Routines

The Routines view is the operating surface for scheduled automation: every routine, its schedule, its run history, and the scheduler daemon that drives it all. A routine is a Markdown specification; each due run spawns a fresh agent process that reads the current spec, does the job, writes its output, and exits.

Open it from the sidebar: Routines. The view has three tabs — Routines, Execution Log, and Claims.

The scheduler bar

Across the top sits the daemon status bar: whether the scheduler is online, its uptime, and how many routines are currently scheduled (n/m scheduled). Controls: Start, Stop, and an autostart toggle so the scheduler comes up with the app.

The one fact to internalize: routines only run while the scheduler runs. If the daemon is stopped, nothing fires — which is also the bluntest way to halt all automation at once.

The Routines tab

Routines are listed in rows, grouped into collapsible categories. Each row carries:

• The schedule switch and interval — enable or disable the routine, and set its cadence inline. When you enable a routine, a one-off warmup run fires about a minute later, so a freshly enabled routine proves itself immediately instead of waiting for its first slot.
• Run now — trigger an immediate run, independent of the schedule. It works even when an instance is already running, so use it deliberately.
• Edit — opens the routine editor (below).
• View log — opens the latest run's log inline, below the table, with a copy button.
• A favorite star — starred routines surface on the Dashboard for one-click access.

Schedules

Three schedule forms cover everything:

• Interval presets — 15min, 30min, 1h, 2h, 4h, 12h, 24h
• Daily at a time — daily:09:00
• Full cron — a standard five-field cron expression, for anything irregular

Set them in the editor's schedule builder or type them directly.

Edit or create a routine

The in-app editor covers the whole specification without touching a file: name, model (the per-routine cost/quality dial), description, timeout, the schedule builder, the full instruction body, and editable configuration rows. Saving writes the routine's Markdown file, and the daemon picks the change up on its own — the next run uses the new spec, nothing to restart.

Because each run is a fresh process, improving a routine is a tight loop: read a log, sharpen the instruction body, save, and the next run behaves accordingly.

How runs behave

• Tracked jobs. Every run is a job with a start time, status, and a log written to documents/routine_logs/.
• Retries with backoff. A failed run retries; a routine that keeps failing disables itself rather than looping — automation here degrades loudly, not silently.
• Bounded concurrency. The scheduler caps concurrent routine jobs (eight at once), so a crowded schedule cannot swamp your machine.

The Execution Log tab is the run history across all routines, with live output for in-flight jobs — your first stop when checking what automation did overnight. The Claims tab shows the work-item claims runs have taken: routines that process discrete items (news, drafts) claim each item as they handle it, which is how repeated runs avoid processing the same thing twice.

Cost and authority — the honest part

Two properties of routines deserve to be stated plainly, because they are easy to forget once a schedule is humming:

• Routines spend real Claude usage. Every run is a fresh Claude session on your own subscription, and an unattended schedule spends while you are away. Cadence is a cost dial: a routine at 15min runs 96 times a day. Start conservative, read the logs, and widen the cadence only after a routine has earned trust.
• Routines run unattended, with tool prompts auto-approved. There is no one at the keyboard to confirm each step, so a routine body executes with real machine authority. Write routine instructions as deliberately as you would a script running as you — scoped, explicit, and no broader than the job needs.

The standing safety gate is unchanged by scheduling: anything outbound — an email, a post, a newsletter — that a routine drafts lands in a review queue for your approval. Automation drafts; you send.

Routine, agent, or mission?

The quick router, worth re-checking before you automate anything:

Where to go next

• Routines & automation — the concept and where you stay in control
• Routines infrastructure — the routine catalog and scheduler reference
• Agents — the roster routines draw their postures from