Scheduled jobs plus (delivery, chaining, pause-all, batch runs)
Scheduled jobs plus lets a scheduled job say what it found and send it where you want. A job can hand its output to the next job, skip itself when a script says there is nothing to do, and mark a run silent or failed. You can pause every job at once, ask the doctor what is wrong, and type a schedule in plain words. Many prompts can run as one batch and be exported as training lines. Nothing here is on until you switch it on in Lab.
On this page
What it is
Scheduled jobs already run on a timetable, but a job's answer used to be thrown away. This adds the controls for jobs that run when you are not looking:
- Reports. A script job's report is what it printed. A new-session job's report is the final answer its session gave at the end of its first turn.
- Delivery. The report goes to the targets you choose.
- Chaining and continuity. A job can receive its parent's output, other jobs' latest reports, or its own last report.
- A skip check. A script decides whether a run happens at all.
- Markers. A report can say "nothing to report" or "this failed".
- Pause all and the doctor. One switch stops every job; one check lists what is wrong.
- Describe-it schedules, batch runs and ShareGPT export.
Where to find it
- Lab switch: Settings, Lab, Scheduled jobs plus (off by default).
- The job editor (Cron Jobs screen, + Add Job or edit): the new sections are Deliver to, Chain input, Skip check and Limits, plus a Describe it field next to the schedule.
- The Cron Jobs header: Pause all / Resume all, Check jobs and Batch runs.
- With the switch off, none of these appear and stored options do nothing.
How it behaves
Delivery targets
A job can have up to 8 targets. Each is sent once per fire, after any retries have settled.
- Inbox: one card per job, replaced on each run, with an Open job action.
- The session that made the job: a quiet note in that session, never a paid wake-up.
- A file: a new timestamped
.mdfile in a project folder you registered or the default reports folder. It never overwrites a file and never writes into the app's own folders. - Email: sent from your connected Gmail.
- Slack, Telegram, Discord: sent with a credential you saved in Automations. Messages are cut to each service's limit (40,000, 4,096 and 2,000 characters).
- Anything that leaves the computer (email and the chats) has secrets masked first.
- A report that is empty or starts with
[SILENT]sends nothing. A failing target never fails the job; after three failures in a row on one target you get one card, cleared by the next success. - Delivery happens only for an approved job (or one that needs no approval), never for the proving run an approval triggers.
Chained output and continuity
- Chain input lets a run-after child receive its parent's final output. It is not offered when the parent is a session job, because that run ends at launch; read the parent's report instead.
- Reports from other jobs (up to 5) adds each named job's latest report.
- Its own last report adds the job's previous report, so a monthly job can compare itself.
- Each piece is capped at 32 KB and labelled. The context a run was given is stored on the run, so a retry gets exactly the same context. A prompt that looks like planted instructions is held for review like any flagged scheduled prompt.
- A chain stops at depth 10, and a parent that was skipped or held starts no children.
- Webhook body: with the option on, an automation rule's
run_cron_jobaction can pass its body to the job as upstream text; such runs launch at the inbound permission floor.
The skip check (wake script)
For script and new-session jobs, Skip check runs a shell command before each fire, inside the job's slot, for at most 60 seconds.
- Answer format: the last line of its output is JSON.
{"wakeAgent":false}skips the run; no paid work happens and the run is recorded as skipped.{"wakeAgent":true,"context":"..."}runs the job and passes the text as context. - Fails open. An error, a timeout or unreadable output runs the job. There is no forced run after many skips; the doctor reports a long skip streak instead.
- A chained child skips its idle check and wake script, but never the prompt hold.
The markers
Start a report with one of these on its first line:
[SILENT]means success with nothing to send. It is not counted as a job that quietly stopped working.[CRON_FAILURE] reasonfails the run with that reason. It never starts a self-heal session, and it retries only if the job's Retry when the job reports a failure option is on.
Limits
- Permission ceiling: the most powerful level the job's session may run at, on the local and cloud roads. Refused for a provider that ignores permission levels.
- Withheld tools (Claude new-session jobs): tools the session may not use.
- Skills: a prompt line asking the session to use named skills.
- Recipe jobs get the permission ceiling only.
Pause all and the doctor
- Pause all stops every start: scheduled fires, retries, chained children, recipe schedules, plugin fires, self-heal sessions, webhook triggers and proving runs. Sweeps for stuck runs keep going, and overdue alarms stay quiet. A banner shows who paused and offers Resume all.
- Run Now from the desktop still works while paused, and says the engine is paused.
- Resume all re-arms every schedule from now, so each job fires at most once afterwards; a job set to run if missed gets its one catch-up run.
- If the pause file cannot be read, jobs stay paused and a card offers one-click Resume.
- Check jobs (the doctor) changes nothing. It lists each problem with a code, the job and a plain fix: paused, will never fire, overdue, stuck run, project gone, long skip streak, failing delivery target, report deadline missed, stale tick loop, unreadable pause state.
- Ops tasks keep their own pause.
Describe-it schedules
Type a schedule in words in Describe it and see the real schedule it becomes. Examples: in 30m,
tomorrow 9am, every 2h, every 15 minutes, hourly, daily at 7am, weekdays 9am,
weekends 10, mondays and fridays 8:30, monthly on the 1st 9am, a cron expression, or an ISO
date with an offset. A phrase it cannot read is refused with a reason, never guessed.
Batch runs and ShareGPT export
Batch runs launch many prompts as separate sessions.
- Up to 500 prompts, at most 5 at once, each through the normal session create so every spend gate applies. Sessions launch unattended at the permission ceiling you set.
- A dollar cap is checked before each launch, a batch stops when the account is parked on capacity, and an item is never relaunched after a restart.
- Export as ShareGPT writes one line per finished item: the batch's system prompt and the scrubbed transcript. An item with no local transcript is listed as skipped.
Limits you will hit
- A report is kept to 64 KB. A new-session job's report must arrive within its timeout, or six hours if it sets none, otherwise the run records "no report".
- A session's report is read at the end of its first turn; tell the prompt that its first answer is the final one.
- Report, delivery and context apply to script and new-session jobs. Recipe, workflow, automation and plugin jobs do not report.
- Changing a job's options sends an approved job back for approval, and the approval card shows every option.
Related
- Cron jobs — creating and managing a scheduled job.
- Session jobs — jobs that start a Claude session.
- Cron failure alerts — the cards a failing job raises.
- Self-healing — the repair session a declared failure never starts.
- Run if missed — the catch-up run a resume allows.
- Import from Hermes — brings Hermes delivery and chaining across.
Last verified 2026-10-29