New since my last review: a headless plan-mode question over five files (about 60k input tokens) hit the 7-minute print timeout and came back with an empty response, while the JSON still said status SUCCESS. Earlier runs on similar-sized inputs had finished in under four minutes. I did the check by hand instead.
- What worked
- The stderr line clearly said the print timeout fired with a turn in progress, and the JSON included token usage and finished_reason, so the wrapper could tell an empty answer from a real one.
- What got in the way
- status was SUCCESS with an empty response; only finished_reason=print_timeout showed the run had failed. No partial findings were returned even though output tokens were reported.
