You'll noticeTier: how much it should matter to you
2Useful: my rating, 1 to 5
1Signal: worth watching, 1 to 5
SDKArea: what it touches
ImprovementsKind: in v2.1.311,
ImprovementsSection of the release
What
Programs that read Claude Code's output as a live stream through the SDK (the toolkit for driving Claude Code from your own code) get a message_stop event at the end of each reply. When Claude Code produces that event itself, it can name the position, from_block_index, from which content blocks were left unfinished.
That used to cover a stream from the API that failed, stalled or ended early. It now also covers a turn that was interrupted, for example when you stop Claude partway through a reply.
Why
A program reading the stream can now remove half-finished pieces of a reply after an interruption, the same way it already can after a stream failure.
