New Features:
managed-settings.d/ drop-in directory for deploying independent policy fragmentsCwdChanged and FileChanged hook events for reactive environment managementsandbox.failIfUnavailable setting to exit with error when sandbox unavailabledisableDeepLinkRegistration setting to prevent claude-cli:// protocol handler registrationCLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 to strip credentials from subprocess environments/ key and n/N navigation in transcript modeCtrl+X Ctrl+E as alias for external editor[Image #N] chip at cursor for positional referenceinitialPrompt in frontmatter to auto-submit first turnchat:killAgents and chat:fastMode rebindable via keybindings.jsonFixes:
--mcp-config bypassing managed policy enforcement--print modecaffeinate process not terminating on exit!-prefixed suggestions/config menu showing both search and list selection--channels showing unavailable message on first launchCtrl+B interfering with readline backward-charImprovements:
--resume memory usage and startup latencyWebFetch to identify as Claude-UserWebFetch peak memory usageclaude -p startup with unauthenticated MCP servers (~600ms)/status now works while Claude respondsXDG_DATA_HOME for protocol handler registrationCtrl+F to Ctrl+X Ctrl+KTaskOutput tool in favor of ReadCLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK env varCtrl+L clears screen and forces full redraw--bare -p pattern ~14% fasterMEMORY.md truncates at 25KB as well as 200 linesAskUserQuestion and plan-mode tools with --channelsAdditional Fixes:
rg ... | wc -l hanging in sandbox mode on Linux--worktree hanging with forward slash in nameNew Features:
--bare flag for scripted -p calls (skips hooks, LSP, plugin sync; requires ANTHROPIC_API_KEY or apiKeyHelper)--channels permission relay for channel servers to forward tool approval promptsFixes:
CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS not suppressing structured-outputs beta header--channels bypass for Team/Enterprise orgs/btw not including pasted text during active response/rename not syncing title for Remote Control sessions/exit not reliably archiving sessionImprovements:
! bash mode discoverabilityFixes (continued):
New Features:
rate_limits field to statusline scripts for displaying Claude.ai rate limit usagesource: 'settings' plugin marketplace source for inline plugin declarationseffort frontmatter support for skills and slash commands--channels (research preview) for MCP servers to push messagesFixes:
--resume dropping parallel tool results/remote-control appearing where it cannot function/sandbox tab switching not responding to Tab or arrow keysremote-settings.json cachedImprovements:
@ file autocomplete in large repositories/effort display/permissions tab switchingNo details provided in changelog entry.
Increased default output token limits for Claude Opus 4.6 to 64k tokens and upper bound to 128k tokens. Added allowRead sandbox filesystem setting and optional index parameter to /copy command. Fixed permission rules saving incorrectly, auto-updater memory accumulation, conversation history truncation with --resume, PreToolUse hooks bypassing permission rules, line ending conversions, memory growth in long sessions, cost tracking in fallback mode, beta tool-schema stripping, Bash errors with spaces in temp paths, paste handling, and numerous UI/terminal issues. Fixed IDE integration in tmux/screen, CJK character rendering, teammate panes not closing, and iTerm2 auto mode detection. Improved startup and --resume performance (45% faster, 100-150MB less memory), Esc abort functionality, and plugin validation. Renamed /fork to /branch. VSCode: Improved plan preview titles and added macOptionClickForcesSelection setting reference.
Added StopFailure hook event for API errors, ${CLAUDE_PLUGIN_DATA} variable for plugin persistent state, and effort, maxTurns, disallowedTools frontmatter for plugin-shipped agents. Terminal notifications now reach outer terminal in tmux, and response text streams line-by-line. Fixed git operations in sandboxed Bash, conversation history truncation in large sessions with subagents, infinite loop in stop hooks, MCP server permission rules not blocking tools, sandbox filesystem permissions, protected directories writability, voice mode keybindings, --worktree flag not loading skills/hooks, git instructions suppression, Bash PATH issues, terminal color support, and various other fixes. Added ANTHROPIC_CUSTOM_MODEL_OPTION env var and fixed ANTHROPIC_BETAS being ignored with Haiku models. Improved memory usage and startup time for resuming large sessions. VSCode: Fixed login screen flash and rate limit error in model dropdown.
Added MCP elicitation support for structured input mid-task, Elicitation and ElicitationResult hooks, -n/--name CLI flag for session naming, worktree.sparsePaths for git sparse-checkout in large monorepos, PostCompact hook, /effort slash command, and session quality survey. Fixed deferred tools losing input schemas after compaction, slash commands showing 'Unknown skill', plan mode re-approval issues, voice mode keypresses, Bash permission rules with quoted arguments, MCP reconnection spinner, LSP plugin registration, clipboard copying in tmux over SSH, transcript auto-scrolling, and several Remote Control issues. Improved --worktree startup performance, background agent behavior, model fallback notifications, blockquote readability, and Remote Control session titles. Updated --plugin-dir to accept repeated flags. VSCode: Fixed gitignore patterns with commas.
Added 1M context window for Opus 4.6 by default for Max, Team, and Enterprise plans, /color command to set prompt-bar colors, session name display on prompt bar, and last-modified timestamps to memory files. Fixed voice mode activation on fresh installs, Claude Code header not updating model name, session crashes with undefined attachments, Bash tool ! character handling in pipes, managed-disabled plugins appearing in plugin list, token estimation over-counting, corrupted marketplace config, /resume losing session names, and various UI issues. Improved startup performance on macOS and suppressed async hook completion messages by default (visible with --verbose).
Added actionable suggestions to /context command with optimization tips, autoMemoryDirectory setting for custom auto-memory storage, and fixed multiple issues including memory leaks in streaming API response buffers, managed policy ask rules being bypassed, full model IDs being ignored in agent frontmatter, MCP OAuth authentication hanging, refresh token expiration handling, voice mode on macOS missing audio-input entitlement, SessionEnd hooks timeout configuration, marketplace plugin installation and git submodule syncing, RTL text rendering in Windows Terminal/VS Code, LSP server Windows file URI handling, and Hebrew/Arabic text rendering. Also fixed permission prompt deadlocks, skill file change deadlocks, bash tool output loss in multiple sessions, subagent model downgrades on Bedrock/Vertex/Foundry, background bash process cleanup, session resume picker, IDE auto-installation crashes, /loop and /resume availability, hook message injection, voice mode session corruption, Linux sandbox ripgrep errors, Linux native module loading on glibc 2.26, image media_type errors via Remote Control, /heapdump Windows errors, and improved up-arrow restoration after interrupting Claude. Changed default Opus model on Bedrock/Vertex/Microsoft Foundry to Opus 4.6, and deprecated /output-style command in favor of /config.
Added modelOverrides setting to map model picker entries to custom provider model IDs, actionable guidance for OAuth/SSL errors, and fixed multiple issues including freezes/100% CPU loops from permission prompts, deadlocks from skill file changes, bash tool output loss in multiple sessions, subagent model downgrade on Bedrock/Vertex/Foundry, background bash process cleanup, /resume session picker, IDE auto-installation crashes, /loop and voice mode availability, hook double-firing, session corruption on slow connections, Linux sandbox ripgrep errors, Linux glibc 2.26 compatibility, image media_type API errors, /heapdump Windows errors, and improved up-arrow history, IDE detection speed, clipboard pasting performance, /effort availability, voice mode connection retry, and Remote Control spawn mode selection. Changed default Opus model to Opus 4.6 and deprecated /output-style command.
Added w key in /copy to write selections to file, optional description argument to /plan, ExitWorktree tool, CLAUDE_CODE_DISABLE_CRON environment variable, and expanded bash auto-approval allowlist. Restored model parameter on Agent tool, simplified effort levels (low/medium/high with new symbols), improved /config and up-arrow history, improved voice transcription accuracy for technical terms, switched bash parsing to native module, and reduced bundle size by ~510 KB. Fixed multiple issues including slow exits with background tasks, stuck agent progress, double skill hook firing, voice mode input lag and false speech detection errors, --continue not resuming properly, security parsing edge cases, marketplace plugin installation failures on Windows, marketplace blocking issues, plugin cache directory creation, plugin.json loading with marketplace-only fields, feedback survey frequency, --effort flag being reset, backgrounded Ctrl+B query corruption, /clear killing background tasks, worktree isolation issues, /model display during processing, digit key selection in plan mode, sandbox permission issues, prompt cache invalidation in SDK queries, Escape key responsiveness, double Ctrl+C exits, team agent model inheritance, permission rule matching (wildcards, heredocs, env vars, narrowing), oversized images from bash, Bedrock API error crashes, validation errors on Edit/Bash/Grep, multi-line session titles, queued message images, parallel tool call cascading failures, and VSCode scroll speed and keybinding issues. Added VSCode effort indicator on input border and URI handler.
Added /loop command for recurring intervals, cron scheduling tools, rebindable voice:pushToTalk keybinding, and expanded bash auto-approval allowlist with additional commands. Fixed stdin freeze in long sessions, 5-8 second startup freeze from CoreAudio initialization, startup UI freeze with OAuth token refresh, forked conversations sharing plan files, Read tool oversized image handling, false-positive permission prompts for heredoc commit messages, plugin installation loss with multiple instances, claude.ai connector OAuth reconnection, duplicate startup notifications, background agent output file path in notifications, bash error message duplication, Chrome extension detection stuck state, marketplace plugin merge conflicts, marketplace URL parsing with @ refs, duplicate /permissions entries, --print hanging with team agents, ToolSearch REPL noise, Windows git command path prompting, and improved startup time deferral of image processor loading, bridge session reconnection speed, /plugin uninstall scope handling, plugin-provided MCP server deduplication, updated /debug mid-session toggle, and removed unauthenticated connector startup notifications.
Fixed API 400 errors when using ANTHROPIC_BASE_URL with third-party gateways by properly detecting proxy endpoints and disabling tool_reference blocks, fixed API Error: 400 This model does not support the effort parameter when using custom Bedrock inference profiles or non-standard Claude model identifiers, and fixed empty model responses immediately after ToolSearch caused by server-rendered tool schemas with system-prompt-style tags.
New Features:
/claude-api skill for building applications with Claude API and Anthropic SDK/mcp command in VS Code chat panel for native MCP server management (enable/disable, OAuth authentication)/remote-control for custom session titles/reload-plugins command to activate pending plugin changes without restartingInstructionsLoaded hook event when CLAUDE.md or .claude/rules/*.md files are loadedagent_id and agent_type to hook eventsworktree field to status line hook commandsgit-subdir and new OAuth config optionssandbox.enableWeakerNetworkIsolation setting for macOSincludeGitInstructions setting to remove built-in git workflow instructions${CLAUDE_SKILL_DIR} variable for skill referencesImprovements:
/rename to work while Claude is processing/poll rate to once per 10 minutes (cutting server load ~300×)/resume picker to show most recent prompt instead of first oneFixes:
instructions connects after first turnSet-Clipboard#NNN references pointing to wrong repository/plugin/security-review command failing with older git versions/color command—now supports default, gray, reset, none to restore defaultAskUserQuestion preview dialogpermissions.defaultMode settings being applied in Claude Code Remote environments--resume.credentials.json losing subscriptionType on transient token refresh failures/stats crash with malformed timestamps/plugin operations--mcp-config filesacceptEdits modeallowManagedDomainsOnly enabled/clear not fully clearing all session caches.mcp.json servers on first runclaude remote-control crashing on npm installs--model claude-opus-4-0 and --model claude-opus-4-1/model to change settings./simplify and /batch bundled slash commandsENABLE_CLAUDEAI_MCP_SERVERS=false env var to opt out from making claude.ai MCP servers available/model command to show the currently active model in the slash command menu/copy picker/clear not resetting cached skills/copy command with interactive picker for selecting individual code blocks or full response