| name | ops-cron |
| description | Schedule recurring tasks that run automatically via a system-level watchdog, even when Claude Code is closed. Job definitions persist in cron/jobs/ as the registry. The watchdog (launchd on Mac, Task Scheduler on Windows) checks every hour and runs due jobs headlessly via claude -p. Triggers on: "schedule a job", "cron job", "run this every morning", "automate daily", "recurring task", "set up a scheduled job", "check scheduled jobs", "list scheduled jobs", "install watchdog", "uninstall watchdog". Does NOT trigger for one-off tasks or in-session reminders.
|
Scheduled Jobs
Recurring tasks that persist across sessions and run automatically. Job definitions live in cron/jobs/. Execution uses a system-level watchdog that runs claude -p headlessly -- jobs keep running even when Claude Code is closed.
How It Works
The watchdog is a background process registered with your OS scheduler (launchd on Mac, Task Scheduler on Windows). Every hour it scans cron/jobs/, checks what's due, and runs each due job via claude -p with the model and budget from the job file. Each execution is stateless and self-contained. Install once, forget about it.
Outcome
Job definition files in cron/jobs/. Automatic execution in the background. Logs for completed runs in cron/logs/.
Context Needs
| File | Load level | How it shapes this skill |
|---|
context/learnings.md | ## ops-cron section | Known issues with specific jobs |
No brand context needed -- this is infrastructure.
Step 1: Determine the Action
| Action | User says | What to do |
|---|
| Create | "schedule a job", "run X every morning" | Build a job file (Step 2) then offer to install watchdog (Step 3) |
| List | "list jobs", "what's scheduled" | Read all files in cron/jobs/, show status table |
| Install | "install watchdog", "start scheduled jobs" | Run bash scripts/install-watchdog.sh |
| Uninstall | "uninstall watchdog", "stop all jobs" | Run bash scripts/uninstall-watchdog.sh |
| Stop job | "stop the X job", "disable X" | Set enabled: false in the job file |
| Remove | "remove the morning briefing job" | Delete the job file |
| Logs | "check job logs", "did the job run" | Read latest from cron/logs/ |
| Status | "is the watchdog running" | Check for launchd plist (Mac) or scheduled task (Windows) |
Step 2: Create a Job File
Ask the user (max 3 questions):
- What should it do? -- The task in plain language
- How often? -- "every 30 minutes", "every 2 hours", etc.
- Any constraints? -- Budget cap, specific model
Create the job file at cron/jobs/{job-name}.md. See references/job-format.md for the format.
Key rules for job prompts:
- Self-contained -- each execution has no memory of the last
- Reference specific file paths
- Say where to save output
- Keep prompts focused -- one clear task per job
Step 3: Install the Watchdog
If the watchdog isn't already installed, offer to install it:
Mac:
bash scripts/install-watchdog.sh
Windows (PowerShell as admin):
powershell scripts/install-watchdog.ps1
The installer:
- Checks
claude CLI is available
- Calculates estimated daily cost from enabled jobs
- Registers with the OS scheduler
- Starts running immediately
Step 4: Verify
Show the user:
Watchdog: installed (checking every 60 minutes)
Scheduled jobs:
trending-research every 2h enabled sonnet $0.50 cap
inbox-check every 30m enabled haiku $0.25 cap
Estimated daily cost: $0.40 - $1.50
Logs: cron/logs/
Schedule Reference
| User says | schedule value | How often it runs |
|---|
| Every 10 minutes | every_10m | 6x/hour |
| Every 30 minutes | every_30m | 2x/hour |
| Every hour | every_1h | 1x/hour |
| Every 2 hours | every_2h | Every 2 hours |
| Every 4 hours | every_4h | Every 4 hours |
| When I open Claude | session_start | Once per session (watchdog skips these) |
No cron expressions. Human-readable intervals only.
Heartbeat Integration
At session start (CLAUDE.md heartbeat step 9):
- Check if watchdog is installed (launchd plist exists on Mac)
- If installed: report status: "Watchdog is active -- your N jobs run automatically"
- If not installed but enabled jobs exist: suggest installing the watchdog
Rules
Updated automatically when the user flags issues. Read before every run.
Self-Update
If the user reports a job failing or a scheduling issue -- update the ## Rules section immediately with the fix and today's date. Also log to context/learnings.md under ## ops-cron.
