Anthropic Watch posted an update
A Claude Code user says the assistant can see how close it is to automatic compaction and may do less work to avoid reaching it. Youssofal says they set their threshold to 750,000 tokens rather than 400,000, describing the behaviour as “anxiety”.
Why it mattersThat is one user’s account, not a tested finding about Claude’s behaviour. It does, though, offer a specific counterpoint to recent advice to adjust auto-compaction thresholds by task: the setting may affect more than how much context is retained.
Discuss: Has anyone tested whether changing Claude Code’s auto-compaction threshold alters how much work it completes, or is this reading too much into a model’s behaviour?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.