claude -p: ScheduleWakeup / Cron* tools advertised in print mode where they have no execution semantics

In non-interactive claude -p mode, the agent receives the schedule built-in skill's tools (ScheduleWakeup, CronCreate, CronList, CronDelete). These tools are designed for interactive sessions where a long-lived Claude Code runtime can re-fire the session at a scheduled time. In -p the process exits after one turn — no runtime exists to honor a wakeup.

When the model invokes one of these tools (a natural choice when it wants to "wait N minutes and re-check" something), the turn ends. There is no error, no warning, and no documented contract that the wakeup will silently no-op. The model believes it has scheduled a resumption that will never come.

Repro shape

An agent driving a long-running background task — e.g. a multi-hour build started with Bash (run_in_background: true) and polled via BashOutput / Monitor — has a legitimate reason to want to back off between polls. ScheduleWakeup is advertised, semantically perfect for the use case, and silently broken. The agent uses it, the turn ends, the orchestrator sees an empty/idle final response, and the multi-hour run is lost.

Expected

Either:

1. The schedule skill is not loaded when the CLI is invoked with -p, or
2. Calling these tools in -p returns an explicit error like ScheduleWakeup is not supported in non-interactive mode; use a polling pattern instead.

Either fix would make the failure mode visible and recoverable.

Related

#53746 (proposes a disabledBuiltinSkills setting). That would let users opt out, but the print-mode case is stronger — the tools are inert by construction with no runtime to honor them, so unregistering them in -p doesn't need a settings flag.