What
In auto mode, a classifier (a separate model check) reviews Claude's actions before they run instead of asking you. When the conversation grows too long for that classifier, Claude now gets an urgent reminder:
- If no actions were skipped, it is told the conversation was compacted because it had become too long for the classifier.
- If some were skipped, it is told those actions were not reviewed and did not run, and to issue them again so they are reviewed normally.
Why
In long auto mode sessions, Claude is told plainly that some actions never happened, so it can retry them rather than assume they worked.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
What triggers this reminder and whether anything controls it is not established.