What changed
Claude Code 2.1.328 shipped on November 19, 2026, with more than 140 changes. The headline is auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry and Claude apps gateway: interactive sessions with no permission mode start in it. /code-review narrows to bugs and CLAUDE.md violations, Code Mode gets more than 20 fixes and improvements, and there are many fixes for self-hosted runners, Remote Control and terminal input.
New features
-
Auto mode for Opus 4.6 and Sonnet 4.6 on third-party providers
Auto mode wasn’t available for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry and Claude apps gateway. It is now, and interactive sessions with no permission mode start in it. Plan mode and Bash prompts for these models also use and offer auto mode, as on newer models;
disableAutoModeturns both off. -
"onFailure": "block"formcp_toolhooksCommand and HTTP hooks could already block an action when the hook’s call failed;
mcp_toolhooks couldn’t. They now take"onFailure": "block"too, so a hook whose call fails blocks the action. -
A notice when an OpenTelemetry exporter has no endpoint
An exporter with no endpoint would likely lose its data without saying so.
/statusandclaude doctornow show a notice. -
CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUMEclaude -p --resumeand--continuecould run a tool call that a hook had deferred. This variable lets a host stop that. -
Warnings about held cross-session messages
A message from a session that skips permission prompts may be held and dropped by a cloud session, and there was no sign of it. Such messages now get a warning.
claude -ptext output says on stderr when a message from another session is held or dropped, and how to allow it, and VS Code shows a notice with a link to the settings file that can allow it. -
Policy settings for administrators
- Apps that embed Claude Code can switch on the policy plugin’s
allowManagedModsOnlyoption; an admin’s managed settings still win. - Claude apps gateway shows a startup warning when a policy’s
maxEffortLevelis not a valid effort level.
- Apps that embed Claude Code can switch on the policy plugin’s
-
Self-hosted runner flags and metrics
--repository-foldernames the folder that sessions use for a repository at--capacity 1, in place of its path under the base folder, which can’t be a symbolic link.- A start-up warning when the host’s settings.json sets a default permission mode that sessions ignore.
--hook-labelonclaude self-hosted-runner orchestratorhands chosen session labels to thespawn-runnerhook, and--metadata key=valueonclaude --environmentlabels a session for it.claude_code_self_hosted_orchestrator_session_queue_wait_by_reason_secondsin the orchestrator’s/metrics, so a re-opened session’s wait can be told from a new one’s.claude_code_self_hosted_orchestrator_sessions_given_up_totalcounts by cause how often Anthropic gave up starting a runner for a session.
Key improvements
-
/code-reviewfor bugs,/simplifyfor cleanups/code-reviewused to mix reuse, simplification and efficiency suggestions in with bugs. On most models it now reports only bugs and CLAUDE.md violations; run/simplifyfor reuse, simplification and efficiency cleanups. Each finding is marked Important, Pre-existing or Nit, each summary ends with a suggested fix, and a clean review’s list reads “Code review found no issues.” -
Code Mode speed and token use
- In auto and bypass-permissions modes, the 0.75 second pause after most scripts is gone.
- Prompt caching is better for new Code Mode sessions that have MCP servers, and Claude is told when a server is still connecting and when the list of servers changes.
- Shorter tool descriptions use fewer tokens.
-
What Code Mode shows and tells Claude
- The command a script is running, and its output, show while it runs.
- An edit that changes nothing no longer stops a script, and Claude is told where a failed edit matches or comes closest.
- When a script fails, Claude is told which tool calls were still running.
- In auto and bypass modes, Claude is told to run dev servers as background tasks that show in
/tasks. - Windows: Claude is told that scripts’ shell commands run in Git Bash and that PowerShell aliases run instead of functions with the same name.
-
Resuming and memory
Large sessions resume faster, resuming a session with many images uses less memory, and long SDK and desktop sessions that run slow commands or tools use less memory and respond sooner.
-
Managed-settings plugins
claude -pand SDK sessions fetch managed-settings plugins that have their own git repository in the background, for a later session.- Managed-settings plugins from a git marketplace load sooner on a fresh machine in
claude -p, SDK and Claude Desktop sessions; a short run may exit up to 30 seconds later.
-
Hook and sandbox behavior
- Code Mode no longer runs UserPromptSubmit hooks, managed ones included, on scripts’ output. To redact that output, use a PostToolUse hook with
updatedToolOutput. - Managed
PostToolUsehooks no longer fire for calls refused before they ran: no log entry, rewrite, block or stop there.PostToolBatchgets most, not marked as refused. - A shell command rewritten by a PermissionRequest hook to a sandbox-excluded command runs outside the sandbox.
- Unless
sandbox.failIfUnavailableis set, a sandbox that can’t start is reported by refusing the next sandboxed Bash command, with a message that the sandbox is off for the session. deniedMcpServersentries of the formclaude.ai <name>apply to that connector in cloud and self-hosted runner sessions too.
- Code Mode no longer runs UserPromptSubmit hooks, managed ones included, on scripts’ output. To redact that output, use a PostToolUse hook with
-
Subagents and teammates
- With
CLAUDE_CODE_FORK_SUBAGENT=0, subagents inclaude -p, the Agent SDK, desktop and VS Code wait for their own subagents by default. - In that setup,
claude -pcounts a background subagent and those it waits for against one 10-minute wait; raiseCLAUDE_CODE_PRINT_BG_WAIT_CEILING_MSfor longer chains. - Subagent progress summaries in terminal sessions on the Anthropic API are written by the default small fast model from short notes, also with telemetry off, using far fewer tokens.
- Starting several in-process teammates at once: far fewer now fail with “Lock file is already being held”.
- With
-
Prompts and explanations
- The computer use approval dialog shows each app’s level of access and plainer warnings.
- Cross-session messaging: Claude explains that Remote Control is off when a cloud session or another computer is not listed.
- The Claude in Chrome welcome screen and
/chromesay that Claude Code asks before Claude acts on a site, unless your permission mode or a rule in/permissionsanswers. /bugasks Yes or No before sending, so repeated Enters no longer send a report or open GitHub./list-agentsshows its list to you alone; Claude lists sessions itself when needed.- Asked who can open an artifact it published, Claude is told to check the artifact again instead of repeating the last publish’s answer.
-
Third-party providers, or telemetry or feature flags off
- Claude Desktop and Agent SDK sessions end the turn when closed with a question open, instead of continuing without an answer.
- Pasted text is marked for Claude as pasted.
-
Self-hosted runner changes
- The built-in
gh apiis off by default where settings guard Claude’s GitHub tools by name;--use-anthropic-gh trueturns it on. - A
claudethat a wrapper script runs before the session’s ownclaudeis up, or after it has ended, has no Anthropic-managed git. CLAUDE_RUNNER_SESSION_UUIDandCLAUDE_CODE_REMOTE_SESSION_UUIDare both set for the session and itscheckout,post-sessionandspawn-runnerhooks.- At
--capacity 1, a session is refused earlier when the path of a repository it pushes to is a symbolic link. - In resumed sessions, Claude checks whether its earlier work is still there and says what is missing.
- The built-in
Bug fixes
Shell and commands
- Bash commands in zsh stopping with a “not found” error at a word that starts with
=, such asecho ====or[ a == b ]—=cmdno longer expands to a command’s path. - The Bash sandbox on Linux renaming
.git/HEADafter each command when the branch name is not valid UTF-8. - An image path pasted into a
!command becoming an attachment.
Conversations, compaction and resumed sessions
- Automatic compaction reusing the session’s earlier system prompt instead of rebuilding it the way
/compactdoes. - A resumed cloud session losing its compaction summary or recent messages when a message was saved late.
- Cloud sessions resuming a cleared conversation after a rewind followed by
/clear. - A remote session losing a turn’s tool calls and results when its runner or worker was lost after a dropped model stream had been retried.
- Some Claude Desktop sessions on third-party providers, other than scheduled tasks, ending for good with “Prompt is too long” after large tool results — the largest are now left out.
Scheduled tasks
- Saving a scheduled task erasing the project’s other saved tasks when the scheduled tasks file can’t be read, on third-party providers or with telemetry or feature flags turned off.
- A scheduled task with a typo in the project’s scheduled tasks file being deleted the next time that file was saved.
Third-party providers and gateways
- Auto mode failing every turn outside Claude Desktop with
CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETASset behind a gateway rejecting betas, on Bedrock, Agent Platform, Foundry or Claude apps gateway. CLAUDE_STREAM_IDLE_TIMEOUT_MSnot lengthening the 5-minute connection limit on gateways, custom base URLs and cloud providers.- MCP tools whose input schema starts with
anyOf,oneOforallOfgoing missing on third-party providers or with telemetry or feature flags turned off. - An MCP tool with an invalid input schema failing requests in the same setups — the tool is now left out and Claude is told why.
- Effort being turned off for a subagent’s model for the session if Bedrock, Google Cloud’s Agent Platform or Foundry refused its progress summary’s effort setting (regression in 2.1.327).
Code Mode
- A script stopping when a permission prompt outside auto or bypass mode, a plan approval or a question stayed open for ten minutes.
- Scripts in auto and bypass-permissions modes failing to call an MCP server that had connected while the previous script was running.
- Scripts not running in the background in auto or bypass permissions mode in a session chosen in the
claude --resumepicker. - Losing a loaded skill, hook context or a note typed with a permission approval when resuming a session in which one reply ran several scripts.
- Claude not being told that a script was cut off when the session was resumed from the picker or in a mode that asks for permission.
- Silently using
undefinedfor top-level variables that an earlier failed script never set. - Scripts not being told when a failing command’s long output was cut.
- Refusing valid scripts that start with a
{block, and scripts that use both a top-levelawaitand a top-levelreturn. - An @-mentioned small file not being attached after a script had read or written it.
- Offering
write()andedit()to read-only subagents that can’t call them. - Claude starting a turn by itself when Esc was followed by a command such as
/model. - Claude not being told to run workflows only when you ask, or the workflow size you set in
/config.
Remote Control, cloud and cross-session messaging
- Remote Control repeating the same disconnect message, and telling you to retry a session that can’t reconnect.
- A waiting permission prompt being lost, and its answer ignored, when a remote session restarted while the service was slow.
- Cloud and Remote Control sessions that just started, or have an @ in their title, being unreachable in cross-session messaging from a session with Remote Control on.
- Claude being told that cloud sessions can never reply — it now learns when a reply can come and how else to get the result.
/diff, the Changes pane of Remote Control sessions and the diff beside an edit leaving out committed work after a command moved into another repository.- Remote Control not showing a running workflow’s agents when a session run by an IDE extension or other SDK app is opened over an hour after the workflow last made progress.
- An idle Desktop or VS Code session showing as running after Remote Control is turned on.
Hangs and crashes
- Markdown rendering freezing on some very long or malformed input, and rare screen freezes.
- A crash at startup when
FORCE_COLORis set to a negative number, and a crash in secret masking when the machine is almost out of memory. - Agent hooks running past their timeout while compacting after a request was refused as too large.
Input and terminals
- Esc often being lost at a permission prompt or question when you typed straight after it — in terminals that send Alt as Esc and a key, an unused Alt or Option key now declines there.
- A paste or a terminal’s reply being typed into the prompt when a slow connection cut it in two.
- Cmd+letter being typed into the prompt in terminals that report Cmd.
- In fullscreen mode, Ctrl+click opening a link twice in GNOME Terminal, side mouse buttons acting as clicks, and the first wheel notch in the other direction being ignored after a pause.
- Shift+Enter in Konsole outside tmux and ssh, and Alt shortcuts on non-Latin keyboard layouts in terminals that speak the kitty keyboard protocol.
- Linux: pasting in other apps freezing while Claude Code is suspended with Ctrl+Z, when
xcliporwl-copyis installed, and Ctrl+V missing the clipboard image, pasting an old one or hanging. - Text copied in fullscreen mode including the quote bar, losing spaces at line wraps when a drag scrolled upward, or missing words under the “Jump to bottom” pill.
- Ctrl+C showing no exit warning while the
?shortcuts list is open. - Ctrl+R search in fullscreen mode and the
/themecolor editor picking the wrong row when keys arrive together, and keys being dropped there and in the search of/configand/plugin. - Wrong or missing colors inside GNU screen, and many background notifications cutting into the top border of open dialogs in fullscreen mode.
Slash commands
/goalremoving a skill’s Stop hook, or ending early on another hook’s result, when both had the same wording./bugsaying cancelled after the report was sent, and Esc not dismissing the transcript share question for Pro and Max users./hookslosing your place when going back, and/exportwriting a file named.txtwhen the name was left empty./terminal-setupin Zed replacing a symlinkedkeymap.jsonwith a plain file./install-github-appadding a code review workflow when asked to update the existing workflow file./pluginreopening a skill, or closing the open page, when a slow save or removal finished after you went back, and asking to uninstall the wrong plugin, or ignoring Esc, after leaving a plugin while a change was still running.- Typing and arrow keys not working on the
/pluginInstalled tab after a search with no results or a click on the tab row, and holding Enter in/pluginturning off several installed plugins. - The
/usageplugin total stopping at 2,000 tokens on 1M context models, and low token estimates in/usageand/pluginwhen no model is picked. - Esc on the Bedrock access keys step closing
/logininstead of going back. /exitand Ctrl+D describing running work differently, and/exitwarning about a workflow that had stopped.
Status, counts and telemetry
- The token counts on the spinner and for subagents counting a retried reply twice, missing replies that were not streamed, and standing still while Claude thinks.
- The error shown when usage credits run out mid-response, which named a rate limit.
- One malformed
OTEL_RESOURCE_ATTRIBUTESentry dropping the variable’s other labels from metric points, events, and spans — now only that entry is left off. - Claude calling an artifact private after reading one in a shared folder or publishing to another organization’s, and saying only named people can open one when its folder’s people can too.
- A plugin’s toast sometimes showing another plugin’s text, and a mod saved just after a turn began sometimes reloading mid-turn instead of at the turn’s end.
- A message to a background agent showing a long file path in place of a pasted image.
- The model saying background agents were still running after an SDK interrupt had stopped them, and Send now in the desktop app and SDK hosts telling Claude that a stopped web search found no results.
Self-hosted runner
- All running sessions being stopped when a proxy or web filter answered a poll with HTTP 200 and a body such as
{}— the runner now retries such a poll. - Deletion of a kept clone’s objects or refs, and of commits that existed only there, after a temporary disk error — that start now fails instead.
- Runners stopping when their token ran out on a host whose clock is slow or was set back.
- Sessions with Anthropic-managed git losing git access to GitHub after a script or hook in the session ran
claudeagain, and being unable to fetch or push where the machine’s git config setshttp.proactiveAuth. - A runner that exited without telling the server when setup failed after it had registered — it now deregisters with a reason, and an on-demand runner reports the failed start.
- Archiving or requeueing a session not taking effect while it waited for another session to prepare the same repository (
--capacityabove 1). - A session with a damaged
.claude.jsonfailing every later start under--trust-workspace false— it now starts on defaults, with MCP servers turned off there on again. - A failed start’s message losing the words that say what failed when a checkout hook or git printed a lot.
- Skipping the post-session hook when a disk stalled for a few seconds — on a dead disk a session’s end can take 25 seconds longer by default.
- A session that resumes in a new Claude config directory losing its task list, and Claude not being told when a self-hosted session starts without one of its repositories, and why.
--correlation-idofclaude --environmentnot reaching thespawn-runnerhook.- The setup wizard’s key rotation advice, which now warns that revoking a key stops the sessions running under it.
VS Code
- The permission prompt cutting off its No button, reply field or title when the panel is low or a command, URL or description is long.
- Failed or blocked tool calls showing no reason or a reason cut short, and Glob and Grep counting an error as results.
Notes
- If you use Opus 4.6 or Sonnet 4.6 on Bedrock, Google Cloud’s Agent Platform, Foundry or Claude apps gateway, interactive sessions with no permission mode now start in auto mode. Plan mode and Bash prompts also use and offer auto mode;
disableAutoModeturns both off. - If you relied on
/code-reviewfor cleanup suggestions, on most models it now reports only bugs and CLAUDE.md violations. Run/simplifyfor reuse, simplification and efficiency cleanups. - If you redact Code Mode script output with a UserPromptSubmit hook, that hook no longer runs on it. Use a PostToolUse hook with
updatedToolOutputinstead. - If managed
PostToolUsehooks log or block calls, they no longer fire for calls refused before they ran.PostToolBatchgets most of those, not marked as refused. - If you run
claude -pwithCLAUDE_CODE_FORK_SUBAGENT=0, a background subagent and those it waits for share one 10-minute wait. RaiseCLAUDE_CODE_PRINT_BG_WAIT_CEILING_MSfor longer chains. - If the sandbox can’t start in your environment, the next sandboxed Bash command is now refused with a message that the sandbox is off for the session, unless
sandbox.failIfUnavailableis set. - If your self-hosted runner settings guard Claude’s GitHub tools by name, the built-in
gh apiis now off by default;--use-anthropic-gh trueturns it on. - If you automate short
claude -pruns, loading managed-settings plugins from a git marketplace on a fresh machine may make a run exit up to 30 seconds later. - This post covers the main items only. See the GitHub release notes for the full list.