Automate recurring work

Three ready-to-run recipes for Schedules — recurring, autonomous jobs on a cron cadence. Each gives you a prompt to copy, how to set it up, and what happens. For what a Schedule is — the pages, run history, limits — see the Schedules concept.

Nightly backlog sweep

Every night, turn “ready” tickets into draft PRs while you sleep — the recurring counterpart of the Runs backlog sweep.

The prompt — paste this into the schedule’s prompt field:

For each open issue labeled `ready-for-dev` that doesn't already have an open
pull request, implement it in its own DevBox, run the test suite, and open a
draft PR. Skip anything you can't finish cleanly.

Set it up

1. Open the Schedules page → + Schedule, and paste the prompt above.

2. Set the cadence to 0 2 * * * (daily at 2am) — the preview confirms the next run time.

3. Pick the engine (e.g. Codex), a target worker, and an image, then Save.

How it works — it appears in the list enabled, next run counting down. Each night, autonomously and without asking you anything, it works every ready ticket end-to-end: a DevBox per ticket, test suite, draft PR, then torn down. Test it immediately with Run now, and open its run history to watch any execution’s timeline. Disable it any time without losing the configuration.

Weekly dependency audit

A different cadence and a different job — every Monday morning, find risky dependencies and open one PR with the safe upgrades, so security patches never pile up.

The prompt — paste this into the schedule’s prompt field:

Audit this repo's dependencies for known vulnerabilities and outdated major
versions. On a new branch, apply only the safe, non-breaking upgrades, run the
full test suite, and open one pull request summarizing what you changed and what
you skipped (and why). Never push anything that breaks the build.

Set it up

1. + Schedule, and paste the prompt above.

2. Set the cadence to 0 7 * * 1 (Mondays at 7am) — a weekly rhythm, unlike the nightly sweep above. The preview confirms the next Monday.

3. Pick the engine — Claude Code is a good fit (it reasons well over changelogs and breaking-change notes) — plus a target worker and image, then Save.

How it works — Run now once to see it end-to-end without waiting for Monday. On the timeline you watch it scan the dependency tree, open a branch, apply the safe bumps, run the suite, and call gh pr create — finishing with a single audit PR you can review at your leisure.

Nightly flaky-test sweep

Wake up to a triaged list of flaky specs instead of chasing intermittent failures by hand. This one files issues and changes no code.

The prompt — paste this into the schedule’s prompt field:

Run the end-to-end test suite three times. Identify any specs that pass and fail
non-deterministically across the three runs. For each flaky spec, open an issue
with the failing output, which run it failed in, and a link to the spec. Don't
change any code.

Set it up

1. + Schedule, and paste the prompt above.

2. Set the cadence to 0 3 * * * (nightly at 3am).

3. Pick the engine (e.g. Codex), a target worker, and an image, then Save.

How it works — every night it runs the suite three times, diffs the pass/fail results across runs to isolate the non-deterministic specs, and files one issue per flaky spec with the trace — no code changes, just triage waiting for you in the morning.

Tip

Mix cadences and engines: a nightly backlog sweep clears tickets, a weekly audit keeps dependencies healthy, a nightly flaky-test sweep keeps the suite trustworthy. Each is one Schedule.

Next

• Schedules — how Schedules work (pages, run history, limits).
• Run an agent on demand — the same ideas, fired once.
• Engines — pick the right engine per job.