What changed
Claude Code 2.1.326 shipped on November 16, 2026. Most of its 367 changes fix the display and input handling, but the changes that affect how Claude Code behaves are concentrated in permissions and the sandbox. Some sessions now start in auto mode by default, and some Bash, PowerShell, cp and mv commands and some tool calls in bypass permissions mode now ask first or are refused. Response time and idle CPU use also improved.
New features
-
Code Mode with agent teams
Code Mode can now be turned on while agent teams are enabled. If you already turned it on, sessions with teams on now start in Code Mode.
-
Saying why features are off
/statushas a new row that says when your provider or an environment variable such asDISABLE_TELEMETRYturns features off. Commands that are off for those reasons now say why instead of “Unknown command”. -
Warnings for configuration mistakes
- A startup warning when an
--allowedTools,--disallowedToolsor--toolsvalue has an entry that doesn’t look like a complete rule. - A message explaining why MCP servers stop loading when an administrator adds
managed-mcp.jsonduring a session. - A note in
/pluginandclaude plugin listwhen a malformedplugin.jsonfield such asauthororhomepageis ignored.
- A startup warning when an
-
Hidden notices and command help
Ctrl+O shows hidden startup notices.
/help <command>is new, and/clear --helpand other built-in commands now show usage instead of running at an idle prompt. -
More hints
The empty prompt says what to do after a request fails, and a subagent’s view shows an
esc to stopfooter hint where Esc would stop it. -
OTEL_RESOURCE_INCLUDE_HOSTIncludes
host.nameandhost.idin OpenTelemetry resource attributes. Off by default. -
Mod layout options
overflowX="scroll"on a mod’sBox, andmaxLinesandwrap="clip"on itsText. -
[VSCode]
claudeCode.promptSuggestionssettingA setting that turns prompt suggestions off in the extension.
Key improvements
-
More sessions start in auto mode
- Terminal and VS Code sessions in organizations with the Claude Code + Cowork HIPAA configuration start in auto mode by default;
disableAutoModeturns it off. - Terminal and VS Code sessions with no mode set start in auto mode, not Manual mode, when your organization’s policy hasn’t loaded.
- Sessions that start in auto by default take a server-managed
defaultModeonce settings load, or Manual if settings fail with no stored copy and policy fails or requires it. - For HIPAA-configured organizations on a direct Anthropic API connection, the server-side classifier now checks actions, as for other organizations.
- Terminal and VS Code sessions in organizations with the Claude Code + Cowork HIPAA configuration start in auto mode by default;
-
Stricter permission checks
Some calls that used to run without asking now ask first, or are refused where no one can answer.
- Bash: some commands, commands that reach Claude Code’s sign-in files, commands with very long values the check can’t finish, and some
cpandmvcommands now ask. Some Bash commands on macOS that ran before now ask or are refused. - PowerShell: more commands ask first, and some ask even in bypass permissions mode or are refused. Malformed commands are denied, and commands that define a function named like a built-in command are refused so Claude can rename it.
- Bypass permissions mode: some tool calls ask for approval, or are denied where no one can be asked.
- File tools: writes to Claude Code’s sign-in files such as
.credentials.jsonare refused, as are some device and process files. - Auto mode: wildcard shell allow rules that name no program or start with
$HOMEare ignored so the safety check decides; wildcard rules whose program is a Windows backslash path are now judged (exact rules keep their allow). Auto mode also stops some shell commands that ran before. - Code Mode: more sandboxed commands ask or are refused, and the
$shell helper refuses some values with an error. - A hook’s
"decision": "block"holds in more cases, so a hook that always prints one may start blocking calls. --add-dirand/add-dirin local terminal sessions show the trust prompt for a folder you have not trusted yet;claude --bgrefuses such a folder.
- Bash: some commands, commands that reach Claude Code’s sign-in files, commands with very long values the check can’t finish, and some
-
Bash sandbox changes
- Sandboxed commands on macOS, Linux and WSL no longer write to some sensitive home folders by default; some can be allowed with
sandbox.filesystem.allowWritein user or managed settings. - Unexpected entries in a repository’s
.gitfolder are set aside after a command. - Commands that a permission hook or SDK host rewrites while approving stay in the sandbox if the original was checked as sandboxed; the approver can add
dangerouslyDisableSandbox: true. - With auto-allow off, the sandbox asks before some lines that set environment variables.
- Edits in a session’s mods folder ask more often, and sandboxed commands write to fewer folders. If every edit asks, remove any link there to an unreadable or network path.
- On Linux and WSL, some missing
--add-dirfolders are created before each sandboxed command, and/rewindwith the sandbox on leaves some files untouched and reports them as skipped. claude mcp serveignoressandbox.excludedCommands.
- Sandboxed commands on macOS, Linux and WSL no longer write to some sensitive home folders by default; some can be allowed with
-
Worktree protections
Subagents isolated in a worktree refuse, in any repository, some git commands that can rewrite branches shared with your checkout (sessions are not affected).
ExitWorktreerefuses to remove a worktree while a background subagent or shell command may still be running in it, unless Claude passesdiscard_changes. Worktree creation,/teleportand IDE checkouts run git with a stricterPATH. -
Speed and resource use
- Typing, streaming and resize are more responsive, especially in long sessions.
- Response time after reading or editing files improved, most noticeably on Windows, where a session’s first Bash command is also about a second faster.
- An idle session wakes the machine less often. With more than 20 subagents running, token counts of background agents whose transcript isn’t open update every 5 seconds.
- The first message after
/compactin git repositories and messages queued in Claude Desktop get faster responses. New Claude Desktop sessions no longer wait up to 500 ms on the first turn for local MCP servers whose tools load on demand. /update-configin the terminal uses about 13,000 tokens of context instead of 93,000.
-
Remote Control
Replies from
claude remote-control, Claude Desktop and VS Code sessions now stream to your phone or browser as they are written. Connections through a proxy now use a CONNECT tunnel, so a proxy that refuses CONNECT stops Remote Control with an error. Session debug logs written under--verbosego to~/.claude/child-logs/, and secrets are masked better in push notifications. -
Hook results
Hook failures, blocks and timeouts are easier to tell apart in the transcript: failures are red instead of grey, a hook that stops Claude is yellow, and the reason is shown under them. A PreToolUse hook that exits 127 now gets an error saying its check may not have run, and a PreToolUse hook’s
updatedInputis kept in more cases when more than one hook answers a tool call. -
Plugin changes
- Enabled plugins that were refused for a malformed
plugin.jsondisplay field now load, with their hooks and MCP servers; administrators should check the plugins they deploy. - Plugin uninstall also deletes the files extracted from the plugin’s MCP bundles (
.mcpb,.dxt), unless another plugin uses the same bundle. - HTTP hooks ignore
Content-LengthandTransfer-Encodingin theirheaderssetting, and howallowedHttpHookUrlsandhttpHookAllowedEnvVarsapply to hooks from managed settings changed. /pluginuninstall, remove and switch questions are now lists answered with the arrow keys and Enter, or Esc to cancel;y,nanduno longer answer them.
- Enabled plugins that were refused for a malformed
-
Self-hosted runners
Linux and macOS sessions with no
ghget a built-ingh apiby default, via Anthropic, which adds their GitHub sign-in, if any;--use-anthropic-gh falseturns it off. Under--trust-workspace false, runners ignore more repository settings. -
Changes that affect cost and usage
- Automatic compaction in plan mode uses the model answering you, such as Opus with
opusplan, so it can cost more. - Cowork cloud sessions hold an organization’s OpenTelemetry export until the first turn; a session that gets no message now sends nothing, so
claude_code.session.countcan fall. - “Continue automatically at usage limit” stops with a notice, sending nothing, when your login changes or ends during the wait.
- Automatic compaction in plan mode uses the model answering you, such as Opus with
-
Clearer messages
- The folder trust prompt, at launch and in
/cd, names the project settings files Claude Code will load; at launch it also marks files that run commands, and it says why Claude Code exited. - Errors for a settings or
.mcp.jsonfile with broken JSON give the line and column. - The
CERT_SIGNATURE_FAILUREconnection error names a possible cause after a CA renewal and what IT can change. - When a team’s pooled group budget is used up, the message says when it resets.
/doctorshares findings as it works instead of staying silent until the full report.- In cloud sessions, when an MCP call through the session’s proxy is cut off after it was sent, Claude is told the outcome is unknown. When a Monitor watch expires with no events and its command wrote to stderr, Claude is told so and which file holds the output.
- The folder trust prompt, at launch and in
-
Sign-in screens and terminal UI
The
/logincode goes in a proper text field, and a rejected code says why under it. Pasting a verification code or a web address where it doesn’t belong now says what to do instead. The/pluginDiscover, Marketplaces, Installed and Errors tabs fit short windows and take Home, End, PgUp, PgDn and the mouse. Errors, warnings and confirmations in many dialogs now start with ✘, ⚠ or ✔. -
Other changes
- The
/claude-apiskill builds agents on Managed Agents by default, unless you want to host the agent loop yourself. - Claude in Chrome’s long
browser_batchcalls now have 150 seconds to answer, up from 90. Its permission requests from a terminal session over Remote Control name the site Claude wants to act on in claude.ai. - The card for approving a subagent’s plan now has two choices: “Yes, and switch to default (ask each time) for this session” and “No”.
- In agent view, quitting with Ctrl+C twice, Ctrl+D twice or
/exitafter pressing←in an empty conversation ends that session instead of leaving it running for 5 minutes. - First-run setup preselects the “Auto (match terminal)” theme instead of Dark mode, and fullscreen
/usageshows plan limit bars before session cost for subscribers. - Removed the tmux notice saying the mouse wheel needs
set -g mouse onin fullscreen, and the spinner tip that suggested/clear30 minutes into a running turn.
- The
Bug fixes
Permissions, settings and sandbox
cpof a file being refused when only its source was covered by anEditdeny rule.- Bash commands that delete or overwrite files (
find -delete,cat a > b,sort -o) being summed up in the transcript as a search or a read. - A misleading “Dangerous rm operation” warning on some commands with no rm.
- Rules added in
/permissionsor at a tool prompt, and/add-dirfolders, looking saved when the settings file couldn’t be written — the terminal now warns and names it. - Managed settings with certain key names stopping Claude Code from starting, settings files with values nested thousands of levels deep crashing Claude Code (values past 64 levels are now dropped with a warning), and a settings file saved with an error mid-session being skipped silently.
- Up then Enter answering Yes at the folder trust prompt shown under an already trusted parent folder, and an extra Enter at the end of setup accepting the offer to make auto mode the default.
- PreToolUse hooks in managed settings leaving no OpenTelemetry
tool_decisionevent orpermission_denialsentry when they refused a direct tool call. - Auto mode staying on its local classifier, which counts toward usage, because of a stale saved value.
- The subagent model set in managed settings not applying in Claude Desktop’s Code tab.
- Linux: sandboxed commands failing to start when the home or config directory is reached through a symlink into
/netand sandbox settings deny reads of it, and empty placeholder files such as.git/config.lockbeing left in the project when Claude Code was stopped during a sandboxed command.
Sign-in and background sessions
- Background sessions losing the LLM gateway your shell exported after a crash, a reopen from another folder, or when the background service lacks it — they now stop before any model request.
- The background service ending your own processes, such as an SSH login, when a process number on its list had been reused by another user’s process.
- A background session being killed and its request lost when you opened, stopped or removed the older session whose conversation it had resumed.
- “Sign in with your Console account” in
/loginreplacing your Claude subscription login when the browser sign-in used your Claude plan instead of a Console organization. /loginending on “Login successful” while an approvedANTHROPIC_API_KEYwas still used, and “Not logged in” staying after turning on the API key in/config— these screens now name an unusedANTHROPIC_API_KEY.claude auth logoutleaving credentials behind on a drive that rejects folder timestamps.- Model requests failing on a one-off authentication error (401) in Claude Desktop’s Code tab — they are now sent once more first.
forceLoginOrgUUIDandforceRemoteSettingsRefreshstartup errors telling you to sign in again when a network block page (HTTP 403) was the cause.- The
/privacy-settingstoggle sometimes saving the opposite of what the screen showed on a quick double press, or saving on the wrong account after signing in to a different account in another terminal, and your OFF answer to the updated terms being replaced when keys were typed while it was saving.
Conversations, compaction and resumed sessions
- Resumed sessions losing recent turns after a crash when the previous turn ended with an empty reply, or losing earlier history when Claude Code ran low on memory.
- Compaction on Bedrock, Vertex and Foundry, and with telemetry turned off, sometimes replying to one of your messages instead of summarizing, or leaving Claude’s last reply out of the summary.
- Image-heavy sessions on Vertex AI getting stuck on a 400 error about the request size exceeding a 30 MB limit — they now compact and retry.
- An error when
max_tokensexceeded the model’s limit — Claude Code now retries with the maximum the Anthropic API reports. - Valid PDFs being dropped after the API rejected a different PDF in the same batch, and
grepin the Bash tool silently skipping some text files. - The automatic session recap being sent after a reply billed to usage credits —
/recapstill works. /plan <description>ignoring the description when plan mode was already on, and a prompt typed after--add-dir,--allowedToolsor another list option being dropped without a message./rewindsaying “No code changes” when a shell command changed files it can’t restore — such a turn now reads “Code changes unknown”.
Fixes that could lose data
claude purgehiding its prompts when stdout is redirected, where one unseen Enter could delete a project’s data — they now show on the terminal.- The spend limit menu removing the monthly limit without asking when Backspace and Enter were pressed as it opened.
- Ctrl+C or Ctrl+D twice in
/skillsquitting Claude Code and dropping changes — it now saves and closes the panel. - A re-sent file overwriting Claude’s changed copy in Remote Control and cloud sessions on Linux with no feature flags loaded.
/feedbackand/bugsending the report when you typed a space on the confirmation screen.
MCP
- Parallel MCP tool calls reading another call’s saved result when two long results arrived in the same millisecond (anthropics/claude-code#105138).
- Tool search sometimes loading another MCP server’s tool when asked for a tool by its exact name — you may be asked again for that tool.
- MCP prompts disappearing when two servers have nearly identical names.
- MCP tools with a root
anyOforoneOfshowing Claude only the first branch’s form of an argument — text such as42can now arrive as a number where another branch takes one. - A failed MCP
headersHelper, or one not run in an untrusted folder, making other sessions start without that server. /mcpnot showing why a local (stdio) MCP server failed to start, andclaude mcp add-jsonanswering most mistakes with “Invalid configuration: : Invalid input”.- Esc doing nothing while an MCP server reconnects — it now closes the wait and the reconnect continues in the background.
Code Mode
- Skills and slash commands failing to load in Code Mode under auto mode when an inline
!command needs approval, and background subagents started in Code Mode failing when they preload a skill with an inline!command. - Mods being unable to use Bash, Read and other tools through
$.tool.calland$.tool.checkin Code Mode.
Remote Control, cloud and artifacts
- A Remote Control session disconnecting mid-reply when the server was briefly unavailable, and the QR code in
/remote-controlbeing cut off in short terminal windows such as 80×24. - Terminal sessions showing as idle on your phone while background subagents were still running — Stop on your phone or the web now ends them together with the turn.
- Stop in cloud and self-hosted runner sessions waiting for slow MCP servers to connect, and file uploads failing at once when a proxy answers with HTTP 502, 504 or 520 to 524 (they now get a second try).
claude self-hosted-runnerasking for new tokens nonstop when a refreshed token was already due, such as when the machine’s clock runs ahead.- Artifact publishing refusing files in a background session’s temporary folder, subagents whose
disallowedToolslistsArtifactstill being offered artifact tools in some cloud sessions, and the terminal not showing an artifact’s link after Claude publishes or updates it.
Subagents and agent rows
- The main conversation waiting in silence for a background subagent that another subagent had messaged — it’s now told which agent the subagent reports to.
- Up and Esc moving the main conversation’s queued messages into the prompt box of a subagent or teammate you are viewing.
- Agent rows showing an agent as busy after its turn ended, and stopped or failed agents looking like running ones — rows now say what an agent is waiting on, and
/tasksnames the subagent behind each background command. - The “background agents launched” line stopping at two when Claude starts three or more agents in one reply.
Display, input and terminals
- The transcript dropping rows or the prompt jumping as a reply started streaming, and lines of long lists or code blocks going missing from scrollback.
- Some very long Bash or PowerShell commands, and long newly created files in the permission prompt, freezing the screen for seconds.
- Shift+Enter or Option+Enter pressed right after a paste sending the prompt, and Up, Down, End and mouse clicks putting the cursor on the wrong row of a prompt that wraps.
- Asterisks disappearing from your own prompts in the transcript, such as
src/**/*.tsor2*3*4, and pasted code or diffs losing their indentation, with a pasted diff’s added lines showing as removed. - Bold text showing as raw
**in Japanese, Chinese and Korean replies when it starts or ends with punctuation. - Windows paths in Claude’s responses losing the backslash before folders like
.claudeor_build. - Wide tables breaking apart in plans and subagent prompts, and code keywords, added diff lines and “thinking” being hard to read in some themes and terminals.
- Text being drawn out of place after skin-tone or joined emoji, symbols such as ✌ and ☝, and Hindi, Tamil and other Indic vowel signs.
- Desktop notifications not being sent inside tmux (with
allow-passthroughon) or to iTerm2 over SSH, and typing and spinners redrawing the whole pane in tmux 3.7. - The terminal progress bar disappearing in Ghostty after 15 seconds while Claude was still working, and Program Status Protocol (OSC 7501) terminals not seeing Claude Code as waiting on you while the folder trust prompt is open (anthropics/claude-code#105801).
- The screen scrambling after
/tui defaultin a long conversation (anthropics/claude-code#94879), and a custom status line keeping its old width, folder and session name after a terminal resize,/cdor/rename(anthropics/claude-code#107291). /btwshowing only a lead-in such as “Let me check that file.” when the model tried to use a tool (anthropics/claude-code#90314)./resumesearch saying “No sessions match” for older sessions that exist, and search in/resume,/plugin,/permissionsand/configignoring IME input, dictation and fast typing./modelshowing your saved model as “Default (recommended)” — Default now names your plan’s default model.
VS Code
- “Rewind code to here” in a forked conversation sometimes undoing one turn too many and deleting files created in that turn.
- The Sandbox dialog creating empty files and folders in the home directory and the project’s
.gitfolder while the sandbox is off. - The saved sign-in being lost on some networks, and sign-in ending when a verification code, a web address or a code from another sign-in attempt is pasted.
/idein a second Claude Code session silently losing the editor connection back to the first session — the newest connection now wins, and the displaced session shows how to switch back.- Saved memories not being listed in the Memory dialog, or RAM use growing without end, when the memory folder held symlink loops.
Notes
- If your organization uses the Claude Code + Cowork HIPAA configuration, or its policy hasn’t loaded at startup, sessions can start in auto mode. In the HIPAA configuration,
disableAutoModeturns it off. - If unattended runs rely on Bash, PowerShell,
cp,mvor bypass permissions mode, some of those commands now ask, and where no one can answer they are refused. Check such runs after updating. - If a hook always prints
"decision": "block", it may start blocking calls, since the decision now holds in more cases. - If you use a git filter or helper, worktree creation,
/teleportand IDE checkouts now run git with a stricterPATH, so it may need its full path. - If sandboxed commands need to write to sensitive home folders, some can be allowed with
sandbox.filesystem.allowWritein user or managed settings. - Administrators: plugins refused for a malformed
plugin.jsondisplay field now load with their hooks and MCP servers, so check the plugins you deploy. A managedautoMode.environmentlist without"$defaults"now replaces every built-in entry in Claude Desktop and Agent SDK apps too; add"$defaults"or copy the sensitive-target ones. - If you use
opusplan, automatic compaction in plan mode now runs on Opus and can cost more. - If Remote Control runs behind a proxy, it now uses a CONNECT tunnel and stops with an error if the proxy refuses CONNECT.
- If your self-hosted runners have no
gh, they get a built-ingh apivia Anthropic by default; pass--use-anthropic-gh falseto turn it off. - If you track Cowork cloud sessions in OpenTelemetry, sessions that get no message now send nothing, so
claude_code.session.countcan fall. - This post covers the main items only. See the GitHub release notes for the full list.