Bumping this with a cross-reference, since it’s now at 10 likes and several people in-thread are hitting the same wall.
The two changes @douglascaetano and @kevinn already identified remain the right fix: (1) make the timeout tunable, including fully off, and (2) report a timeout to the model as a timeout, not as a skip.
Worth linking the confirmed-bug side of point (2): Questionnaire timeout falsely treated as “skipped”. The problem there is that the timeout return string is byte-identical to a real user skip, so the agent treats “ran out of time” as “user consented to defaults” — which on a high-impact prompt is a genuine safety hazard, not just an annoyance. A distinct return string plus a user-configurable (or disable-able) timeout would resolve both threads together.