Anthropic

What's changed

• Bug fixes and reliability improvements

What's changed

• Bug fixes and reliability improvements

Fixed a Bedrock streaming bug behind proxies that stripped response Content-Type headers, silently doubling billed API calls, plus a startup hang with SSO profiles behind HTTPS proxies. Added /claude-api upgrade to migrate Python projects from anthropic 0.x to 1.x, a one-time fullscreen renderer offer on Bedrock/Vertex/Foundry setups, and cost estimates that now include the 1.1× US-only inference premium for data-residency workspaces. Another four dozen fixes span vim mode, fullscreen rendering, remote sessions, and /resume behavior.

Fixed unbounded memory growth in long interactive sessions by releasing subagent tool results once they leave the recent display window, and added a keybindingFlavor setting with a "readline" option for Bash-style Ctrl+W. Also fixed output styles drifting back to default voice mid-session and dozens of Remote Control, MCP, and cross-session messaging fixes.

Fixed prompt caching for sessions using an LLM gateway or custom base URL. Added a built-in "Concise" output style that leads with results and skips preamble, selectable under Output style in /config.

New ANTHROPIC_DEFAULT_MODEL env var sets a session's starting model while preserving /model overrides, and cross-session SendMessage gains one-shot idle notifications. Also fixed sandbox read-deny wildcards being bypassed via renaming, plus clipboard/renderer/crash regressions across fullscreen, /model picker, and remote control.

Fixed Shift+Tab in permission comment fields no longer approving edits prematurely, and fixed whole-prompt-cache invalidation on language server reconnect. Also improved memory/CPU use for background cloud sessions, hardened permission dialogs, and added an optional spellcheck setting.

Remote file reads, session restore, CLAUDE.md includes, and workflow scripts now reject Windows NT-namespace paths, fixing the remaining pre-approval file-access NTLM credential-leak vector. Sessions now resume automatically when a claude.ai usage limit resets, with a /config toggle to disable; 30+ other fixes cover shell-command queueing, permission previews, and Remote Control.

Todo and task-tracking tools are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models — set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to bring them back. Also patched an NTLM credential-leak vector involving Windows NT device-prefix paths, fixed a 2.1.232 auto-mode approval regression, and reverted the 2.1.232 Bash permission changes for Cygwin symlinks and input redirections. Plus improved worktree GitLab MR support, memory limits for Bash, and faster self-hosted runner startups.

Subagent forking is now on by default: subagent_type: "fork" inherits the full conversation and prompt cache, and non-teammate agent spawns in interactive sessions run in the background by default. Also fixed PowerShell and Git Bash permission bypasses that could overwrite settings and redirect file access, plus hardening around nested git trust and cross-session messaging sockets.

Fixed MCP OAuth sign-in failing with a redirect URI mismatch for servers that use a pre-registered OAuth client, such as Slack.

Fixed long responses being printed twice and vanishing mid-stream, plus crashes from non-string tool arguments, narrow terminal rendering, and extended-length Windows paths. Added server-supplied hooks for self-hosted runners, SSE keepalive pings for gateway streams, and plugin marketplace command sources.

Interactive sessions no longer stop redrawing after a rare layout error, and /tui no longer reverts to an earlier model after /model changed. Self-hosted runners skip repos whose checkout hook fails, session cleanup no longer deletes project memory folders, and synced skills no longer shadow local commands.