ap-cron

ap-cron

When to use

• Work must run on a schedule with nobody prompting it (nightly audits, periodic sweeps).
• A long-running mission needs periodic check-ins that keep its context alive.
• You want a delayed one-shot action ("in 20 minutes, poke the deploy").

Create a job

cron_create takes a 5-field cron expression in local time and exactly one action. label is the human-readable name; recurring defaults to true.

{
  "schedule": "0 9 * * 1-5",
  "label": "morning-standup",
  "action": {
    "kind": "agent",
    "adapter": "claude-code",
    "prompt": "Summarize overnight commits and draft the day's priority list.",
    "cwd": "/repo",
    "model": "claude-sonnet-4-5"
  }
}

Three action kinds:

• command — run an allowlisted executable. command is the basename (git, not /usr/bin/git) and must be listed in <workspace>/.agentproto/allowed-commands.json.

{
  "schedule": "*/30 * * * *",
  "label": "health-check",
  "action": { "kind": "command", "command": "curl", "args": ["-fsS", "http://localhost:8080/health"], "cwd": ".", "timeoutMs": 15000 }
}

• agent — spawn a brand-new session on each firing. Use for self-contained periodic work.
• prompt-session — re-ping an EXISTING living session by id.

{
  "schedule": "0 8,14,20 * * *",
  "label": "mission-checkin",
  "action": { "kind": "prompt-session", "sessionId": "sess_abc123", "prompt": "Check-in: post progress, blockers, and next step." }
}

Inspect, test, delete

cron_list({})          // every job, active + inactive, with schedule, lastResult, nextRunAt
cron_run({ "jobId": "job_01J..." })     // fire NOW, bypassing the schedule — the test path
cron_delete({ "jobId": "job_01J..." })  // permanent removal

After creating a job, call cron_run once and read lastResult from cron_list before trusting the schedule. A fired cron does not notify you — lastResult is the only feedback.

Durable check-ins on a long mission