Anthropic published a postmortem yesterday covering three separate changes that degraded Claude Code output between 4 March and 20 April. Two were deliberate product decisions. One was an actual bug. The postmortem treats all three as the same category of problem, and that framing is doing a lot of work. The bug is the interesting part only because it made the other two visible. On 26 March, Anthropic shipped a change that cleared older thinking blocks from sessions idle for more than an hour, to avoid a full cache miss when users resumed. The implementation cleared thinking every turn for the rest of the session instead of once. That is a straightforward regression, it took over a week to find, and they fixed it on 10 April. Fine. The other two are not bugs. On 4 March, Anthropic silently changed Claude Code’s default reasoning effort from high to medium because some users were seeing long enough latency on high that the UI looked frozen. On 16 April, they added a system prompt…
No comments yet. Log in to reply on the Fediverse. Comments will appear here.