Claude docs changes for October 23rd, 2026 [diff]

Executive Summary

  • Claude Code 2.1.309 ships a batch of sandbox and permission-check hardening fixes, and claude ultrareview now requires workspace trust (--trust-workspace for scripts and CI).
  • /rewind now also undoes Bash-command changes to files Claude has already edited with its file tools. Checkpointing, glossary and Agent SDK docs were updated to match.
  • New API docs split "Claude Models on Google Cloud" into Anthropic-managed and Google-managed offerings, with an IAM permissions reference.
  • A new set of customer-managed encryption key (CMEK) docs covers how it works, coverage, key lifecycle, the Admin API and troubleshooting.
  • Claude Code docs add OpenTelemetry safety_event telemetry, MCP idempotency hints and stdio reconnect behavior, and new VS Code and Desktop troubleshooting sections.
  • Note: most files changed only because in-site links were rewritten (relative .md paths vs. /docs/en/...), and API reference pages were re-indented. These are not listed below.

New Claude Code versions

2.1.309

New features

  • claude ultrareview requires workspace trust: it asks on a terminal and otherwise refuses. Scripts pass --trust-workspace, or set CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1
  • Skills from an --add-dir folder that share a name with an existing skill are now offered as <folder>:<name>
  • Failed shell/hook environments now give clearer errors: full-disk and quota messages name the folder to free up, and hooks-module nesting errors name the file
  • Bundled dataviz skill accepts APCA as a text-contrast measure and adds an apca() function to its validator scripts
  • /claude-api skill builds Managed Agents subagent teams on multiagent_20261001 with dynamic workflows on by default, and offers API tracing alongside SDK tracing for OpenTelemetry
  • [VSCode] Warning in the log that explains what to do when dictation or extension requests fail certificate verification (for example behind a TLS-inspecting proxy)
  • [VSCode] Ctrl+J in the chat input can be left to VS Code and the OS by setting "ctrl+j": null in the Chat context of ~/.claude/keybindings.json

Existing feature improvements

  • Faster git commands on Windows, and git errors show in English on every platform
  • sed -i now gets the standard Bash prompt and runs as written, so it is no longer restored by /rewind
  • Bypass-permissions mode in the Desktop app and VS Code denies dangerous rm/rmdir immediately with a rewrite hint instead of prompting
  • Commands excluded from the sandbox ask before writing to write-list paths when the sandbox is enabled but cannot run
  • Worktree creation no longer lets git fetch objects on demand (partial clones fetch first)
  • Nested CLAUDE.md files over the size limit now load only their start
  • [VSCode] Faster Edit and Write in the chat panel (0.75 s wait for new problems, down from 1 to 1.5 s)
  • [VSCode] Long slash-command messages such as /goal fold behind Show more
  • [Cloud sessions] The Changes pane says whether it lists only pushed commits or only Claude's edits
  • Self-hosted runner: sessions that move to a new runner show the interrupted tool call as outcome unknown

Major bug fixes

  • Sandbox: shell startup files no longer run outside the sandbox, macOS denyWrite globs apply through symlinks, and sandboxed commands can't modify the --mcp-config file
  • Permission checks: fixed several Bash and PowerShell bypasses (variables set by commands with failing redirects, wrapped cd before git, Edit(path) deny rules with subshell redirects, Set-Location/cd with extra arguments)
  • Worktree-isolated sessions no longer accept locale-prefixed or zsh arithmetic commands that the guard should refuse
  • PowerShell hooks now block on exit code 2, and PreToolUse/PermissionRequest PowerShell hooks no longer fail open without Git for Windows
  • Fixed repeated compaction caused by large nested CLAUDE.md files
  • Fixed /rewind showing "No code changes" for turns changed via /code-review --fix, subagents or shell commands
  • Split-pane agent teammates now inherit the leader's --tools and deny rules
  • Fixed Claude notifying unrelated GitHub users by writing role names or placeholders as mentions
  • Fixed macOS repeatedly opening the "Install Command Line Developer Tools" dialog
  • [Claude Tag] Fixed Slack drafts posting to a DM instead of the approved channel, Allow/Deny cards doing nothing after the session stopped, and untagged messages being ignored after a workspace moved organizations
  • [Code Review] Fixed reviews showing no check on the PR after failed creation, and pushes queued behind a review restarting it after @claude review stop

Claude Code changes

Changed documents

accessibility [Source]

  • New troubleshooting entry: project settings can't turn on screen reader mode. Claude Code ignores axScreenReader and CLAUDE_AX_SCREEN_READER in project .claude/settings*.json files; set "axScreenReader": false in user settings to silence the startup line.

agent-sdk/file-checkpointing [Source]

  • Rewinding now undoes Bash-command changes to files the main conversation already edited with Write or Edit.

agent-sdk/observability [Source]

  • Events that become a per-user audit trail for SIEM forwarding (tool_decision, tool_result, mcp_server_connection, permission_mode_changed).

agent-sdk/skills [Source]

  • Clarified the skills list: own skills use the directory name, plugin skills use plugin:skill. The same change is in agent-sdk/slash-commands.

agent-sdk/typescript [Source]

  • Added the allowNestedUserNamespaces sandbox option.

agent-teams [Source]

  • Teammates inherit the lead's --disallowedTools and --tools (v2.1.309+).
  • Added a subagents vs. agent teams comparison sentence.

checkpointing [Source]

  • Replaced "Bash changes are not tracked" with new behavior: /rewind undoes Bash changes to files Claude already edited with its file tools. Files changed only by Bash are still untracked; use Git.
  • The rewind menu counts changes to already-edited files regardless of what made them.

claude-apps-gateway [Source]

  • New section on how hook and skill keys in parent settings take effect (allowedHttpHookUrls, httpHookAllowedEnvVars), requiring v2.1.309+.

claude-apps-gateway-deploy [Source]

  • New section on retrying unanswered requests sooner with CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS.

claude-code-on-the-web [Source]

  • /clear is now supported in cloud sessions and starts a fresh conversation in the same session.

claude-md [Source]

  • Per-file size limits now apply to on-demand files (subdirectory CLAUDE.md and rules): 40,000 characters (200K window) or 150,000 to 200,000 (1M window). Larger files load only their start.
  • New troubleshooting entry "Claude ignores the end of my subdirectory CLAUDE.md". The same content was added to memory.

claude-projects [Source]

  • Reworked moving a cloud session into a project: Move to project now has a submenu with Continue as new project or an existing project.

cloud-environments [Source]

  • "Add a setup script" is now its own section.

code-review-pull-requests [Source]

  • New section "When a push keeps the earlier result": Code Review may keep an earlier result when the PR's diff is unchanged and no Important thread was resolved. The check title is Code review skipped — no new changes.
  • Scripts parsing severity counts should handle the no new changes title first.

costs [Source]

  • Explained the recommendation share: it shows how much usage a row applies to, not the savings, and shares can exceed 100%.

desktop [Source]

  • New "Allow jump hosts" section for sshHostAllowlist. Jump hosts from ProxyJump or ssh -W ProxyCommand must be listed (Desktop v3.1623.0+), with a step-by-step way to find the names. The same content was added to desktop-changelog.
  • /config opens Settings > Claude Code, and a disk-space error message was renamed.

env-vars [Source]

  • Added CLAUDE_CODE_WARN_DISK_FREE_GIB and CLAUDE_CODE_WARN_MEMORY_FREE_GIB (v2.1.304+).

errors [Source]

  • New entries: MCP stored credentials unreadable (locked macOS keychain), MCP connection timed out during sleep or clock change, MCP connection closed before the tool call was answered, cached plugin MCP connection failure, and claude ultrareview workspace-not-trusted errors.
  • Rewrote disk-full and quota messages and fixes, now naming the volume and CLAUDE_CODE_TMPDIR.
  • Renamed the sandbox mask file warning ("left behind at write-deny paths") with updated cleanup steps.
  • Reworked feedback error handling and noted that CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS applies to gateway sessions from v2.1.307.

features-overview [Source]

  • Clarified that only CLAUDE.md files loaded at session start load in full.

glossary [Source]

  • Checkpoint entry updated for the new Bash-change rewind behavior.

hooks [Source]

  • Reworked guidance for hooks that fail without Git Bash: exec-form hooks for .exe programs, and "onFailure": "block" plus "shell": "powershell" for PowerShell guard hooks. The message text changed in v2.1.309.

ide-integrations [Source]

  • The same VS Code additions as vs-code (below).

managed-mcp [Source]

  • Reworded the upgrade warning for allowedMcpServers users: add deniedMcpServers entries before upgrading. ${VAR} expansion in env no longer counts toward the check.

managed-settings [Source]

  • Managed settings files may be symlinks. Claude Code follows them and warns in /status and claude doctor when the target is outside the system directory.

mcp [Source]

  • New "Stdio servers that exit or fail to start" table covering reconnect behavior by run mode.
  • New section "Mark tools that are safe to call twice" for idempotentHint and readOnlyHint annotations.
  • Revised the retry description for dropped remote servers and 401 refresh.

monitoring-usage [Source]

  • New claude_code.safety_event with attributes, categories and an example, plus which calls don't emit one.
  • ccr.session.id identifies cloud sessions, and session.id is replaced by /clear.

network-config [Source]

  • Removed the note about the 5 to 6 minute body idle timeout on other connections.

permissions [Source]

  • Write(path), NotebookEdit(path) and Glob(path) rules never match. Use Edit(path) and Read(path) instead.
  • Reorganized path-pattern docs: anchoring, depth, special characters and ! negation.

plugins/mods/interface [Source]

  • Users can collapse the mod band to one row. Mods can't expand it or detect it, so don't make the band the only place to ask something.

plugins/mods/reference [Source]

  • Svg props listed (source, alt, width, height, isInteractive); new Button rules for hotkeys; more band-supported props (position offsets, alignSelf, borderDimColor, hover).

plugins/mods/troubleshoot [Source]

  • New entries for "engine ran out of stack" hooks-module errors and "plugin panel hidden", and an expanded hotkey-not-working checklist.

routines [Source]

  • New "Permission prompts during a run" section and an "Artifacts in Accept edits mode" subsection. The same content is in web-scheduled-tasks.

security-guidance [Source]

  • Reviews now default to "a Claude Opus model" instead of a pinned version.

self-hosted-environments-configuration [Source]

  • Wrapper stdin guidance changed to file descriptor 9. Runners on v2.1.309+ reserve fds 4 and 5.
  • New "Exit the hook without waiting for the runner" section (--hook-timeout kills the process tree).
  • The Stop hook script is hardened (core.fsmonitor and hooks path disabled, --ignore-submodules=dirty), with new submodule behavior notes and replacement instructions for older copies.

self-hosted-environments-deploy [Source]

  • --base-dir /workspace added to deployment examples.
  • New "Let running turns finish before a restart" section (--drain-wait-sec).
  • New troubleshooting for malformed poll responses from intercepting proxies and for an unwritable base directory.

self-hosted-environments-reference [Source]

  • New section on joining a session's metrics to the runner's by session_id.

sessions [Source]

  • Documented the "Resume this conversation?" dialog and its usage estimate.

settings-reference [Source]

  • Unix socket sandbox setting now also allows AF_VSOCK VM sockets on Linux and WSL2.

skills [Source]

  • New section on running a same-named skill from an added directory (/cli:verify) and on managed settings that lock hooks to plugins. The same content is in slash-commands.

sub-agents [Source]

  • Subagent permission mode inheritance now lists only default and dontAsk (plan removed).

tools-reference [Source]

  • New "Task-tracking tools in subagents and agent teams" section.

ultrareview [Source]

  • New "Trust the folder before the review runs" section (claude ultrareview 1234 --trust-workspace) and a revised exit-code list (130 for Ctrl-C).

voice-dictation [Source]

  • VS Code dictation stops after 15 seconds without recognized speech or at a 2-minute limit.

vs-code [Source]

  • New sections: "The conversation is still open somewhere else" (Open here anyway), "Ctrl+J adds a new line in the prompt box" (unbind with keybindings.json), "The Claude Code panel is blank or couldn't load", and "Dictation or the Web tab fails behind a TLS-inspecting proxy" (NODE_EXTRA_CA_CERTS).

whats-new [Source]

  • Weekly entries gained date ranges (September 7–11 back to March 23–27, 2026).

worktrees [Source]

  • New troubleshooting entry for worktree creation failing in a partial clone.

API changes

New Documents

api/claude-models-on-google-cloud-anthropic-managed-iam-permissions [Source]

Reference of Google Cloud IAM permissions (claude.googleapis.com), the three predefined roles and audit log types for Claude Models on Google Cloud (Anthropic managed).

build-with-claude/claude-models-on-google-cloud/anthropic-managed [Source]

Setup guide for the Claude API run by Anthropic and reached through a Google Cloud project, with Google Cloud handling access and billing.

build-with-claude/claude-models-on-google-cloud/google-managed [Source]

Guide for Claude operated by Google as part of Google Cloud's Agent Platform, called with Anthropic's client SDKs (prerequisites, gcloud, SDK usage).

build-with-claude/claude-models-on-google-cloud/overview [Source]

Compares the two Google Cloud offerings and explains how to tell which one you use.

manage-claude/cmek-admin-api [Source]

Create, validate and attach a customer-managed encryption key using three Admin API calls.

manage-claude/cmek-aws-kms-workspaces [Source]

How to limit an AWS KMS key to specific workspaces with a key policy condition.

manage-claude/cmek-coverage-claude-enterprise [Source]

Lists which Claude Enterprise data a CMEK key encrypts, what turns off or changes, and what stays under Anthropic's key.

manage-claude/cmek-coverage-claude-platform [Source]

The same coverage reference for Claude Platform data.

manage-claude/cmek-data-retention [Source]

Explains why a workspace with a key retains data, what that means for zero data retention (ZDR), and what Anthropic can preserve outside your key.

manage-claude/cmek-how-it-works [Source]

Explains envelope encryption: your KMS key wraps Anthropic's data keys, the key never leaves your KMS, and you can log each call. Also covers when it takes effect.

manage-claude/cmek-key-lifecycle [Source]

What rotating, disabling or deleting your key does to your data, how soon, and whether it can be undone.

manage-claude/cmek-troubleshooting [Source]

Causes and fixes for CMEK errors, failed verification and unexpected log entries.

Changed documents

about-claude/models/model-ids-and-versions [Source]

  • "Google Cloud" renamed to "Claude Models on Google Cloud (Google managed)" in model ID format notes. The rename also appears in pricing and api/errors.

about-claude/models/optimizing-for-cost-and-intelligence [Source]

  • New measurements (October 20, 2026): a max_tokens: 0 request refreshes an existing cache entry on Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. The Haiku 5.5 keep-alive guidance was rewritten.
  • Removed the cost-of-one-broken-turn table, and added October 2026 cache-break test results (tokens read and written).

about-claude/use-case-guides/commerce-agents [Source]

  • Guide intro now covers building a shopping agent and a merchant agent, with runnable examples for retail, travel, telecom and entertainment, plus a Claude Code plugin that scaffolds the design.

about-claude/use-case-guides/customer-support-chat [Source]

  • Go example now imports anthropic-sdk-go/v2.

agents-and-tools/agent-skills/claude-api-skill [Source]

  • Platform availability for Managed Agents is no longer listed in the page. It now points to which platforms are supported.

api/errors [Source]

  • Clarified how to pass thinking and redacted_thinking blocks back with tool use, and updated the per-SDK maxRetries option names.

api/migration-2026-10-21 [Source]

  • Removed the Text Completions API removal section (CLI and per-SDK before/after examples) from the migration guide.

api/compliance/activities/list [Source]

  • New activity types: account_migration_started, account_migration_completed, account_migration_failed, claude_code_review_promo_enrolled, claude_skill_auto_enabled and claude_skill_auto_disabled. The Claude Code Security scan project description now includes pausing and unpausing scheduled scanning.

api/messages [Source]

  • The reference was restructured so message params nest under UserMessageParam and MessageContentParam. The same restructuring applies to the beta, CLI and per-language SDK reference pages (api/beta/*, api/python/*, api/typescript/*, and so on).
  • Removed the note about a "system" role from the messages system prompt guidance.

api/claude-code/routines-fire [Source]

  • Response examples are now marked as json code blocks.