Unclear It is not clear how far the extra saving step during compaction reaches.
What
The Agent SDK lets programs run Claude Code and read back what happened in each turn, a turn being one request and Claude's response. Previously the turn's result text was always taken from Claude's last message. Now, if that message is an API error but the turn itself did not end in error, it is not used as result.
Two smaller changes come with it. When a conversation is compacted, meaning shortened so it fits, Claude Code now saves state first. Notices that a command has been taken in now go through their own callback.
Why
Programs reading the result of a successful turn no longer get an error message where they expect Claude's answer.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
It is not clear how far the extra saving step during compaction reaches.