Claude docs changes for October 28th, 2026 [diff]
Executive Summary
- New
CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONSenvironment variable (v2.1.311+) letsbypassPermissionsmode runrm/rmdiron critical paths with no prompt, with startup notices and troubleshooting for when it has no effect - Claude Code 2.1.312 adds
oidc.user_id_claimfor the Claude apps gateway, click-to-send for queued messages, and caps concurrent prompt/agent hooks at 24 per session - Managed machines can now stop offering the keyless Console sign-in by deploying
forceLoginOrgUUID;forceLoginMethod: "claudeai"also removes it - Mods docs gain
Buttonvariant,roleandchildren, multi-lineInputrows, and clearersettings.readhook guidance - Claude Desktop and claude.ai settings paths are now documented as Settings > Code or Settings > Claude Code, and cloud routines can pick an effort level
New Claude Code versions
2.1.312
New features
- Added
oidc.user_id_claimto the Claude apps gateway, to identify people by an id_token claim other thansub(such as Microsoft Entra ID'soid) - Added click to send: in fullscreen, clicking the "to send now" hint under a queued message sends it right away
- [Cloud sessions] Added an effort level to the model picker when creating or editing a cloud routine
- [Cloud sessions] Added a Bash editor with colors, line numbers and a Copy button for the setup script in cloud environment settings
Existing feature improvements
- Improved the prompt's effort indicator: a mid-task change now says it may wait for your next message
- Improved WebFetch errors in cloud sessions: an unresolvable host is explained as possibly outside the allowed network domains
- Bedrock, Vertex, Foundry and Mantle now use the shorter system prompt for unrecognized model names (custom deployments, inference profile ARNs); set
CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0to keep the full prompt - Automatic model switches after a flagged message now also cover unrecognized model names; set
CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1to turn them off - Prompt and agent hooks now run at most 24 at a time per session (sub-agents included), plus at most 12 that have waited longer than their event's timeout; a
timeoutunder 1 second counts as 1 second - A subagent's
autoCompactWindownow lowers the window set byCLAUDE_CODE_AUTO_COMPACT_WINDOW - OpenTelemetry
mcp_server.namenow reports the real name of fixed-URL remote MCP servers declared by official-marketplace plugins; theclaude_code.safety_eventevent is no longer sent - Usage limit messages now say "5-hour limit" instead of "session limit"
- [Cloud sessions] Cloud sessions started from the CLI get names and branch names from the server
- [Claude Tag] Slack retry notices now name the model in use or say rate limits were hit
Major bug fixes
- Fixed managed settings and organization policy not being fetched after signing in from the VS Code extension
- Fixed Claude overwriting a very large CLAUDE.md, rules file or @-included file after seeing only its start; Write now asks Claude to read it all or edit in place
- Fixed per-directory and
paths:skills not being found after a Bashcddeeper into a git repository - Fixed a command with an invalid name stopping all commands from loading, which could leave
claude -pwith no output - Fixed the Read tool sometimes failing on PDF pages ("pdftoppm failed") during old session file clean-up
- Fixed the spinner, OpenTelemetry
effortvalues and transcripts naming a newly picked effort level before the model applied it - [Code Review] Fixed Code Review reviewing an outdated commit when GitHub delivered a push event late
- [Cloud sessions] Fixed resending an unchanged message after a failed send sometimes running it twice
- [Claude Tag] Fixed Claude in Slack not replying in channels whose environment has network access set to None, and a channel's switch/settings resetting when a private channel goes public
Claude Code changes
New Documents
None.
Changed documents
admin-setup [Source]
- Added a link to the new "Manage plugins for your organization" page for installing shared plugins through managed settings. [line 190] [Source]
agent-sdk/agent-loop [Source]
max_turnsnow explained as counting each tool-call turn plus one for the final text-only response; otherwise the run ends with anerror_max_turnsresult and no final reply. [line 40] [Source]
agent-sdk/configuration [Source]
- Switching to a model auto mode doesn't support moves the session to
defaultmode. [line 160] [Source]
agent-sdk/hosting [Source]
agent-sdk/permissions [Source]
agent-sdk/python [Source]
Taskalias for theAgenttool is accepted only when the tool is available to the session. [line 2523] [Source]allowAllUnixSocketsalso allows VM sockets (AF_VSOCK) on Linux. [line 3592] [Source]
agent-sdk/streaming-output [Source]
- Clarified what a cut-off stream delivers:
message_stopalways arrives,content_block_stoparrives except for open tool calls, andmessage_deltamay be missing after a dropped connection or non-streaming fallback. [lines 115-118] [Source]
agent-sdk/typescript [Source]
startup()keeps only the newesthook_progressmessage fromSessionStart/Setuphooks and delivers held messages once you call.query(). [lines 104-105] [Source]- Returning
allowwith more than 4,096updatedPermissionsitems is treated as invalid and denied. [line 1404] [Source] agent_idbehavior for nested subagents expanded. [line 1695] [Source]maxTurnsnow links to the definition of a turn. [line 525] [Source]
agent-view [Source]
/resumein agent view no longer documented as needing v2.1.212+. [line 319] [Source]/forksection split into subsections: prompting the copy, where the copy runs, and sessions that can't be forked. [lines 453-472] [Source]
authentication [Source]
forceLoginMethod: "claudeai"(orforceLoginOrgUUID) removes the keyless sign-in;"console"alone still allows it after a failed sign-in. [lines 126-127] [Source]- New section "Stop offering the keyless sign-in on managed machines", covering migration of keyless, environment/helper and claude.ai logins. [lines 140-149] [Source]
iam [Source]
- Same keyless sign-in changes as in authentication. [lines 126-149] [Source]
claude-code-on-the-web [Source]
- New troubleshooting section for files missing from the diff view (unpushed command-made changes, changes after the last push). [lines 263-271] [Source]
/configand sharing settings now point to Settings > Code or Settings > Claude Code (or Privacy). [lines 278-348] [Source]
web-quickstart [Source]
claude-projects [Source]
- Move to project renamed Continue in project throughout. [line 69] [Source]
- New Keep this computer awake for Remote Control option and restructured recovery steps for a lost folder connection. [lines 274-278] [Source]
cloud-environments [Source]
- Network secrets: on every plan you can add and delete secrets on your own environments; without an organization admin role, Team and Enterprise users get the shorter single-host form. [lines 98-135] [Source]
cross-session-messaging [Source]
- Messages Claude hasn't received yet appear under
Waiting to be deliveredat the end of the transcript (v2.1.309+). [lines 173-175] [Source]
desktop [Source]
- Settings paths updated to Settings > Code or Settings > Claude Code across the page. [line 68] [Source]
- Computer use is enabled under This computer > System or Capabilities, depending on the Settings menu. [lines 351-359] [Source]
- Connectors are managed via Customize > Connectors. [line 584] [Source]
- Cloud diffs can leave out unpushed changes. [line 163] [Source]
desktop-changelog [Source]
- Mirrors the desktop page updates (settings paths, computer use, connectors, cloud diff caveat). [line 68] [Source]
desktop-scheduled-tasks [Source]
- Keep computer awake is under Settings > Desktop app or Settings > This computer > System. [line 134] [Source]
env-vars [Source]
- New
CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS: runs dangerousrmcommands with no prompt inbypassPermissionsmode. [line 290] [Source] - New
CLAUDE_CODE_RESUME_SHOWS_CUT_OFF_CALLS: keeps interrupted tool calls, marked as cut off, when a resumed turn continues. [line 368] [Source]
errors [Source]
- New entry "Feedback payload too large" with the fix (set Send transcript to
no). [lines 2211-2222] [Source] - New entry "A command-line deny rule cannot be handed to a teammate intact" for split-pane agent teams; switch Teammate mode to
in-process. [lines 4691-4703] [Source] - "Couldn't send feedback" now means the service didn't confirm receipt; treat the report as not received. [line 2191] [Source]
hooks [Source]
- Clarified WorktreeCreate path resolution and rejection rules (relative paths,
./..segments, symlinks). [lines 3195-3197] [Source]
headless [Source]
- Reorganized the
mcp_server_errors/ stderr warning explanation. [lines 283-287] [Source]
large-codebases [Source]
- Local deny rules like
Read(./**/vendor/**/*)anchor at the working directory; use//-absolute paths if you start from subdirectories. [lines 157-161] [Source] - Plugins can be installed for everyone through managed settings. [line 377] [Source]
mcp [Source]
- New "Connection failures and unconfigured servers" subsection; failure details are redacted and never include the expanded server URL. [lines 277-283] [Source]
permission-modes [Source]
- New section "Turn off the critical-path removal prompt in bypassPermissions mode" with per-shell examples, a danger warning, and scope limits (not plan mode; Bash and Monitor only). [lines 959-1015] [Source]
- Startup notices explain whether the variable is active, and "Why the critical-path removal prompt stays on" lists causes (permission-prompt tools,
PermissionRequesthooks, VS Code/Desktop, deny/ask rules). [lines 982-1015] [Source] - Other sections now cross-reference the opt-out. [line 30] [Source]
permissions [Source]
- Compound commands: every subcommand must be allowed to run without a prompt (Bash and PowerShell). [line 215] [Source]
- Windows network shares: example
claude --add-dir Z:\. [line 521] [Source] - New "Where output styles, hooks, and local settings load from" section;
settings.local.jsonloads from the git repository root. [lines 708-712] [Source]
plugins/mods/events [Source]
- New "Send a request to a different model": aliases like
haikuinturn.stepneed v2.1.309+, andnext({ ...e, model: 'haiku' })redirects a request. [lines 312-317] [Source] - Reworked
settings.readhook explanation: when it first runs, how to detect origin, and what returns do. [lines 320-381] [Source]
plugins/mods/gallery [Source]
Inputrowsset to 2 or more makes a multi-line field; requires v2.1.301+, ignored in the VS Code extension, and keyboard rules apply to the terminal. [lines 245-288] [Source]
plugins/mods/interface [Source]
rowsand Up/Down behavior inInputnow scoped to the terminal. [lines 420-545] [Source]
plugins/mods/reference [Source]
Buttongainschildren(v2.1.295+),variant(primary/secondary) androle(dismiss). [lines 266-293] [Source]- New limits:
$.process.runoutput capped at 4 MiB with truncation flags, and 28,000-level nesting depth for API call arguments,nextevents and hook return values (v2.1.299+). [lines 372-394] [Source] $.ui.toastin the classic renderer can replace an older toast. [line 390] [Source]settings.readrequirements consolidated into the events table. [line 115] [Source]
plugins/mods/troubleshoot [Source]
settings.readrefusal messages may come from another mod; added step-by-step procedure to find the offending mod with--debug-file. [lines 363-385] [Source]- Returned permission rules and
envon Claude Code's own settings read are not applied. [line 340] [Source]
routines [Source]
- New step "Select a model and effort level" for cloud tasks, and the effort level is selectable in the CLI instructions box. [lines 114-139] [Source]
web-scheduled-tasks [Source]
- Same model and effort level step as in routines. [lines 114-139] [Source]
sandboxing [Source]
- A
git clonenaming an absolute,~or..path isn't covered by an entry; use a relative path. [line 319] [Source]
server-managed-settings [Source]
- Non-interactive runs apply approval-requiring settings for that run only and re-fetch them at each startup until approved interactively. [lines 287-290] [Source]
sessions [Source]
- Resumed sessions: interrupted tool calls aren't rerun (details moved to
CLAUDE_CODE_RESUME_SHOWS_CUT_OFF_CALLS). [line 45] [Source]
settings-reference [Source]
forceLoginMethod: "claudeai"also removes the keyless Console sign-in. [line 5172] [Source]CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONSis ignored in every settings file. [line 2391] [Source]- A
<<EOFheredoc counts as a redirection that keeps a Bash call sandboxed. [line 1292] [Source]
skills [Source]
- The claude-api skill's
upgradesubcommand now described as upgrading the Anthropic SDK to a new major version and updating code for breaking changes (not just Python 0.x to 1.x). [line 82] [Source]
slash-commands [Source]
sub-agents [Source]
- The per-invocation
modelparameter on the Agent tool takes only an alias (sonnet,opus,haiku,fable), not a full model ID orinherit. [line 351] [Source]
worktrees [Source]
"fresh"base falls back to localHEADwhen there's no remote ororigin/HEADcan't be fetched. [lines 159-163] [Source]- New "Worktree locks and manual removal" heading. [line 138] [Source]
- New troubleshooting entry: with
git replace, created worktrees hold the original objects' content; create the worktree manually instead. [lines 327-332] [Source]
workflows [Source]
- Save-dialog and symlink-check wording tightened. [lines 197-203] [Source]
API changes
New Documents
None.
Changed documents
about-claude/models/overview [Source]
- Models API capability fields now link to the API reference; a pointer to the full List Models response was added. [lines 67-81] [Source]
api/claude-platform-on-aws-iam-actions [Source]
- Added the
DeleteUserProfileaction (DELETE /v1/user_profiles/{id}) to the action list and route table. [line 175] [Source] - Noted that
*Filewildcards also over-matchDeleteUserProfile. [line 180] [Source]
build-with-claude/batch-processing [Source]
request_countsis now described as counting requests still processing plus requests with each result type; field references link to the API reference. [lines 704-706] [Source]
build-with-claude/claude-models-on-google-cloud/anthropic-managed [Source]
- The workspace-less base URL also applies to the user profile routes under
/v1/user_profiles. [line 218] [Source]
build-with-claude/refusals-and-fallback [Source]
fallbackcontent blocks now include atriggerobject (type: "refusal"andcategory, which can benull); the example now shows Claude Fable 5 declining a request whosefallbackslist names Claude Opus 4.8. [lines 633-651] [Source]
cli-sdks-libraries/overview [Source]
antCLI and Go SDK versions bumped from 2.3.0 to 2.4.0 (also in the CLI quickstart and Managed Agents quickstart/worker docs). [line 59] [Source]
manage-claude/cmek-coverage-claude-enterprise [Source]
- Removed the "Chat search" and "Project knowledge search" rows from the CMEK coverage list. [line 31] [Source]
manage-claude/compliance-activity-feed [Source]
organization_idisnullfor events not tied to an organization; the examples of such events (sign-in, sign-out, Compliance API calls) were removed. [line 141] [Source]
manage-claude/inference-hooks-configuration [Source]
- A
DlpWebhookStatusErrorentry can show the HTTP status returned and the request body size (such asHTTP 413ยท1.3 MB). [line 124] [Source]
manage-claude/user-profiles [Source]
- User profiles are now supported on Claude Models on Google Cloud (Anthropic managed): create, retrieve, list and update, with the
anthropic-user-profile-idheader. Profile IDs are checked on every request there. [lines 41-373] [Source]
manage-claude/workspaces [Source]
- Workspace Limited Developer role now documented as unable to open or start Claude Managed Agents sessions, open memory stores, manage deployments, or download/delete files in the Console (Console only, not API). [line 42] [Source]
managed-agents/self-hosted-sandboxes [Source]
antCLI install: Homebrew instructions moved up, and a newgo install github.com/anthropics/anthropic-cli/v2/cmd/ant@latestoption (Go 1.25+) withPATHsetup. [lines 180-215] [Source]