pb-boss-checkins

pb-boss-checkins — a cron boss re-prompts my own session

Goal

A long mission outlives your attention. Install a cron boss that fires on a local-time schedule and re-prompts YOUR OWN session with a status question, so every tick forces a real report: tasks done, children alive, blockers, next action. Delete the boss when the mission ends.

Prerequisites (reference by name): ap-cron, ap-lifecycle, ap-tasks, ap-spawn-agent (only for the agent-kind variant). The family map is the agentproto skill; the supervise-long-missions group covers the wider supervision playbook.

Steps

1. Get your own session id

The boss targets a session id, so know yours (from your spawn descriptor or session_list on the daemon).

2. Create the boss

cron_create({
  schedule: '30 4 * * *',
  action: {
    kind: 'prompt-session',
    sessionId: '<my session id>',
    prompt: 'report status: tasks done? agents alive? blockers? next action?'
  }
})

Five-field cron in LOCAL time: minute hour day-of-month month day-of-week. The response returns the job id and nextRunAt — note both. A morning tick (04:30 Paris in the example) fits the credit-reset cycle; pick whatever cadence keeps the mission honest.

3. Answer every tick with a real status

When the boss fires, it wakes your session with the prompt. Answer it with facts, not intentions:

• task_update({taskId, rev, note}) notes on the task board for progress.
• session_tree({}) to confirm children are alive or finished.
• Name the next concrete action, or declare the mission complete.

4. Delete the boss when the mission ends

cron_delete({ jobId: '<jobId>' })

The job is a persistent host-level job that survives daemon restarts — a forgotten boss pings your session forever. This step is not optional.

Variants

• Machine check-in (kind: 'command'): run an allowlisted command on each tick — a build, a test, a heartbeat script. Good when the status you need is green/red rather than a report.
• Independent reviewer (kind: 'agent'): spawn a BRAND-NEW agent session on each tick with a review prompt. Costs more per tick, but the reviewer has no stake in your session's narrative and audits from a fresh context.
• Follow-up on a child: kind: 'prompt-session' also works against a child session id, turning a periodic check-in into a per-agent heartbeat.