durable-jobs - SKILL.md Agent Skill

name: durable-jobs description: Use when the user asks for durable or recurring background work, work that should survive peer death, retry/cancel/result tracking, spawned-on-demand workers, agent folders, daily briefs, or deciding between repowire jobs and schedule.

Durable Jobs

Use repowire jobs when work needs durable lifecycle state, recovery after peer death, retry/cancel/result inspection, spawned-on-demand execution, or recurrence. Use schedule only for future ask/notify wakeups to an existing target.

Agent folders are a convention, not a registry. Scaffold them with:

repowire agents create <name> --backend <runtime>

Then create jobs targeting the folder's absolute path with --path <abs-path> --backend <runtime>. The folder's AGENTS.md is the source of truth; CLAUDE.md is only a shim for Claude Code. Other supported runtimes load AGENTS.md directly.

--result-surface is metadata only until delivery routing exists. Do not claim jobs send Telegram, email, or dashboard notifications automatically. Workers must update job results explicitly.

Daily email brief

User intent: "Every morning, summarize important email and send me a brief."

Route:

repowire agents create daily-email-brief --backend codex
repowire jobs create "Daily email brief" \
  --path "$(pwd)/.repowire/agents/daily-email-brief" \
  --backend codex \
  --cron "0 8 * * *" \
  --prompt "Prepare today's email brief. Use the job_id and attempt_id from this prompt when updating lifecycle state." \
  --result-surface telegram

Put standing worker guidance in .repowire/agents/daily-email-brief/AGENTS.md: email tool expectations, privacy boundaries, what counts as important, and output format. Keep credentials outside the folder and outside job records.

One-time durable task

repowire jobs create "Inspect billing webhook compatibility" \
  --path /path/to/billing-repo \
  --backend codex \
  --prompt "Inspect webhook compatibility risk and update this job with a concise result."
repowire jobs run <job_id>

Use a job instead of notify_peer when the result must survive peer death, be retried, or be inspected later with repowire jobs result.

Wake-up reminder

repowire schedule self 30m "Check whether the release peer replied."

Use schedule, not jobs, when this is just a future message to an existing session.