[FEATURE] AskUserQuestion is not available as a tool in routine (scheduled/triggered) cloud sessions

Summary

AskUserQuestion is not available as a tool in a routine (scheduled/API/GitHub-triggered
cloud session) at all — it's absent from the tool list entirely, not merely refused or
auto-resolved. A routine that needs to ask a real question can only stop and report that
it's blocked; it has no way to pause and wait for an answer.

Reproduction

1. Create a routine with a GitHub-event trigger (issue labeled).
2. Give it a task whose first instruction is to call AskUserQuestion before doing

anything else.

1. Fire the trigger.

Observed: the run stops immediately with a message stating AskUserQuestion is not
an available tool in this session — confirmed absent from the tool list, and a
ToolSearch-style lookup for it finds nothing. No file created, nothing committed, no PR
opened. The run correctly declines to guess rather than fabricating an answer, which is
the right failure mode — but there's no mechanism to actually ask and wait.

Expected / requested: a routine should be able to call AskUserQuestion, pause the
run, and resume from an answer delivered later (e.g. via a reply from the mobile app or
web session), the same way an ad hoc cloud session (claude --cloud) already does when
it asks a question and "sits idle" until answered.

Why this matters

The routines docs state runs happen with "no approval prompts during a run," which reads
as being about tool-permission prompts — a different thing from a conversational,
ask-the-user tool. The actual behavior confirms the tool is unavailable either way, which
means every routine prompt has to be written as a complete spec with all uncertainty
resolved in advance, since there's no fallback to ask. That's a real constraint worth
designing around, but it's also a gap: routines are the one surface that's both unattended
and doesn't require a machine to stay on, and giving them a way to pause-and-resume
(rather than only "complete or give up") would make a much wider range of tasks safe to
hand to one.

Full write-up with screenshots

I documented two shakedown tests (a working mechanical one, and this one) in detail,
including the exact repro and the run's own output:
https://oharu121.com/blog/claude-code-routines-github-issue-pr-sandbox-askuserquestion/