+Addedclaude remote-control subcommand for external builds, enabling local environment serving for all users.
+Added support for custom npm registries and specific version pinning when installing plugins from npm sources
+AddedCLAUDE_CODE_ACCOUNT_UUID, `CLAUDE_CODE_USER_EMAIL`, and `CLAUDE_CODE_ORGANIZATION_UUID` environment variables for SDK callers to provide account info synchronously, eliminating a race condition where early telemetry events lacked account metadata.
~Updated plugin marketplace default git timeout from 30s to 120s and added `CLAUDE_CODE_PLUGIN_GIT_TIMEOUT_MS` to configure.
~BashTool now skips login shell (`-l` flag) by default when a shell snapshot is available, improving command execution performance. Previously this required setting `CLAUDE_BASH_NO_LOGIN=true`.
~Tool results larger than 50K characters are now persisted to disk (previously 100K). This reduces context window usage and improves conversation longevity.
~The /model picker now shows human-readable labels (e.g., "Sonnet 4.5") instead of raw model IDs for pinned model versions, with an upgrade hint when a newer version is available.
~Managed settings can now be set via macOS plist or Windows Registry. Learn more at https://code.claude.com/docs/en/settings#settings-files
✓Fixed a security issue where statusLine and `fileSuggestion` hook commands could execute without workspace trust acceptance in interactive mode.
✓Fixed a bug where duplicate control_response messages (e.g. from WebSocket reconnects) could cause API 400 errors by pushing duplicate assistant messages into the conversation.
✓Fixed slash command autocomplete crashing when a plugin's SKILL.md description is a YAML array or other non-string type
+Added support for startupTimeout configuration for LSP servers
+AddedWorktreeCreate and `WorktreeRemove` hook events, enabling custom VCS setup and teardown when agent worktree isolation creates or removes worktrees.
+Added support for isolation: worktree in agent definitions, allowing agents to declaratively run in isolated git worktrees.
~CLAUDE_CODE_SIMPLE mode now also disables MCP tools, attachments, hooks, and CLAUDE.md file loading for a fully minimal experience.
~Opus 4.6 (fast mode) now includes the full 1M context window
↑Improved memory usage during long sessions by clearing internal caches after compaction
↑Improved memory usage during long sessions by clearing large tool results after they have been processed
↑Improved startup performance for headless mode (`-p` flag) by deferring Yoga WASM and UI component imports
✓Fixed a bug where resumed sessions could be invisible when the working directory involved symlinks, because the session storage path was resolved at different times during startup. Also fixed session data loss on SSH disconnect by flushing session data before hooks and analytics in the graceful shutdown sequence.
✓LinuxFixed native modules not loading on systems with glibc older than 2.30 (e.g., RHEL 8)
✓Fixed memory leak in agent teams where completed teammate tasks were never garbage collected from session state
✓FixedCLAUDE_CODE_SIMPLE to fully strip down skills, session memory, custom agents, and CLAUDE.md token counting
✓Fixed/mcp reconnect freezing the CLI when given a server name that doesn't exist
✓Fixed memory leak where completed task state objects were never removed from AppState
✓Fixed bug where MCP tools were not discovered when tool search is enabled and a prompt is passed in as a launch argument
✓Fixed a memory leak where LSP diagnostic data was never cleaned up after delivery, causing unbounded memory growth in long sessions
+Added--worktree (`-w`) flag to start Claude in an isolated git worktree
+Added Ctrl+F keybinding to kill background agents (two-press confirmation)
+AddedConfigChange hook event that fires when configuration files change during a session, enabling enterprise security auditing and optional blocking of settings changes.
~Agent definitions support `background: true` to always run as a background task
~Plugins can ship settings.json for default configuration
~Simple mode (CLAUDE_CODE_SIMPLE) now includes the file edit tool in addition to the Bash tool, allowing direct file editing in simple mode.
~Permission suggestions are now populated when safety checks trigger an ask response, enabling SDK consumers to display permission options
~Sonnet 4.5 with 1M context is being removed from the Max plan in favor of our frontier Sonnet 4.6 model, which now has 1M context. Please switch in /model.
~SDK model info now includes `supportsEffort`, `supportedEffortLevels`, and `supportsAdaptiveThinking` fields so consumers can discover model capabilities.
↑ImprovedMCP OAuth authentication with step-up auth support and discovery caching, reducing redundant network requests during server connections
↑Improved performance in non-interactive mode (`-p`) by skipping unnecessary API calls during startup
↑Improved performance by caching authentication failures for HTTP and SSE MCP servers, avoiding repeated connection attempts to servers requiring auth
↑Improved startup performance by caching MCP auth failures to avoid redundant connection attempts
↑Improved startup performance by reducing HTTP calls for analytics token counting
↑Improved startup performance by batching MCP tool token counting into a single API call
↑Improved permission prompts for path safety and working directory blocks to show the reason for the restriction instead of a bare prompt with no context
✓Fixed file-not-found errors to suggest corrected paths when the model drops the repo folder
✓Fixed Ctrl+C and ESC being silently ignored when background agents are running and the main thread is idle. Pressing twice within 3 seconds now kills all background agents.
✓Fixed prompt suggestion cache regression that reduced cache hit rates.
✓Fixedplugin enable and `plugin disable` to auto-detect the correct scope when `--scope` is not specified, instead of always defaulting to user scope
✓Fixed verbose mode not updating thinking block display when toggled via `/config` — memo comparators now correctly detect verbose changes
✓Fixed unbounded WASM memory growth during long sessions by periodically resetting the tree-sitter parser
✓Fixed potential rendering issues caused by stale yoga layout references
✓Fixed unbounded memory growth during long-running sessions caused by Yoga WASM linear memory never shrinking
+Addedlast_assistant_message field to Stop and SubagentStop hook inputs, providing the final assistant response text so hooks can access it without parsing transcript files.
+Addedchat:newline keybinding action for configurable multi-line input (anthropics/claude-code#26075)
+Addedadded_dirs to the statusline JSON `workspace` section, exposing directories added via `/add-dir` to external scripts (anthropics/claude-code#26096)
~Search patterns in collapsed tool results are now displayed in quotes for clarity
~Use ctrl+f to kill all background agents instead of double-pressing ESC. Background agents now continue running when you press ESC to cancel the main thread, giving you more control over agent lifecycle.
~Simplified teammate navigation to use only Shift+Down (with wrapping) instead of both Shift+Up and Shift+Down.
~Increased initial session count in resume picker from 10 to 50 for faster session discovery (anthropics/claude-code#26123)
~Moved config backup files from home directory root to `~/.claude/backups/` to reduce home directory clutter (anthropics/claude-code#26130)
~The /rename command now updates the terminal tab title by default (anthropics/claude-code#25789)
↑ImprovedVS Code plan preview: auto-updates as Claude iterates, enables commenting only when the plan is ready for review, and keeps the preview open when rejecting so Claude can revise.
↑Improved memory usage in long-running sessions by releasing API stream buffers, agent context, and skill state after use
↑Improved startup performance by deferring SessionStart hook execution, reducing time-to-interactive by ~500ms.
↑Improved performance of @ file mentions - file suggestions now appear faster by pre-warming the index on startup and using session-based caching with background refresh.
↑Improved memory usage by trimming agent task message history after tasks complete
↑Improved memory usage during long agent sessions by eliminating O(n²) message accumulation in progress updates
✓Fixed FileWriteTool line counting to preserve intentional trailing blank lines instead of stripping them with `trimEnd()`.
✓Fixed Windows terminal rendering bugs caused by `os.EOL` (`\r\n`) in display code — line counts now show correct values instead of always showing 1 on Windows.
✓Fixed a bug where bold and colored text in markdown output could shift to the wrong characters on Windows due to `\r\n` line endings.
✓Fixed compaction failing when conversation contains many PDF documents by stripping document blocks alongside images before sending to the compaction API (anthropics/claude-code#26188)
✓Fixed an issue where bash tool output was silently discarded on Windows when using MSYS2 or Cygwin shells.
✓Fixed the bash permission classifier to validate that returned match descriptions correspond to actual input rules, preventing hallucinated descriptions from incorrectly granting permissions
✓Fixed user-defined agents only loading one file on NFS/FUSE filesystems that report zero inodes (anthropics/claude-code#26044)
✓Fixed plugin agent skills silently failing to load when referenced by bare name instead of fully-qualified plugin name (anthropics/claude-code#25834)
+Added support for reading enabledPlugins and `extraKnownMarketplaces` from `--add-dir` directories
+AddedspinnerTipsOverride setting to customize spinner tips — configure `tips` with an array of custom tip strings, and optionally set `excludeDefault: true` to show only your custom tips instead of the built-in ones
+AddedSDKRateLimitInfo and `SDKRateLimitEvent` types to the SDK, enabling consumers to receive rate limit status updates including utilization, reset times, and overage information
↑Improved startup performance by removing eager loading of session history for stats caching
↑Improved memory usage for shell commands that produce large output — RSS no longer grows unboundedly with command output size
↑Improved collapsed read/search groups to show the current file or search pattern being processed beneath the summary line while active
↑VSCodeImproved permission destination choice (project/user/session) to persist across sessions
✓FixedAgent Teams teammates failing on Bedrock, Vertex, and Foundry by propagating API provider environment variables to tmux-spawned processes (anthropics/claude-code#23561)
✓Fixed sandbox "operation not permitted" errors when writing temporary files on macOS by using the correct per-user temp directory (anthropics/claude-code#21654)
✓FixedTask tool (backgrounded agents) crashing with a `ReferenceError` on completion (anthropics/claude-code#22087)
✓Fixed autocomplete suggestions not being accepted on Enter when images are pasted in the input
✓Fixed skills invoked by subagents incorrectly appearing in main session context after compaction
✓Fixed excessive .claude.json.backup files accumulating on every startup
✓Fixed plugin-provided commands, agents, and hooks not being available immediately after installation without requiring a restart
✓Fixed a crash when agent teams setting changed between renders
✓Fixed a bug where commands excluded from sandboxing (via `sandbox.excludedCommands` or `dangerouslyDisableSandbox`) could bypass the Bash ask permission rule when `autoAllowBashIfSandboxed` was enabled
+Added plugin name to skill descriptions and `/skills` menu for better discoverability
+VSCodeAdded support for remote sessions, allowing OAuth users to browse and resume sessions from claude.ai
+VSCodeAdded git branch and message count to the session picker, with support for searching by branch name
~Errors from invalid managed settings are now surfaced
↑Improved error messages for API connection failures — now shows specific cause (e.g., ECONNREFUSED, SSL errors) instead of generic "Connection error"
✓Fixed agent teammate sessions in tmux to send and receive messages
✓Fixed warnings about agent teams not being available on your current plan
✓Fixed an issue where submitting a new message while the model was in extended thinking would interrupt the thinking phase
✓Fixed an API error that could occur when aborting mid-stream, where whitespace text combined with a thinking block would bypass normalization and produce an invalid request
✓Fixed API proxy compatibility issue where 404 errors on streaming endpoints no longer triggered non-streaming fallback
✓Fixed an issue where proxy settings configured via `settings.json` environment variables were not applied to WebFetch and other HTTP requests on the Node.js build
✓Fixed/resume session picker showing raw XML markup instead of clean titles for sessions started with slash commands
✓VSCodeFixed scroll-to-bottom under-scrolling on initial session load and session switch
+Added research preview agent teams feature for multi-agent collaboration (token-intensive feature, requires setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
+Added"Summarize from here" to the message selector, allowing partial conversation summarization.
+VSCodeAdded spinner when loading past conversations list
~Claude now automatically records and recalls memories as it works
~Skills defined in .claude/skills/ within additional directories (`--add-dir`) are now loaded automatically.
~Updated --resume to re-use --agent value specified in previous conversation by default.
~Skill character budget now scales with context window (2% of context), so users with larger context windows can see more skill descriptions without truncation
✓Fixed@ file completion showing incorrect relative paths when running from a subdirectory
✓Fixed: Bash tool no longer throws "Bad substitution" errors when heredocs contain JavaScript template literals like `${index + 1}`, which previously interrupted tool execution
✓Fixed Thai/Lao spacing vowels (สระ า, ำ) not rendering correctly in the input field
✓VSCodeFixed slash commands incorrectly being executed when pressing Enter with preceding text in the input field
↑Improved system prompts to more clearly guide the model toward using dedicated tools (Read, Edit, Glob, Grep) instead of bash equivalents (`cat`, `sed`, `grep`, `find`), reducing unnecessary bash command usage
↑Improved PDF and request size error messages to show actual limits (100 pages, 20MB)
↑Reduced layout jitter in the terminal when the spinner appears and disappears during streaming
✓Fixed PDF too large errors permanently locking up sessions, requiring users to start a new conversation
✓Fixed bash commands incorrectly reporting failure with "Read-only file system" errors when sandbox mode was enabled
✓Fixed a crash that made sessions unusable after entering plan mode when project config in `~/.claude.json` was missing default fields
✓FixedtemperatureOverride being silently ignored in the streaming API path, causing all streaming requests to use the default temperature (1) regardless of the configured override
✓Fixed LSP shutdown/exit compatibility with strict language servers that reject null params
+Addedpages parameter to the Read tool for PDFs, allowing specific page ranges to be read (e.g., `pages: "1-5"`). Large PDFs (>10 pages) now return a lightweight reference when `@` mentioned instead of being inlined into context.
+Added pre-configured OAuth client credentials for MCP servers that don't support Dynamic Client Registration (e.g., Slack). Use `--client-id` and `--client-secret` with `claude mcp add`.
+Added/debug for Claude to help troubleshoot the current session
+Added support for additional git log and `git show` flags in read-only mode (e.g., `--topo-order`, `--cherry-pick`, `--format`, `--raw`)
+Added token count, tool uses, and duration metrics to Task tool results
+Added reduced motion mode to the config
+VSCodeAdded multiline input support to the "Other" text input in question dialogs (use Shift+Enter for new lines)
~Changed/model to execute immediately instead of being queued
↑Improved memory usage for --resume (68% reduction for users with many sessions) by replacing the session index with lightweight stat-based loading and progressive enrichment
↑ImprovedTaskStop tool to display the stopped command/task description in the result line instead of a generic "Task stopped" message
✓Fixed phantom "(no content)" text blocks appearing in API conversation history, reducing token waste and potential model confusion
✓Fixed prompt cache not correctly invalidating when tool descriptions or input schemas changed, only when tool names changed
✓Fixed 400 errors that could occur after running `/login` when the conversation contained thinking blocks
✓Fixed a hang when resuming sessions with corrupted transcript files containing `parentUuid` cycles
✓Fixed rate limit message showing incorrect "/upgrade" suggestion for Max 20x users when extra-usage is unavailable
✓Fixed permission dialogs stealing focus while actively typing
✓Fixed subagents not being able to access SDK-provided MCP tools because they were not synced to the shared application state
✓Fixed a regression where Windows users with a `.bashrc` file could not run bash commands
