Perpetual's on_complete handler always returns True after rescheduling, which tells the Worker "I handled it" — so the Worker skips its normal error logging. When a perpetual task fails (e.g. a dependency raises during __aenter__), the exception is completely invisible: no error log, no traceback, just a fast completion time and a quiet reschedule.

I hit this when CurrentInferenceBackend started raising a KeyError after switching inference backends. The backup task was silently failing every 5 minutes with no indication anything was wrong — just ↪ [5ms] backup(){backup} followed by ↫ [7ms] backup(){backup}.

The on_complete method should probably log the exception from outcome when there is one, even if it still reschedules. Something like:

if outcome.exception:
    logger.error(
        "↪ [%s] %s failed: %s",
        format_duration(outcome.duration.total_seconds()),
        execution.call_repr(),
        outcome.exception,
        exc_info=outcome.exception,
    )

The reschedule-on-failure behavior is fine (you want perpetual tasks to keep trying), but the silent part makes debugging really painful.