cron: scheduled prompts in Loom

At each match of a cron expression, Loom submits a fixed prompt to an existing session. The reply lands in that session like any other turn, marked ⏰ cron. Jobs live in ~/.loom/cron.json; never edit it — go through Loom.Cron over bin/loom rpc, which needs no running session of yours and no HTTP.

Calling it

B="$(dirname "$LOOM_ENV_FILE")/_build/prod/rel/loom/bin/loom"
$B rpc "$(cat <<'EOF'
IO.inspect(Loom.Cron.add(%{name: "Morning news", cron: "0 1 * * *", target: "a_xxx", prompt: ~S"""
The prompt, verbatim: quotes, What's, newlines and #{} all stay as written.
"""}))
EOF
)"
$B rpc 'IO.inspect(Loom.Cron.list())'
$B rpc 'IO.inspect(Loom.Cron.delete("c_xxx"))'

• Wrap every call in IO.inspect: rpc prints nothing otherwise.
• The quoted heredoc plus ~S""" keeps the prompt literal; it only must not contain """.
• add returns {:ok, job} with next_at (first fire, UTC), or {:error, reason}.

Fields

• cron: five fields (minute hour day month weekday), always UTC. *, a,b, a-b, */n, a-b/n, and names like MON, JAN. No @daily, no seconds. With both day and weekday set, either one matching fires.
• target: the id of an existing session. Find it by title: $B rpc 'IO.inspect(Loom.Agent.list_persisted() |> Enum.map(&{&1[:id], &1[:title]}))'. If it is unclear which session ("this one", "here"), confirm the title with the user.
• name: optional; defaults to the prompt's first line.
• prompt: what the target sees when it fires, with none of this conversation around it — write it to stand on its own.

Time zones

The user speaks local time; cron is UTC. Convert it yourself (Beijing = UTC+8, no DST: 9:00 there is 0 1 * * *, and a weekday can shift across midnight). If the zone is unclear, ask. After adding, tell the user the next fire time back in their own zone so they can check it.

Behaviour

• A fire missed while Loom was down is skipped, not caught up.
• A busy target gets the prompt like a typed message: steered into the running turn, or queued.
• Deleting the target session leaves the job; it records last_error until deleted.
• /cron typed in any session lists every job.

Waiting on long work

For work that runs for hours outside this conversation (a benchmark, a batch job, a remote build), don't rely on background-task notifications: they die with the backend process, which Loom releases when idle or restarts. Add a job targeting your own session (Loom gives you its id) whose prompt says it is a cron check-in, how to check progress, what to do when done (report, clean up), and to delete this job (give its id) once finished. Keep a hard failsafe on the resources themselves too, e.g. shutdown -h +N on a temporary instance.