Retry after a safety-classifier refusal is now on by default, with CLAUDE_CODE_DISABLE_REFUSAL_RETRY to turn it off

What

The recovery Claude Code runs when a safety classifier stops a response no longer needs to be switched on. A safety classifier is an automated check that can halt a reply partway through. Until this release the recovery ran only if you set the environment variable CLAUDE_CODE_OCHRE_KITE, or if the server-side flag tengu_ochre_kite was on, and that flag counted as off when the server gave no value. An environment variable is a named setting you give a program when you start it. Both the variable and the flag are removed. Now:

• It retries once by default: when safeguards stop a response in the main conversation, Claude Code shows "safeguards stopped the response above · continuing once with that noted" and tries once more on the same model.
• Tool calls in the stopped response still get a recorded result, such as "Not run: the response that made this tool call was stopped by a safety classifier." A tool call is an action Claude asked to take, like running a command or editing a file.
• CLAUDE_CODE_DISABLE_REFUSAL_RETRY is a new environment variable. When it is set, the retry does not run.
• Some requests are still excluded: the retry does not run for background helper requests (those classed as "auxiliary"), or when the caller passes an internal argument that skips it.

Why

A refusal no longer simply ends the turn for everyone. The model gets one more attempt, with a note that a safeguard was triggered. If you would rather a stopped response end the turn as before, set CLAUDE_CODE_DISABLE_REFUSAL_RETRY.