Cron Doctor

---
name: cron-doctor
description: "Diagnose and validate cron expressions before they ship. Catches the five silent death-traps: impossible dates that never fire, OR-semantics that fire too often, midnight spikes, uneven step drift, and leap-year February 29."
category: devops
risk: safe
source: community
source_repo: takeaseatventure/devops-skills
source_type: community
date_added: "2026-06-26"
author: takeaseat
tags: [cron, crontab, scheduling, devops, debugging, kubernetes, validation]
tools: [claude, cursor, codex, gemini, opencode]
license: "MIT"
license_source: "https://github.com/takeaseatventure/devops-skills/blob/main/LICENSE"
---

# cron-doctor

## Overview

Cron is deceptively error-prone. The failure mode is **silent** — a syntactically
valid expression that simply never fires, or fires far more often than intended.
`0 0 30 2 *` parses cleanly and then sits dead forever (February has no 30th).
`0 0 1,15 * 1` looks like "1st and 15th if Monday" but actually means "1st, 15th,
**OR** every Monday" — ~6 fires/month instead of ~2.

This skill teaches an agent to catch those before they reach production. It comes
with a zero-dependency validation engine (`scripts/cron-engine.js`, no install
needed) that parses, describes, deep-validates, and computes next fire times.

## When to Use This Skill

- Use when a user writes, edits, reviews, or deploys a cron expression — in a
  crontab, a Kubernetes `CronJob`, a GitHub Actions `schedule`, an Airflow DAG,
  a Celery beat schedule, a systemd timer, or any scheduled task.
- Use when debugging a job that "didn't fire" or "fired at the wrong time."
- Use when a user asks "what does this cron expression mean?" or "when will this
  run next?" or "how often does this run per year?"
- Use when reviewing a CI/CD pipeline or infrastructure config that contains a
  `schedule` field.
- Use when a user pastes a 5-field cron expression and asks for a sanity check.

## Scheduler and calculation limits

Identify the scheduler/dialect and timezone first. systemd OnCalendar is not a five-field cron expression; Quartz and library day semantics differ. This engine uses the Node process local timezone, not CRON_TZ/spec.timeZone, and scans at most 600,000 minutes (~416 days): fewer than requested results are partial, not proof a schedule never fires. It does not model platform delays, DST policy, missed-job recovery or overlap control. Confirm next runs in the actual scheduler. No scripts were runtime-tested during this content review.

## How It Works

### Step 1: Parse the expression

Split on whitespace into 5 fields: minute, hour, day-of-month, month, day-of-week.
Confirm valid ranges:

| Field | Position | Range | Notes |
|-------|----------|-------|-------|
| minute | 1 | 0–59 | |
| hour | 2 | 0–23 | |
| day-of-month | 3 | 1–31 | |
| month | 4 | 1–12 | names (JAN–DEC) accepted |
| day-of-week | 5 | 0–7 | 0 and 7 both = Sunday; names (SUN–SAT) accepted |

### Step 2: Describe it in plain English

State what the user *thinks* it does vs. what it *actually* does. Be explicit
about OR-vs-AND semantics for day-of-month + day-of-week (see death-trap #2).

### Step 3: Run the trap checklist

Check the five death-traps below and flag any that apply.

### Step 4: Calculate next runs and annual fire count

Compute the next 5 fire times as concrete dates so the user can verify the
schedule behaves as expected. Estimate annual fire count — a schedule that fires
365×/year vs. 12×/year is a ~30× cost and load difference.

## The Five Cron Death-Traps

These are the bugs that pass `crontab -l` validation but break in production.

### 1. Impossible dates — the "never fires" bug

```
0 0 30 2 *
```

**Valid syntax. Never fires.** February has no 30th. This schedule is a dead job
that silently sits forever. The same applies to day 31 in any 30-day month:
`0 0 31 4 *`, `0 0 31 6 *`, `0 0 31 9 *`, `0 0 31 11 *`.

**Fix:** use `0 0 28-31 * *` and check for end-of-month in the script, or use `L`
(last day) syntax if your scheduler supports it.

### 2. OR-semantics — the "fires too often" bug

```
0 0 1,15 * 1
```

**Does NOT mean** "midnight on the 1st and 15th if it's Monday."
**Does mean** "midnight on the 1st, the 15th, **OR** every Monday." That's ~6
fires/month instead of ~2.

This is the single most misunderstood cron rule. When **both** day-of-month AND
day-of-week are restricted (neither is `*`), cron uses OR logic, not AND.

**Fix:** if you need "1st and 15th only if Monday," run daily and check in the
script:

```bash
0 0 * * 1 [ "$(date +\%d)" = "01" -o "$(date +\%d)" = "15" ] && your-command
```

### 3. Midnight spike — the "everything at once" bug

```
0 0 * * *
```

Every job scheduled at `0 0` competes for resources at exactly 00:00. Database
backups, log rotations, cert renewals, report generation — all fire simultaneously.
This causes load spikes, connection-pool exhaustion, and cascading timeouts.

**Fix:** stagger jobs across the hour. Use `17 2 * * *` or `43 3 * * *` instead of
`0 0`. Jitter is your friend.

### 4. Uneven steps — the "drift" bug

```
*/7 * * * *
```

**Does NOT mean** "every 7 minutes evenly." It means "every 7 minutes starting at
0, then resets at 60." So: 0, 7, 14, 21, 28, 35, 42, 49, 56 — then 0 again
(a 4-minute gap). The intervals drift: 7,7,7,7,7,7,7,7,**4**.

**Fix:** 60 is not divisible by 7. Use step values that divide 60 evenly: `*/5`,
`*/10`, `*/15`, `*/20`, `*/30`. If you truly need every-7-minutes, use a supervised monotonic/interval scheduler; a sleep loop adds task duration and needs restart/overlap handling.

### 5. Leap-year February 29 — the "annual surprise"

```
0 0 29 2 *
```

Fires only on leap years — February 29, 2024 / 2028 / 2032… If someone writes this
expecting "end of February," they'll be confused for 3 out of every 4 years.

**Fix:** for last day of February, run on 28 and 29 and use an explicit next-day-is-March guard; running only on the 28th cannot select the 29th.

## Using the validation script

This skill ships a zero-dependency engine at `scripts/cron-engine.js` (Node.js, no
`npm install` needed). You can use it programmatically or from the CLI:

```javascript
// Programmatic — Node.js, zero dependencies
const { describe, validate, nextRuns, formatNextRuns } = require('./scripts/cron-engine.js');

// Parse + describe -> returns { text, error, parsed }
const d = describe('0 0 30 2 *');
console.log(d.text);   // "At 00:00, on day-of-month 30 in in FEB"

// Deep validation -> catches the traps
const result = validate('0 0 30 2 *');
console.log(result.valid);              // true (syntax is valid)
console.log(result.observations);       // includes the "never fires" insight
console.log(result.suggestions);        // e.g. "Midnight is a common spike..."

// Next 5 fire times -> returns Date[]
const runs = nextRuns('0 9 * * 1-5', new Date(), 5);
console.log(formatNextRuns(runs, new Date())); // [{ date, relative, formatted }, ...]
```

```bash
# CLI (via the bundled wrapper)
node scripts/cli.js describe "*/5 * * * *"
node scripts/cli.js validate "0 0 30 2 *"
node scripts/cli.js next "0 9 * * 1-5" 5
```

## Common cron presets

| Expression | Description | Use case |
|-----------|-------------|----------|
| `*/5 * * * *` | Every 5 minutes | Health checks, polling |
| `0 * * * *` | Every hour | Hourly aggregation |
| `0 */2 * * *` | Every 2 hours | Semi-frequent sync |
| `0 9 * * 1-5` | 9am Mon–Fri | Business-hours task |
| `0 2 * * *` | 2am daily | Off-peak batch (avoid midnight) |
| `0 0 * * 0` | Midnight Sunday | Weekly maintenance |
| `0 0 1 * *` | Midnight 1st of month | Monthly report |
| `0 0 1 1 *` | Midnight Jan 1st | Annual task |

## Best Practices

- ✅ Always provide the plain-English description AND run the trap checklist.
- ✅ Stagger midnight jobs to avoid the spike.
- ✅ Prefer step values that divide 60 evenly (`*/5`, `*/15`, `*/30`).
- ✅ Add a comment above every crontab line explaining intent.
- ✅ Set an explicit timezone (`CRON_TZ`) on schedulers that support it.
- ❌ Don't trust `crontab -l` validation — it lists the installed crontab; it is not a syntax or semantic validator.
- ❌ Don't restrict both day-of-month and day-of-week without confirming OR-logic.
- ❌ Don't schedule everything at `0 0`.

## Common Pitfalls

- **Problem:** "My cron job isn't running."
  **Solution:** Check for an impossible date (trap #1) and confirm the daemon is
  running (`service cron status` / `systemctl status crond`). Verify the file
  ends with a newline and has correct ownership.

- **Problem:** "My job runs far more often than expected."
  **Solution:** You hit OR-semantics (trap #2). If both day-of-month and
  day-of-week are set, cron ORs them. Move one to `*` or guard in-script.

- **Problem:** "Intervals are uneven — sometimes 7 min, sometimes 4."
  **Solution:** Step value doesn't divide 60 evenly (trap #4). Use a divisor of 60.

- **Problem:** "My job works locally but not in the cluster."
  **Solution:** Timezone mismatch. Kubernetes uses the controller local timezone unless `spec.timeZone` is set; use `spec.timeZone: Etc/UTC` deliberately. Do not embed TZ/CRON_TZ in its schedule string. Check the exact GitHub Actions schedule support and execution delays.

## Limitations

- This skill targets standard 5-field cron as implemented by Vixie-style cron and Kubernetes `CronJob`, GitHub Actions `schedule`, and most libraries. It
  does **not** validate Quartz 6/7-field expressions with seconds/years, nor
  non-standard `@reboot` / `L` / `#` extensions without a note.
- Estimated annual fire counts assume a non-leap reference year; February 29
  schedules (trap #5) are flagged explicitly.
- This skill does not replace environment-specific validation, testing, or expert
  review. Stop and ask for clarification if required inputs, permissions, or
  safety boundaries are missing.

## Related Skills

- `docker-expert` — when the cron job runs inside a container and the issue is the
  container/entrypoint rather than the schedule.
- `kubernetes-deployment` — when validating a `CronJob` manifest's `spec.schedule`
  field alongside the broader resource config.

## Security & Safety Notes

This skill is read-only and `risk: safe`. The validation script performs no file
writes, network calls, or mutations — it only parses and computes. It accepts data only, but inputs/counts must be bounded and execution should use reviewed local files. Parsing cannot guarantee a job will execute.