In auto mode, Claude Code runs a classifier (a separate check that decides whether an action is safe to run without asking you) on what the model wants to do. That check includes a quick early screen of shell commands. The classifier can now be told to retry that screen or to skip it.
Other changes:
- A new remotely controlled switch covers how the classifier treats MCP tools (extra tools that come from add-on servers). The flag server returned off for this site's account and for an anonymous check.
- Deferred tools, whose full descriptions load only when they are needed, are now replaced only when an MCP server is involved.
This changes how auto mode decides whether a shell command or an MCP tool call can go ahead without your approval.
tengu_parsed_creek 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 v2.1.319: not a boolean we can read
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 v2.1.319. It isn't a statement about your account. What a flag value here can and cannot tell you