Claude Code now contains a new set of rules for deciding what to do when a call to the model fails. They cover:
- responses that are refused, unreachable or unauthorized.
- responses that stop partway through.
- 529 overloaded errors and their limits.
- falling back to the next model.
- retrying without streaming.
- a cap on retries for each cause.
These rules run beside the existing retry logic and compare what they would have done with what Claude Code actually did. Mismatches are reported only when a setting is switched on, and that setting is off unless it is switched on remotely.
This points to a coming rewrite of how Claude Code retries failed or overloaded calls.
tengu_attempt_decision_shadow Not enough to sayNothing here resolved what this flag was doing on this version, so nothing here should be read as on or off.
This account: no value returned · anonymous baseline: no value returned · compiled default in v2.1.284: 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 v2.1.284. It isn't a statement about your account. What a flag value here can and cannot tell you
It is not confirmed whether any part of Claude Code acts on the new rules' decisions or whether they only observe.
New in this build: tengu_attempt_decision_shadow