Patient Waiting — wait without burning tokens

One incident to remember: a background until-loop watcher kept getting killed, so the main agent “recovered” by rescheduling itself hourly via ScheduleWakeup to cat a status file. The loop had no idle cap and ran for a week — ~140 full-context premium-model turns spent confirming that nothing had changed. Every rule below exists because of that.

Harness selection

The ladder below uses Claude Code background tasks and Monitor. For Codex wall-clock work, use the until skill (delayed-work plugin): capability-detect codex queue (verified on CLI 0.160.1), then arm a detached one-shot timer targeting the exact current thread. It costs no model tokens while waiting and needs a live queue receiver at fire time. Queue acceptance is not delivery proof. If unavailable, report the missing capability and offer a manual nudge or short active-turn wait. This timed path does not establish a general Codex file/event watcher contract. For Codex queued callback payloads, follow the canonical Codex queued sender labels section in $maestro:conduct.

The waiting ladder (Claude Code, in order — never skip down)

1. Background bash until loop (run_in_background). Zero tokens while waiting; one completion notification when the condition flips.

   until <condition-check>; do sleep 3; done; echo "signal"; <print state>
   

2. Monitor tool, only inside its reported deadline. The harness runs your poll script shell-side and invokes the model ONLY when a line is emitted. Zero tokens idle, but Claude Code 2.1.272 still applies a 30-minute cap when it accepts persistent: true; accepted is not the same as honored. Use it only when the event fits inside the returned deadline. Cover failure states too — emit on the watched thing dying, not just succeeding.
3. Both keep dying → STOP and hand the loop to the human. One line: “the watcher can’t survive in this environment — when you’ve done X, nudge me.” Do NOT fall through to model-in-the-loop polling.

Hard rules

• Never machine-poll a human. If the event is human-triggered (review submitted, doc approved, “when I’m ready”, “after my meeting”), the human can close the loop by speaking. A scheduled model wake adds cost and nothing else. Zero-token watcher or nothing.
• ScheduleWakeup loops are for machine-paced state only (CI runs, deploys, data jobs, remote queues) — and only when neither ladder rung 1 nor 2 can see the state. Even then, three backstops are mandatory:
  • Max quiet iterations: 3. Three polls with no change → stop the loop and leave a one-line resume instruction instead of rescheduling.
  • No off-hours polling. If the human is asleep, nothing you’re waiting on them for is happening. Stop; resume on their next message.
  • Count quiet polls in each reschedule’s reason (“quiet poll 2/3”) so drift is visible in telemetry and to the user.
• A killed watcher is a blocker, not a license to escalate. The failure of a cheap mechanism never justifies substituting an expensive one of the same shape. Name the problem, present options (including “do nothing — you nudge me”), and wait.
• Know what a wake costs. A main-agent wake is a full-context turn on the session’s model — possibly the scarcest premium model. A status poll is worth zero of those. If the work on wake is cat file + “no change”, the model should not be in the loop.

Quick decision table