Unclear It is unclear whether Claude Code's own requests changed in this version or only the guide was updated.
What
Claude Code includes documentation for people running a sign-in server or an LLM gateway, a server that sits between Claude Code and the model. That text changed:
- Servers must ignore request parameters, body fields and headers they do not recognize ("Don't reject what you don't recognize."). They should pass body fields and
anthropic-betavalues through unchanged. - In auto mode, the client adds an
anthropic-betavalue and a top-levelsafeguardsfield to requests, and readssafeguard_resultsfrom responses. A gateway that does not support this should answer with a 400 error namingsafeguards. - The managed settings response is now wrapped as
{uuid, checksum, settings}, with a sha256 checksum used forIf-None-Matchand 304 (not modified) replies. Before, the documentation described returning the baremanaged-settings.jsonwith an ETag.
Why
If you run a gateway, requests can carry fields your server has not seen before and should not be rejected. The findings suggest a managed settings response in the old shape may be treated as a failed fetch.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
It is unclear whether Claude Code's own requests changed in this version or only the guide was updated.