11-09T02:00:00+09:00), unix (1794157200), or a strftime pattern such as %Y/%m/%d. An unknown directive fails dagu validate.

Lists

{ each: 1d } lists the window from interval_start up to, but not including, the slot, as a JSON array. Monday's run of a weekday schedule owns the window from Friday, so its window_days is Friday, Saturday, and Sunday. Load one day per child with parallel:

steps:
  - id: per_day
    action: dag.run
    with:
      dag: load_day
      params:
        DAY: ${ITEM}
    parallel: ${dates.window_days}

• 1bd lists the business days. A step under one day, such as 1h, needs a format.
• A list holds at most 1,000 values, the parallel limit.

Time Zones

An entry is shifted and formatted in its timezone, else the zone of the entry it builds on, else the zone of the slot's schedule (context.run.timezone), else the server's tz.

Business Days

bd, bd_or_next, and bd_or_prev use the calendar of the schedule entry that fires the slot, its own or the DAG's. A sub-DAG uses the DAG's calendar. dagu validate fails when a recurring start schedule has no calendar.

Where Dates Resolve

• A run resolves its dates once, when it is created, and records them. A retry reads the recorded dates, even after the DAG or its calendar changed.
• ${dates.NAME} resolves wherever ${params.NAME} does, including env, except in the fields that resolve while the DAG loads: consts, params[].eval, and the map form of worker_selector. Pass a date to a script through env.
• A sub-DAG resolves its own dates from its parent's slot.
• A run started without a slot, by hand, from a webhook, from a watched folder, or as a sub-DAG of a parent with none, gets the default slot: the latest fire at or before it was created. A monthly load started on the 15th loads what its run of the 1st loaded. The run counts as that slot's run and records that the slot was the default.
• On a DAG without a recurring schedule, scheduled_at is the minute the run was created, and a date that needs a window, from interval_start or next_scheduled_at, or a list, is refused before the run is created. dagu validate warns about such a date.
• dagu validate reports a ${dates.NAME} the DAG does not declare.