Claude Code v2.1.282: Thinking-Block Fixes, Managed Settings, and a maxProseWidth Toggle

Claude Code v2.1.282 is out, and the changelog is dominated by fixes around extended thinking, session resume, and managed settings parsing. There are a handful of new settings too.

New settings and commands

• maxProseWidth caps the width of Claude's prose in wide terminals — tables and code blocks still render full width.
• A startup notice, plus /status and claude doctor entries, now lists telemetry variables in a project's settings files that were ignored or that turned telemetry off.
• allowClaudeInChromeWithManagedMcp is a new managed setting that lets claude --chrome run alongside an exclusive managed-mcp.json. The error shown when Chrome is blocked now names the setting.
• store.readiness_grace_seconds was added to the Claude apps gateway so /readyz can stay ready through a short Postgres outage, such as a database failover.
• The /feedback drafts list in fullscreen mode got a scrollbar that appears while the mouse is over the list.

Thinking-block and session fixes

• Sessions failing every turn with an Invalid data in redacted_thinking block API error now drop the conversation's thinking blocks and retry once.
• Extended thinking dropped by immediate slash commands (/model, /rename, /artifacts) while Claude was working is preserved.
• Continued or resumed conversations (--continue, --resume) no longer lose earlier extended thinking when relaunched with a --tools list that omits a built-in tool offered earlier.
• Resumed sessions were also re-sending earlier messages in a changed form, which could make the API drop Claude's earlier reasoning — more of those cases are fixed.
• Compaction now retries on a fallback model when the summarization request is refused.

API, auth, and managed-config fixes

• Conversations whose history holds web search results the API cannot decrypt (e.g. from a third-party gateway) no longer fail every request with a 400.
• A failed turn reading Effort 'xhigh' isn't available with thinking turned off after a safety-related model switch is fixed.
• Login errors like another Claude Code process is refreshing it no longer persist for up to a minute after that process was killed mid-refresh.
• Sessions started while another window was refreshing sign-in now retry their organization policy fetch.
• Managed settings with a mistyped value for boolean lock keys such as disableClaudeAiConnectors or allowManagedPermissionRulesOnly now apply the lock and name the key at startup.
• Managed permissions, autoMode, worktree, and attribution blocks no longer get dropped entirely when one nested value is invalid.
• Bash permission rules with a mid-pattern :* now work from every source (settings files included), with a startup warning explaining how they match.

Vertex AI web search now works for newer models Claude Code doesn't recognize yet, and Bedrock/Bedrock Mantle safeguard block messages include request and message IDs. If you run resumed sessions or rely on managed settings, this is the release to pull.