Files
pipelines/internal/author/skill/references/schedules.md
T

43 lines
2.5 KiB
Markdown

# Independent saved run schedules
A schedule is not part of a workflow. It pairs an explicit pipeline ID/version
with run information (`input.prompt`, server `workspace`, optional model/stage
overrides, title, notifications, output format, and `write_paths`) and a timer.
One generic research workflow can have many topics/prompts on different timers.
Scheduling an existing workflow never authorizes creating/editing/publishing a
workflow or altering its permissions. Ask before enabling unattended runs unless
the user already requested that schedule.
The web app exposes **Pipeline → Schedules → Add schedule** and **Schedule this
run**. The latter copies original input/options, not output, and preserves
run-level write restrictions. Publishing a workflow creates no timer.
All scheduling commands use the existing JSON envelope through `pipelines rpc -`:
- `list_schedules` with optional `pipeline_id` → `schedules`.
- `save_schedule` with a full `schedule` object → `schedule`. Omitted/zero ID and
revision create; edits require the returned positive `id` and current
`revision`. The record contains `name`, `pipeline_id`, `version`, `enabled`,
`input`, `workspace`, and `rule`.
- `delete_schedule` with `schedule_id` and `expected_revision` removes future
occurrences only. History and already admitted/running work are retained.
For creates, use a new top-level `request_key` for each intentional schedule and
reuse it only to retry that same creation after a lost response. A conflicting
payload fails; retrying a deleted creation cannot resurrect it. On revision
conflict, reload and reconcile—never force-overwrite somebody else's changes.
`rule` uses exactly one of `{"every":"24h"}` or
`{"times":["09:00","17:30"]}`, optionally `days` (`sun` through `sat`, omit for
everyday) and `timezone` (IANA, `UTC`, or server-local `Local`/omitted). Intervals
are at least one minute. Workspace is an existing absolute directory on the
execution server, not the browser/client's filesystem. Validate run/model options
against the pinned workflow; unavailable providers still need runtime checks.
**Pipelines → Has schedules** includes configured paused schedules.
**Runs → Scheduled runs** means a real timer-triggered execution, not a manual
run of a pipeline that happens to have schedules. Provenance stays after pause or
deletion. Missed occurrences are skipped, not replayed; overlapping workspace
execution is refused and recorded. Do not add recovery workflows or rerun work
merely because a timer's dispatch was skipped.