Jobs

Jobs

Durable scheduled work lives in ~/.physiclaw/jobs/jobs.md. The format is regex-parsed by the cron loop — a malformed field breaks every scheduled job in the file. Never edit jobs.md directly.

All job operations go through a single CLI:

uv run python "$SKILL_DIR/jobs.py" <subcommand> [args]

Four subcommands: create, list, get, finish. Each is a thin wrapper around the canonical engine module — the file format stays invariant no matter what you pass.

$SKILL_DIR is provided by Claude Code; it resolves to the directory containing this SKILL.md.

Create a job

uv run python "$SKILL_DIR/jobs.py" create \
  --id <user>-<topic>-<YYYY-MM-DD> \
  --kind <one-time|periodic> \
  --schedule "<5-field cron>" \
  --description "<human text>" \
  --context "<what the next wake needs to know>"

• --kind defaults to one-time if omitted.
• --schedule: standard 5-field cron min hour dom mon dow. Day- of-week is 0=Sun, 6=Sat. All times are naive local time. Examples:
  • Every weekday at 09:00 → "0 9 * * 1-5"
  • At 14:30 on 2026-05-01 → "30 14 1 5 *" (combined with a matching one-time kind)
  • In 30 minutes (compute wall-clock now + 30m yourself): "M H D Mo *" where M/H/D/Mo are the computed minute/hour/day/month.
• --context: at least 10 characters. Write what the next wake needs to do — the engine injects this into the SYSTEM prompt when the job fires.

Id format: <user>-<topic>-<YYYY-MM-DD>. Lowercase letters, digits, hyphens only. The date suffix keeps repeat ids (alice-water-plants-…) unique across days without -v2 churn.

On duplicate id (even terminal), the command errors — pick a fresh id.

Example

uv run python "$SKILL_DIR/jobs.py" create \
  --id alice-water-plants-2026-05-01 \
  --schedule "0 9 1 5 *" \
  --description "Remind Alice to water the kitchen plants." \
  --context "Send Alice a friendly reminder via WeChat to water the kitchen herbs."

List jobs

uv run python "$SKILL_DIR/jobs.py" list [--status pend|fired|done|fail|cancel|all]

Default is all. Prints one job per line with id, status, kind, schedule, next fire, and short description.

Common recovery move: start a wake with jobs.py list --status fired to find orphaned fired jobs from earlier wakes that you need to close.

Get one job

uv run python "$SKILL_DIR/jobs.py" get --id <id>

Prints every field (description, type, status, schedule, context, all timestamps, execution result). Use when the one-liner from list isn't enough.

Finish a fired job

uv run python "$SKILL_DIR/jobs.py" finish \
  --id <id> --status <done|fail|cancel> --recap "<one-line outcome>"

• done — work complete.
• fail — blocked or impossible.
• cancel — no longer needed, or you're rescheduling (do cancel + new create with a fresh id).

For periodic jobs, done/fail automatically reset to pend so the next cycle fires. Only cancel is permanent for periodic.

Errors on already-terminal jobs (re-finishing is a bug, not idempotent).

Reschedule = cancel + create

There is no update. Jobs are append-only; their only state change is finishing them. To reschedule:

1. jobs.py finish --id <old> --status cancel --recap "rescheduling — new id <new>"
2. jobs.py create --id <new> … (use a fresh id — bump the date, or append -v2 if same-day)

WAIT close pattern

When closing with >> WAIT - <reason>, pair it with a fresh job for the resume check. Otherwise the engine auto-schedules a generic 15-min follow-up — usually wrong for your situation. Only fall back to that if you genuinely don't know when to check back.

Example:

uv run python "$SKILL_DIR/jobs.py" create \
  --id alice-confirm-order-2026-05-01 \
  --schedule "$(date -v +30M '+%M %H %d %m') *" \
  --description "Re-check IM: user said they'd confirm in ~30min." \
  --context "Alice reviewed cart before leaving for a meeting; re-open WeChat and look for her OK to place the order."