Released: 2026-10-02
If you accidentally hit a shortcut mid-typing and wipe out your input, you no longer have to stare at a blank terminal in despair and retype everything. Just press the Up arrow key in the empty prompt input box, and the draft you just cleared comes back intact. Previously pasted text and attached images return to the input box along with it.
Whether you want to see every detail of a full code review at once or just control how many issues get reported, you can now set the limit yourself. In Claude's interactive interface, type /code-review --max-findings <n> to specify a count, or /code-review --max-findings all to list every issue found. This choice carries over until you type /code-review --max-findings default to restore the default.
When you have several background tasks running at once, finding a specific task in the agent overview screen is now much faster. You can press Ctrl+F to search sessions by name (a session is an independent conversation flow with the AI), and use Alt+↑ or Alt+↓ to jump quickly between groups. These shortcuts, along with the rename function, can all be remapped in keybindings.json. When filtering with n: or searching with Ctrl+F, pressing Enter opens the most relevant session directly, instead of mechanically opening the first row at the top of the list.
When operating the terminal with voice assistance software, several prompts are now clearer. When you approve an execution plan, the screen reader immediately reads out the new permission mode being applied, and it speaks correctly even when you switch quickly with Shift+Tab. On the /permissions management page, pressing a number key directly selects the corresponding rule, without accidentally triggering the search box. Brief prompts such as word deletion stay on screen until you press any key or the content above changes; questions already answered in a dialog are also clearly marked with the word answered, making it easy to confirm their status. All of these adjustments apply only to screen reader mode.
When digging through history or picking a theme color in full-screen mode, the cursor position now follows your typing in real time. Previously, when searching text in the full-screen history viewer or customizing colors with /theme, the terminal cursor often stayed frozen in place; now it accurately follows the input position. On Windows, the issue where the keyboard could not type after the software auto-restarted due to first-time gateway login, setting up a service provider, or switching the /tui interface has been resolved. After sending a queued message with ctrl+enter, the interrupt prompt line no longer shows a redundant alternative-action hint.
When continuing the same project across sessions over a long period, conversation records and context are now fully preserved. Previously, if the last reply reported zero tokens (the unit of text the AI processes and bills), an overlong conversation would immediately throw a prompt-too-long error; now auto-compaction triggers normally. When resuming a session with --resume, fixes cover restored file context going missing, the last reply not being saved so it looked unanswered, and history being truncated because files were overwritten during loading; restoring conversations created by version 2.1.286 or earlier no longer loses the model's prior thinking either. In addition, the /autocompact auto-compact window size is now stored independently per model, and switching models automatically applies that model's own setting.
When letting the AI run a series of commands on its own, the local safety check mechanism is now steadier and more reliable. When auto mode blocks a non-Bash tool call, it no longer wrongly prompts you to set a Bash permission rule. In the Bedrock and Mantle cloud environments, after auto mode encounters operations that request older models—such as web fetch summarization or a sonnet subagent—the rest of the session no longer falls back to the local classifier. When conversation length exceeds the client safety classifier's checking limit, the conversation history is auto-compacted first, avoiding a confirmation prompt or outright failure on every subsequent tool call; the client classifier also now ignores the lock when ANTHROPIC_DEFAULT_SONNET_MODEL is pinned to Claude Sonnet 5.5 or Opus 5.5, and uses Claude Sonnet 5 instead.
When running tasks in cloud images that do not have GitHub tooling preinstalled, you can now call the built-in API forwarding directly. Cloud sessions now ship with the gh api tool built in, and the issue of filenames, filter conditions, or error messages carrying terminal control characters has been fixed. In self-hosted runners, this built-in tool is further enhanced: when a regular gh command is denied, it prints the corresponding gh api syntax directly; the --paginate flag automatically fetches multi-page repository lists in full; and nested claude invocations no longer accidentally remove the built-in tool.
When collaborating with teammates on cloud tasks across devices, connection state and permission judgments are now more accurate. Previously in Cowork cloud sessions, if an unapproved web fetch permission prompt sat unanswered for five minutes, the session got stuck in a waiting-for-input state; now, after five minutes without a response, the session no longer keeps showing as waiting for input. When a phone joins a Cowork session started from another device, prompt suggestions now display properly; and for messages sent before a /compact compression, the edit-and-retry function is no longer denied. After a cloud session restarts, if the server rejects the newly selected model, replies no longer continue with that model; on restart, models rejected by the organization-enforced model list are no longer restored either.
When building custom interface extensions (mods) for the terminal, you can now precisely capture what the user has highlighted in full-screen mode. The new $.ui.selection() method returns the last selected text; if the selection falls within a single history row, it also returns that row's data. This release also fixes several interface glitches: clicking a button on a screen drawn before the program restarted no longer triggers another button's action; and when an extension displays custom content above the input box, opening the background-task dialog no longer causes an unrecoverable interface error that crashes the app.
Installing plugins on different operating systems or older environments is now more flexible. On macOS and Linux, if the machine lacks a GitHub SSH key, running claude plugin install for a GitHub-sourced plugin will automatically fall back to the HTTPS protocol when cloning fails and print a notice. Installing subdirectory plugins with older Git versions no longer fails or leaves the cache (previously computed or looked-up results stored for direct reuse) incomplete. When a plugin marketplace fails to fetch over both SSH and HTTPS, the error messages from both attempts are shown, with the transport tried first listed on top. Plugins loaded via --plugin-dir now display their settings options properly in /plugin, and claude plugin test no longer misjudges an extension as disabled because of stale settings.
Running complex plugins that include a Language Server Protocol (LSP) server or background timers is now significantly more stable. Plugin LSP settings now correctly substitute ${user_config.*} and ${CLAUDE_PLUGIN_ROOT} with actual setting values or defaults, instead of sending unresolved text variables. When a language server uses dynamic capability registration or stops responding, tool calls automatically time out after 60 seconds instead of hanging forever. For subagents running in a worktree (a Git feature that opens multiple parallel copies of the same project), tool.call hooks no longer cause Bash execution to fail or file searches to read the wrong directory. When these hooks (small programs triggered when specific events occur) fire, they maintain the correct working path. Reloading or disabling a plugin while its timers or read operations are still running no longer forcefully interrupts the background session; if a plugin panel encounters a code diff it cannot parse, it safely falls back to plain-text display.
When running terminal commands that contain scripts or custom variables, safety checks and convenience are now better balanced. If a bash -c or sh -c script contains a dangerous rm deletion targeting the root or home directory, a confirmation prompt is now forced even in bypass-permissions mode or when an allow rule is configured. Commands involving BASHPID arithmetic assignment also proactively prompt before execution instead of silently passing. In sandbox auto-allow mode, heredoc multi-line input with an unquoted delimiter that contains only plain text and simple variables no longer asks for approval every time. When blocking reads outside the working directory is enabled, credential settings in git config files still take effect normally, and commands that were not fully checked now give a more concise reason in the prompt.
When editing project files in a multi-level directory structure, scoped rules now take effect at more precise moments. Previously, path-scoped .claude/rules and nested CLAUDE.md files only loaded when a file was read; now they also load immediately for the corresponding scope when files are created or modified through the write or edit tools. The idle_prompt idle-notification hook no longer fires early while a background agent is still running. If a PreToolUse or PermissionRequest hook fails to match or its input cannot be converted to JSON, the tool call is blocked outright for safety; when a subagent loads rules by accessing files, the InstructionsLoaded hook now fills in the agent ID and type information and correctly reports the thinking effort level.
When dispatching multiple subagents (independent AI assistants that handle specific subtasks) to work together, configuration and message delivery are more transparent. Agent teams launched by name from a plugin now properly apply their own prompts, tool lists, disallowed tools, and thinking effort levels, instead of falling back to global defaults. When a cross-session message is held by the target session and not delivered, the notification clearly states it was undelivered and names that session; in SDK sessions, the AI can also learn the delivery status in real time mid-turn. The lag when launching an agent with an overly long tool list has been fixed; claude mcp serve correctly reports the available subagent types, and prompt annotations in the interface are precisely labeled as us, the main agent, or you depending on the decision maker.
Connecting to external services through MCP (Model Context Protocol, the protocol that lets AI connect to external tools) is now more stable for authorization and tool calls. When an MCP server requests additional authorization scopes during a tool call, a re-authorization prompt now appears proactively. For servers that cannot report authorization completion on their own, the interface waits for you to click done before continuing, giving you plenty of time to work in the browser. When a remote server returns a result over 16 MB or one that cannot be parsed, it no longer triggers repeated calls; subagents in the desktop Code tab can now properly use tools from an MCP server named memory; and cloud sessions no longer wait idly on the first turn for stdio servers configured not to load automatically.
When logging in or rotating cloud keys, status reporting from the credential store (where login proofs are kept) is more accurate. Previously, running /login would still show success even if the credential could not be written to secure storage; now it honestly reports the failure reason and offers a retry option when the login did not take effect. In a --bare clean session, the ineffective login flow no longer runs, avoiding overwriting saved credentials, and you are told which credentials are available. When a laptop wakes from sleep and a background process is already verifying, a second browser login window no longer pops up; pressing stop while Bedrock is querying credentials now ends the request directly instead of mistakenly switching the session to a fallback model; and for Team and Enterprise plan members with credit requests disabled, /usage-credits also shows a clear explanatory message.
Exception handling is more thorough when running long tasks in automated schedules or headless mode (an execution mode with no terminal window or visual interface). When the API times out mid-reply, non-interactive sessions and subagents now continue from the partial reply already received, while timeouts during a pure thinking phase are automatically retried. The watchdog mechanism restarts the stream after a long stream breaks and stops retrying after 3 timeouts. The execution time limit for background commands now applies only to unattended sessions (-p, Agent SDK, CI, cloud); terminal, desktop, and VS Code sessions are unrestricted; and the issue where headless mode ignored termination signals while sending SIGCONT to a monitored process has been fixed. In -p, SDK sessions, and PreToolUse hook approvals, telemetry previously showed the source or decision of a blocked tool as unknown; in -p or interrupted turns, unanswered permission prompts failed to log the tool_decision event. Both have been fixed.
When managing tasks over long periods with the Remote Control feature, connection stability and session retention have been strengthened. When a server credential expires, the session stays connected while the credential is being rotated; even if the server fails and retries stop, the original session is properly preserved instead of being cleared. During background cleanup, sessions that are still connected or were just reattached by another process are no longer mistakenly archived.
Behind Mantle or gateways that reject structured outputs, session titles, memory recall, and prompt hooks (small programs triggered when a prompt arrives) previously failed; you can now disable structured outputs with the CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS environment variable. The issue where Claude 3 Opus and Sonnet errored on every turn after loading a full PDF has been fixed. When the Chrome extension cannot use auto mode, screenshots and reads on approved sites no longer ask repeatedly; typing, navigating, and running JavaScript still prompt. On Team and Enterprise plans or managed machines, creating slides or designs with Artifacts now applies the organization's design system; when the Artifact database reaches its capacity limit, you are told the size limit and how to free up space; and the first request in a fresh environment or after switching models applies the server-configured output limit and compaction window, possibly waiting up to 1.5 seconds.
When using the Claude extension interface inside the VS Code editor, tab handling and conversation management are now tidier. With the new-conversation shortcut enabled, pressing Cmd+N or Ctrl+N now creates a new conversation only in the currently active view, instead of opening a new session in every visible Claude window on screen. After archiving the current conversation in the chat view, the interface cleanly opens a brand-new blank conversation and no longer automatically loads the next saved record. After authorizing a claude.ai connector, if the status remains stuck at needing verification, the MCP server dialog now offers a check-connection option to manually reconfirm the status.
When adjusting cloud features in the admin console or managing self-hosted execution nodes, settings and scheduling controls are more responsive. On the admin settings page, the cloud sessions toggle no longer gets stuck in the off position just because other unrelated security settings are still loading or failed to load. If you press the stop button while a self-hosted runner is starting up, queued messages are now actually canceled, preventing the runner from unexpectedly executing tasks you meant to cancel once it becomes ready.
When collaborating with the Claude Tag feature in the Slack team messaging app, cross-channel information stays in sync more promptly. When a task the AI is working on depends on a relevant thread in another channel that Claude had read but wasn't tracking, it now proactively follows that thread, ensuring newly added context reaches the current conversation in time. For channel settings management, channels whose settings were created automatically no longer show a remove-scope option that was guaranteed to fail; when saving on the channel settings page fails, an overly long custom instruction now triggers a clear request to shorten it, and a save rejected due to lost permissions no longer misleadingly suggests retrying.
For routine project cleanup or tool version updates, commands are shorter to type and installation checks are stricter. claude project purge has been renamed to claude purge; the old name still works and prints a rename notice. When auto-updating packages via npm, if the native binary download fails and the system only has a placeholder stub installed, the update process now honestly reports the failure instead of falsely claiming the update succeeded.
Added recovery for a prompt cleared with Ctrl+C: pressing Up on the empty prompt brings the draft back, including pasted text and imagesAdded `--max-findings <n>|all` to /code-review to report more or fewer findings than the usual limit; the choice is reused until you pass `--max-findings default`Added Ctrl+F to find a session by name and Alt+↑/↓ to jump between groups in the agents view; both, and rename, can be rebound in keybindings.json
Changed the agents view `n:` filter (and Ctrl+F search) so Enter opens the session whose name matches best instead of the top rowAdded a screen reader mode announcement of the new permission mode when you approve a plan, including with Shift+Tab
Fixed `/permissions` in screen reader mode: typing a rule's number now picks it instead of opening the search box
Improved screen reader mode: short announcements, such as a deleted word, now stay on screen until your next key press or until something above them on screen changes
Improved screen reader mode: answered questions in question dialogs now say "answered" beside their boxFixed the keyboard not working on Windows after Claude Code restarts itself (first sign-in to a Claude apps gateway, provider setup, `/tui`)
Fixed the "What should Claude do instead?" hint showing on the Interrupted row after sending queued messages with ctrl+enter
Fixed the terminal cursor not following the typed text in the fullscreen transcript viewer's search and in `/theme`'s custom color searchFixed long conversations failing with "Prompt is too long" instead of auto-compacting when the last reply reported zero token usage
Fixed `--resume` sometimes dropping files and other context that a compaction had just restored
Fixed a resumed session sometimes not saving the last response of a turn, so that the next `--resume` showed the prompt unanswered
Fixed resume occasionally loading a transcript cut short when the same session rewrote the file during the load
Fixed resuming a conversation started on 2.1.286 or earlier dropping the model's earlier thinking
Changed `/autocompact` to save the auto-compact window per model, so each model keeps its own setting when you switchFixed auto mode denials pointing Claude at a Bash permission rule when the blocked tool was not Bash
Fixed auto mode on Bedrock and Mantle switching to the local classifier for the rest of the session after a request to an older model, such as a WebFetch summary or a `sonnet` subagent
Improved auto mode: when a conversation grows too long for the client-side safety classifier to review, it is now compacted instead of prompting for, or failing, every tool call
Changed the client-side auto mode classifier to ignore an `ANTHROPIC_DEFAULT_SONNET_MODEL` pin that names Claude Sonnet 5.5 or Opus 5.5 and use Claude Sonnet 5 insteadAdded a built-in `gh api` to cloud sessions whose image has no GitHub CLI, and fixed the built-in sending control characters from file names, jq filters or GitHub errors to the terminal
Self-hosted runner: Improved the built-in `gh api`: a refused gh command now prints its `gh api` equivalent, `--paginate` follows every page of a repository's lists, and a nested `claude` no longer removes itFixed cloud sessions that restarted on a newly picked model replying with that model after the server refused it
Fixed Cowork cloud sessions staying marked as waiting for input after a WebFetch permission prompt for an unapproved URL went unanswered for five minutes
Fixed prompt suggestions not appearing on a phone that joins a Cowork cloud session started on another device
Fixed Edit and Retry in Cowork cloud sessions refusing a message sent before `/compact` even though its history was still saved
Fixed restarted cloud sessions restoring a model that the organization's enforced model list refusesAdded `$.ui.selection()` for mods: returns the text you last selected in fullscreen mode and, when the selection lies within one transcript row, that row
Fixed a mod's button sometimes running a different button's action when pressed on a view drawn before Claude Code restarted
Fixed fullscreen sessions exiting with "unrecoverable interface error" when opening the background tasks dialog while a plugin or mod showed rows above the promptFixed `git-subdir` plugin installs failing, or caching an incomplete plugin, on older git (before 2.39, e.g. Ubuntu 22.04's 2.34)
Fixed plugins loaded with `--plugin-dir` not showing "Configure options" in `/plugin`
Fixed `claude plugin install` failing for GitHub-source plugins on macOS and Linux machines with no GitHub SSH key: the clone now falls back to HTTPS and prints a notice
Fixed `owner/repo` plugin marketplaces showing only the second attempt's error when both the SSH and HTTPS fetch fail; both errors are now shown, with the transport tried first on top
Fixed `claude plugin test` reporting mods as turned off remotely when it had only read an out-of-date saved settingFixed a plugin's pane showing nothing when one `Code` element held a diff that does not parse; it now draws as plain code
Fixed plugin LSP servers receiving literal `${user_config.*}` and `${CLAUDE_PLUGIN_ROOT}` placeholders in `initializationOptions` and `settings` instead of substituted values or manifest defaults
Fixed a plugin's `tool.call` hook making Bash fail and file searches read the wrong folder in subagents that run in a worktree
Fixed background sessions ending when a plugin was reloaded or disabled while one of its timers or reads was still running
Fixed LSP tool calls hanging indefinitely when a language server uses dynamic capability registration or stops responding; requests now time out after 60s (per-server `requestTimeout`)Fixed sandboxed heredocs with an unquoted delimiter (`python3 <<EOF`) asking for approval on every run under sandbox auto-allow when the body holds only plain text and simple `$VAR` references
Fixed Bash tool permission check to prompt before a `BASHPID` assignment whose value the shell would evaluate as arithmetic, instead of allowing it silently
Fixed `sandbox.credentials.files` entries on git config files not taking effect while `permissions.blockReadsOutsideWorkingDirectories` is on
Fixed a dangerous `rm` (such as one on `/` or the home directory) inside a `bash -c` or `sh -c` script running without a prompt in bypassPermissions mode or under a shell allow rule (anthropics/claude-code#96300)
Improved Bash permission prompts to give a shorter reason when part of a command can't be checked before it runsFixed path-scoped `.claude/rules` and nested CLAUDE.md files not loading when Write or Edit creates or changes a file in their scope (previously only Read loaded them)
Fixed `idle_prompt` notification hooks firing while background agents are still running (anthropics/claude-code#93672)
Fixed PreToolUse and PermissionRequest hooks being skipped when matching them failed or the tool's input could not be serialized to JSON; the call is now blocked
Fixed the InstructionsLoaded hook omitting agent_id and agent_type when a subagent's file access loads a rule or nested CLAUDE.md; rules and nested CLAUDE.md files loaded on file access now also report effortFixed Claude reporting a message to another session as delivered when that session held it: the notice now says it wasn't delivered and names the session, and in SDK sessions Claude can now learn of it mid-turn
Fixed agent teams: a plugin-defined agent spawned by name now runs with its own prompt, tools, disallowedTools and effort instead of the defaults
Fixed a stall when launching an agent whose `tools:` lists very many `Agent(...)` entries
Fixed the Agent tool in `claude mcp serve` always reporting no available agents and rejecting every subagent_type
Improved "You should know" notes to say "we", "the main agent" or "you" depending on who was responsible for a decisionAdded a re-authenticate prompt when an MCP server asks for more OAuth scope during a tool call
Fixed MCP tool calls sometimes running twice when a remote server's result was over 16 MB or could not be parsed
Fixed subagents in Claude Desktop's Code tab getting none of the tools of a user-configured MCP server named `memory`
Improved cloud sessions: a new conversation's first turn no longer waits for a stdio MCP server whose config sets `alwaysLoad: false`
Changed MCP URL prompts from servers that can't report when you're done to wait for "I'm done, continue" before the tool call continues, so you can finish in the browser firstFixed `/login` reporting "Login successful" when credentials could not be saved to secure storage; it now shows the failure, and offers a retry when the new login didn't take effect (anthropics/claude-code#73861)
Fixed a Stop during Bedrock credential lookup sometimes moving the session to a fallback model instead of ending the request
Fixed a second `gcpAuthRefresh`/`awsAuthRefresh` browser sign-in opening when a laptop wakes from sleep while another Claude Code process is signing in
Fixed `/login` in a `--bare` session running a sign-in the session never reads, which could replace your saved login; it now says which credentials work
Improved the `/usage-credits` message shown to Team and Enterprise members whose organization has turned off usage credit requestsFixed mid-response API timeouts failing the turn: non-interactive sessions and subagents now continue from the partial response, and thinking-only responses are retried
Fixed OpenTelemetry `claude_code.tool.blocked_on_user` spans reporting `unknown` source or decision in `-p` and SDK sessions and for PreToolUse hook approvals
Fixed permission asks that ended unanswered, in `-p` or on an interrupted turn, emitting no `tool_decision` event
Fixed unattended sessions (`CLAUDE_CODE_RETRY_WATCHDOG`) retrying for hours after a very long response stream failed; Claude Code now streams again, and gives up after three timeouts
Fixed headless (`-p` / SDK) sessions occasionally ignoring SIGTERM when a supervisor such as `timeout` or systemd sends SIGCONT alongside it
Changed the background command time limit to apply only in unattended sessions (`-p`, Agent SDK, CI, cloud); terminal, desktop app and VS Code sessions have no limitFixed Remote Control cleanup archiving a session that is still connected or was just re-attached by another Claude Code process
Improved Remote Control's recovery from an expired server credential: sessions stay connected during renewal and are kept if it gives up after a server outageFixed session titles, memory recall and prompt hooks failing on Mantle or behind gateways that reject structured outputs; added `CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS` to turn structured outputs off
Fixed Claude in Chrome asking before every screenshot and page read on a site you allowed when auto mode is unavailable (such as with `disableAutoMode` or an older model); typing, navigation and JavaScript still ask
Fixed Claude leaving out your organization's design systems when starting slides or a design with the Artifact tool on Team and Enterprise plans or machines with managed settings
Fixed sessions on Claude 3 Opus and Claude 3 Sonnet failing on every turn after a whole PDF entered the conversation
Fixed the first request in a fresh environment or after a model switch using the built-in output limit and auto-compact window, not the server's; that request may now wait up to 1.5 seconds
Improved the error for an artifact database write refused at the database's size limit: it now states the limit and what frees space[VSCode] Fixed a claude.ai connector staying on "Needs authentication" after you authorize it: the MCP servers dialog now offers Check connection
[VSCode] Fixed the opt-in New Conversation shortcut (Cmd/Ctrl+N) starting a conversation in every visible Claude view instead of only the one you are in
[VSCode] Fixed the chat view resuming the next saved session after you archive the one it shows; it now starts a new conversation instead[Cloud sessions] Fixed the Cloud sessions switch in Claude Code admin settings staying locked off while an unrelated security setting was loading or had failed to load
[Cloud sessions] Fixed pressing Stop while a self-hosted runner was still starting not cancelling the queued message, which could then run once the runner was up[Claude Tag] Fixed Claude Tag admin settings offering a "Remove this scope" option, which always failed, on channels whose settings were created automatically
[Claude Tag] Improved Claude to also follow a related Slack thread in another channel that it only read, so updates there reach the conversation that depends on them
[Claude Tag] Improved save errors on a channel's Configure page: too-long channel instructions now say to shorten them, and a save refused for lost access no longer says to try againFixed the npm auto-updater reporting success when the platform-native binary failed to download and only the placeholder `claude` stub was installed
Changed `claude project purge` to `claude purge`; the old name still works and prints a notice