Apps and editors can tell whether typing ultrathink raises effort, and settings survive brief read failures

What

Typing ultrathink in a message can make Claude think harder for that turn. Programs that host Claude Code, such as editor extensions and apps built on the Agent SDK, can now find out whether that works:

• The session start message gains an optional ultrathinkAvailable field. It is true when typing ultrathink runs the turn at xhigh effort, and also when the session is already at xhigh or max, where the turn stays as it is.
• Hosts are told to highlight the word only while the field is true. When it is absent, the host is talking to an older Claude Code.
• The get_settings response also gains ultrathinkAvailable, and the status report gains ultrathink_available when it is known.

Settings loading also changed:

• If a settings file cannot be read but was read successfully before, Claude Code now returns an unreadable_for_now result with the last values it read. Before, any unreadable file gave a plain unreadable result.
• Certain short-lived read errors no longer throw away the saved copy of the settings.

Smaller additions appear alongside: sessionBeganAsCopy in a schema, ccr_cloud_browser_turn_id in a list, a repoHooksListing entry, voice footer hint counters, and a substitutionHeredocs option in the shell parser.

Why

Editor extensions and other hosts can highlight ultrathink the way the terminal does, and only when it will actually raise effort. A settings file that fails to read for a moment no longer loses your settings for that session.