A reply can stop early when it reaches the most text the model may write in one go. Claude Code then nudges the model to continue. The choice of nudge and the point to stop retrying used to come from settings sent by a server. Those are gone, and one fixed nudge is used.
A second, remotely switched prompt also exists. It applies when the cut-off reply ended in the middle of text, in the main conversation rather than a helper task. The flag server returned off for this site's account and for the anonymous baseline. The error shown when retries run out no longer has a separate case for stopping early.
Cut-off replies are continued in a different way, and the remote setting that tuned this has been removed.
tengu_rosy_pebble Off in both readingsThe flag server returned off for the account this site reads and for the anonymous baseline. A reading of off cannot rule out a rollout these two readings sit outside of.
This account: off · anonymous baseline: off · compiled default in v2.1.314: off
These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.
Read once, for one account on one subscription tier, against v2.1.314. It isn't a statement about your account. What a flag value here can and cannot tell you
New in this build: tengu_rosy_pebble