You'll noticeTier: how much it should matter to you
3Useful: my rating, 1 to 5
3Signal: worth watching, 1 to 5
Extended ThinkingArea: what it touches
ImprovementsKind: in v2.1.296,
ImprovementsSection of the release
What
Every reply has a maximum output length, counted in tokens (small chunks of text). With extended thinking, Claude reasons before answering, and that reasoning can use up the whole limit. A new recovery feature, query_output_limit_recovery, handles this case:
Detection: it spots a max_tokens stop where everything produced was thinking.
Nudge: it picks a different nudge, the follow-up instruction that tells Claude to carry on, for that case.
Message: when recovery runs out of attempts, the message adds "The last attempt reached the limit while still thinking." It then suggests lowering the effort level, naming the command to use when you are in an interactive session.
Outcomes: it records whether the thinking was discarded and nudged, resumed, exhausted, or whether continuing failed.
Why
If Claude runs out of room mid-thought, you are told so plainly and given something to try, rather than getting a reply that just stops.
