Claude docs changes for November 3rd, 2026 [diff]
Executive Summary
- Claude Code 2.1.317 adds an email-sharing check before web or shell calls,
onFailure: "block"for Agent SDK hook callbacks, a Claude apps gateway cost report API, and richer OTel attributes, plus many fixes. - Computer use and browser use toolsets (
computer_toolset_20260801,browser_toolset_20260801) are now documented as available on Amazon Bedrock, Claude Platform on AWS and Microsoft Foundry. Claude 5.5 models reject older computer use tool versions. - LLM gateway docs now describe how Claude Code retries after a rejected beta value, advisor tool,
effortorcontext_management, and how long each rejection is remembered. - Claude Code docs add Bedrock and Vertex AI pin-update prompts,
/btwfor subagent transcripts, cloud-session MCP elicitation cards, and Desktop SSHConnectTimeoutbehavior. - Managed Agents examples now disable
web_searchandweb_fetchthroughagent_toolset_20260401configs.
New Claude Code versions
2.1.317
New features
/schedulewarns when a new cloud scheduled task gets connectors you did not ask for- Asks before a web or shell tool call sends your account email address to the network
- Agent SDK hook callbacks support
onFailure: "block", so a failing callback blocks the action - Notices in Claude Desktop's Code tab when remote managed settings haven't loaded, and when they do
- Notice when auto mode's check has no answer for a subagent's tool call
- OTel metrics, events and spans now carry
anthropic.plan,anthropic.seat_type,anthropic.providerandanthropic.custom_base_url - Claude apps gateway admin API cost report,
GET /v1/organizations/analytics/cost, by total, developer, day or IdP group - Gateways with
admin:set keep a daily record of developers' IdP groups foradmin.identity_retention_days CLAUDE_CODE_DISABLE_NESTED_SESSION_ID=1restores the old session ID and transcript sharing forclaude -prun inside a session on a self-hosted runner- Cloud sessions get a link card for MCP server link requests
Existing feature improvements
- Clearer Microsoft Foundry error when Azure credentials can't be loaded (
cloud_credential_error) /design-syncsays up front when your design system has moved to Claude- Clearer startup error when a required Claude apps gateway is unreachable or its TLS certificate can't be verified
- The "Resume this session with" exit hint repeats the flags you launched with
- Better self-hosted runner recovery from 5xx errors and unanswered heartbeats
- SendMessage to a running background agent arrives earlier, moving its running command to the background
- Subagents and teammates on a different model use that model's saved or default effort level
- Exit skips the "Background work is running" prompt for plugin monitors
- The
plugin-authoringskill offers to help share a mod once it works - Removed the Read tool's "file unchanged since your last Read" stub
- Remote Control fork names keep their "(fork)" ending
- Cloud sessions: the organization environment dialog has a shorter layout, adding a network secret for a host that already has one is refused, and new scheduled tasks start on the organization's default environment
- Claude Tag: pasted artifact links in Slack show a simpler card with an Open button
Major bug fixes
- Fixed project and local MCP servers disconnecting after the first turn in some Desktop sessions
- Fixed auto mode failing every turn when an LLM gateway answers 403 or 422 to its safety check
- Fixed
claude auth statusexiting before a slow token refresh was saved, which could sign you out - Fixed
claude -pexiting 0 when a full disk or quota lost its transcript or~/.claude.jsonwrites - Fixed
claude -pinside cloud or self-hosted runner sessions sharing the session's ID and transcript, and generating its own title - Fixed "Prompt is too long" after automatic compaction with a very large, changed CLAUDE.md
- Fixed very long conversations losing their earlier summary when compaction had to shorten its request
- Fixed the Claude apps gateway hiding Amazon Bedrock's "Input is too long." error
- Fixed Claude treating partial Grep and Glob results as complete after a timeout
- Fixed a region or network block (HTTP 403) showing as a sign-in failure
- Fixed "unable to get local issuer certificate" behind some TLS-inspecting proxies after a root certificate renewal
- Fixed OpenTelemetry export hangs and connection floods behind proxies; exports now time out after 10 seconds
- Fixed
DISABLE_PROMPT_CACHING_OPUSandDISABLE_PROMPT_CACHING_SONNETnot applying to[1m], other-region and undated model spellings - Fixed settings edits made just after launch not taking effect
- Fixed Claude deleting the most recently updated
/team-onboardingguide when none was named - Fixed Ctrl+C at the prompt stopping all background agents
- Fixed several agent team issues with inboxes, re-created teammates and auto mode in panes
- Fixed
claude --from-prreporting "No conversations found" in projects with many sessions - Fixed self-hosted runner sessions losing turns when restarted while a plan awaited approval
- Fixed high memory use and slow startup in headless sessions with many plugins
- VSCode: fixed forked conversations dropping messages and messages lost when Claude was slow to start
- Windows: fixed every tool call refused on some volumes and scratchpad paths hitting the 260-character limit
Claude Code changes
New Documents
None.
Changed documents
admin-setup [Source]
- In WSL sessions, a
~entry in Claude Desktop'sallowedWorkspaceFoldersmeans the distribution's home directory. [line 78] [Source]
agent-sdk/typescript [Source]
- Documents what the
/coreentry exports and which root exports it leaves out. [line 51] [Source] - New startup failure reasons
sandbox_required_unavailableandorg_config_refused. [lines 2004-2028] [Source]
amazon-bedrock [Source]
- Claude Code prompts to update an older pinned model version when the account can invoke the newer one. Managed-settings pins and inference profile ARN pins are not checked. [lines 307-314] [Source]
- New sections on starting a session on a specific model and on remembered model refusals, cached for up to a day. [lines 318-329] [Source]
analytics [Source]
- Clarified the Console dashboard metrics (lines accepted, spend, active users, team insights date range) and the permissions needed to view it. [lines 191-214] [Source]
authentication [Source]
- Expanded the description of
CLAUDE_CODE_OAUTH_TOKENin the credential precedence list, including what happens when you run/login. [line 250] [Source]
claude-md [Source]
- Path-scoped rules also load when Bash views a matching file in a way that counts as a read, such as
catorhead. [line 205] [Source] - Expanded the glob pattern table (
**/*.ts,src/**/*,/*.md, and more). [lines 218-222] [Source]
cli-reference [Source]
--max-turnshas no limit by default, and with--input-format stream-jsona queued message starts a new turn with its own limit. [line 102] [Source]
cloud-environments [Source]
- Cloud sessions include Chrome for Testing at
/usr/bin/google-chrome, withCHROME_PATHset. [line 334] [Source]
commands [Source]
costs [Source]
- Explains that
/usagecomputes the dollar figure locally at list price unless an adminmodelPricingtable is in effect. [line 19] [Source] - On Pro, Max and Ultra plans, reaching your spend limit with credits left prompts you to raise or remove the limit in the CLI. [line 139] [Source]
- Console reporting now points to the Console dashboard and the Claude Code Analytics API. [lines 150-194] [Source]
desktop [Source]
- Browser pane cookies are lost when you quit Desktop in an SSH session. [line 130] [Source]
- The usage ring sits under the prompt box in both layouts and shows plan usage. [line 313] [Source]
- Explains the Worktree toggle, including with third-party providers. [lines 410-412] [Source]
- SSH: built-in SSH client usage on Windows, which SSH config settings Desktop reads, and a new
ConnectTimeouttable with default timeouts through jump hosts andProxyCommand. [lines 1095-1220] [Source] - SSH reconnect behavior: slower retries after a long outage and a message held back when organization settings haven't loaded. [lines 1378-1383] [Source]
- Two new SSH error cards:
[SERVER_NOT_RUNNABLE]/[SETUP_DIR_NOEXEC]and[DOWNLOAD_BUSY]. [lines 2300-2301] [Source]
desktop-changelog [Source]
- Updated Desktop release notes with the same SSH, Worktree and usage ring changes as the Desktop page.
env-vars [Source]
- New
CLAUDE_CODE_DISABLE_CFC_PROMPTomits the Claude in Chrome section of the system prompt while keeping the browser tools. [line 238] [Source] - New
CLAUDE_CODE_GLOB_TIMEOUT_SECONDSsets the Glob and Grep timeout (20 seconds, 60 on WSL). [line 306] [Source] - Documented
CLAUDE_CODE_OAUTH_REFRESH_TOKEN,CLAUDE_CODE_OAUTH_SCOPESandCLAUDE_CODE_OAUTH_TOKEN. [lines 334-336] [Source] - New
FORCE_HYPERLINKenables OSC 8 hyperlinks in terminals that aren't auto-detected. [line 468] [Source]
errors [Source]
- Several error entries now suggest checking that an LLM gateway forwards tool search fields (
defer_loading,tool_reference). Update to v2.1.305+ for the automatic retry. [lines 551-2568] [Source] - Usage-limit warnings (such as "used 85% of your session limit") and the
rate_limitsstatus line fields. [line 836] [Source] - The unrecognized-model stderr line adds
query_sourceand is written once per model ID per process. [lines 6201-6208] [Source]
fast-mode [Source]
- Console users can group by Speed (research preview) on the Usage and Cost pages to see fast mode spend. [line 94] [Source]
google-vertex-ai [Source]
- Same pin-update prompt as on Bedrock: accepting writes the new model ID to user settings, declining is remembered until the next default change, and managed pins aren't checked. [lines 191-196] [Source]
- New sections on starting a session on a specific model and on remembered model refusals. [lines 200-208] [Source]
hooks [Source]
- Prompt and agent hooks default to the model used for background functionality. [lines 579-3946] [Source]
iam [Source]
interactive-mode [Source]
/btwcompared with subagents and forks. New section: with a subagent transcript open,/btwanswers from that transcript. [lines 707-713] [Source]- Usage-limit automatic continue still asks for permissions, and stops after three continuations with a notification. Use
/rate-limit-optionsto choose next steps. [lines 754-756] [Source]
jetbrains [Source]
- Note that the JetBrains Marketplace is third-party software that Anthropic doesn't endorse, maintain or audit. [line 38] [Source]
llm-gateway-protocol [Source]
- Retry behavior for rejected capabilities: advisor tool,
output_config.effortandcontext_management. [lines 236-241] [Source] - New "Retry without a rejected beta value" section: which error message forms (
Unexpected value(s)or abeta_header:token) trigger the retry. [lines 249-272] [Source] - New "How long a rejected beta value stays out" table. [lines 249-272] [Source]
- Tool search retry on v2.1.305+ after a
400or422. [lines 275-277] [Source]
llm-gateway-rollout [Source]
ANTHROPIC_MODELandANTHROPIC_DEFAULT_HAIKU_MODELcontrol which model name is requested for the main session and background traffic. [line 165] [Source]
managed-settings [Source]
useAutoModeDuringPlan,syncClaudeAiSkillsandsyncClaudeAiPluginsare deny-only cross-source keys. Afalsefrom any admin, user or local source turns the behavior off. [line 170] [Source]
mcp [Source]
- How Claude Code reconnects after an MCP OAuth server returns
401, with and without a refresh token. [line 951] [Source] - New section on answering MCP elicitation requests in a cloud session, in form or URL mode. [lines 1657-1667] [Source]
- Tool search through a gateway: expected forwarding of
defer_loadingfields and the v2.1.305+ automatic retry. [lines 1830-1832] [Source]
memory [Source]
- Same path-scoped rule loading and glob table updates as claude-md. [lines 205-222] [Source]
model-config [Source]
network-config [Source]
- Which proxy and sign-in variables Claude Code reads from settings files in Claude Desktop sessions, and how conflicts with the launch environment are handled. [lines 130-135] [Source]
permission-modes [Source]
- New answer option "No, and block reads outside the working directories from now on", which sets
permissions.blockReadsOutsideWorkingDirectoriestotrue. [line 531] [Source]
plugins/mods/api [Source]
- Timers stop on module reload.
next.signalis anAbortSignalfor long-running work in a hook. [line 255] [Source]
plugins/mods/interface [Source]
remote-control [Source]
- Clarified server mode behavior and the QR code toggle (spacebar) in
claude remote-control. [lines 50-151] [Source]
routines [Source]
sandboxing [Source]
- New limitation: services listed in
network.allowMachLookupon macOS can act outside the sandbox. [line 1073] [Source]
settings-reference [Source]
useAutoModeDuringPlan: afalsein any source turns it off, andtrueequals unset. [lines 924-926] [Source]- To lift
permissions.blockReadsOutsideWorkingDirectories, remove the key from every settings file and restart sessions. [line 1063] [Source] network.allowMachLookupnow describes XPC/Mach lookups, such as the iOS Simulator or Playwright, and the sandbox caveat. [line 2073] [Source]- Claude.ai sign-in variables are ignored in
envfrom project settings. [line 2422] [Source] - Idle compaction doesn't run where feature flags can't be fetched on versions before v2.1.314. [line 2491] [Source]
settings [Source]
- Removing
permissions.blockReadsOutsideWorkingDirectoriesmid-session doesn't lift the block in running sessions. [line 183] [Source] useAutoModeDuringPlanadded to the restrictive cross-source keys table. [line 312] [Source]
sub-agents [Source]
- Use
/btwinstead of a subagent for questions about your existing conversation. [line 1047] [Source] - With a subagent transcript open, messages and skills go to that agent and
/btwanswers from its transcript (v2.1.314+). [line 1241] [Source]
tools-reference [Source]
- EnterWorktree
pathmay name a worktree of a repository nested inside the current one. [line 302] [Source]
web-scheduled-tasks [Source]
API changes
New Documents
None.
Changed documents
agents-and-tools/tool-use/browser-use-sdk [Source]
- Now in beta on Claude Platform on AWS, Amazon Bedrock and Microsoft Foundry. [lines 23-26] [Source]
- On other platforms, create that platform's SDK client instead of the Claude API one. [line 86] [Source]
- Availability table split by platform. [line 1331] [Source]
agents-and-tools/tool-use/browser-use-tool [Source]
- Now stable on Claude Platform on AWS, Amazon Bedrock and Microsoft Foundry. Add one
browser_toolset_20260801entry, with noname, totools. [lines 26-44] [Source] - Availability table split by platform. [line 1595] [Source]
agents-and-tools/tool-use/computer-use-tool [Source]
- Computer use is stable on Claude Platform on AWS, Amazon Bedrock and Microsoft Foundry. [lines 32-36] [Source]
- Claude 5.5 and later models use
computer_toolset_20260801and return a 400 for any computer use tool version they don't accept. [line 38] [Source] - New guidance for upgrading a
computer_20251124integration before moving to a model that doesn't accept it. [lines 2070-2130] [Source] - Earlier tool versions and beta headers reorganized. On Bedrock, Opus 5.5, Sonnet 5.5 and Haiku 5.5 also accept
computer_20251124withcomputer-use-2025-11-24. [lines 2134-2137] [Source]
api/errors [Source]
- Claude 5.5 models return a 400 for unsupported computer use tool versions. The same message appears for any unsupported Anthropic-defined tool type. [lines 525-533] [Source]
api/ip-addresses [Source]
- Added inbound range
153.61.224.0/22and outbound range153.61.208.0/20. [lines 19-33] [Source]
build-with-claude/claude-in-amazon-bedrock [Source]
build-with-claude/claude-on-amazon-bedrock-legacy [Source]
- New "Computer use and browser use toolsets on Bedrock" section: available through InvokeModel on several models, with no beta header. [line 724] [Source]
build-with-claude/claude-platform-on-aws [Source]
build-with-claude/prompt-engineering/prompting-claude-opus-4-8 [Source]
- Notes support for
computer_toolset_20260801,computer_20251124and the browser use tool. [line 158] [Source]
build-with-claude/prompt-engineering/prompting-claude-sonnet-5 [Source]
- Notes support for
computer_toolset_20260801,computer_20251124and the browser use tool. [line 152] [Source]
cli-sdks-libraries/cli/apply [Source]
- Agent YAML examples disable
web_searchandweb_fetchthrough toolconfigs. [lines 112-204] [Source]
cli-sdks-libraries/cli/using [Source]
- Explains inlining a file into a string field of a stdin document with an
@-prefixed, quoted path, and updated structured flag examples. [lines 151-160] [Source]
manage-claude/cmek-coverage-claude-enterprise [Source]
- Claude Security (beta) and cloud scheduled tasks (formerly routines) are listed as off under CMEK. [line 42] [Source]
manage-claude/compliance-content-data [Source]
- Message example adds
claude_file_entries. Documentsfiles,generated_filesandartifacts, and says to pick the endpoint by ID prefix. [lines 147-178] [Source] - File downloads always use the RFC 5987
filename*form inContent-Disposition. curl's-Jdoesn't read it, so use--output. [lines 195-208] [Source]
manage-claude/compliance-errors [Source]
- New "Invalid file ID" error, caused by sending an ID to an endpoint that expects a different prefix. [line 135] [Source]
manage-claude/compliance-faq [Source]
- Local session capture covers only requests that reach the Claude API. No endpoint lists Cowork scheduled tasks or Claude Code cloud scheduled tasks. [lines 73-112] [Source]
manage-claude/compliance-sessions [Source]
- Claude Code cloud sessions, including cloud scheduled tasks, are listed as captured sessions. [line 36] [Source]
manage-claude/inference-hooks-configuration [Source]
managed-agents/github [Source]
- Agent creation examples in every SDK now add an
agent_toolset_20260401withweb_searchandweb_fetchdisabled, and use Opus 5.5. [lines 66-294] [Source]
managed-agents/multiagent-orchestration [Source]
- The contract review example now disables web search and web fetch, since reviewing contracts doesn't use the web. [lines 849-965] [Source]
managed-agents/quickstart [Source]
- Quickstart examples now configure the agent toolset with web search and web fetch disabled.
managed-agents/self-hosted-sandboxes-custom-tools [Source]
- Agent examples in all SDKs now disable
web_searchandweb_fetchalongside custom tools. [lines 259-630] [Source]
managed-agents/tools [Source]
- The custom tool example now creates an agent with one custom tool,
get_weather, and no built-in tools. PHP and Ruby examples use the Opus 5.5 model constant. [lines 189-462] [Source]
models/haiku-5-5/whats-new-haiku-5-5 [Source]
- The change table adds the browser use tool and a breaking change:
computer_20250124now returns an error, so usecomputer_toolset_20260801(computer_20251124also works on Bedrock). [lines 17-34] [Source]
models/opus-5-5/overview [Source]
- Summarizes Opus 5.5 as built for long-running agentic coding and knowledge work, and lists breaking changes compared with Opus 5. [line 21] [Source]
models/opus-5-5/migration-guide, whats-new-opus-5-5; models/sonnet-5-5/migration-guide, overview, whats-new-sonnet-5-5; models/haiku-5-5/migration-guide [Source]
- Migration guides and what's new pages for the 5.5 models gain computer use toolset migration notes (
computer_toolset_20260801).