Claude docs changes for October 20th, 2026 [diff]

Executive Summary

  • Claude Code 2.1.304 adds /auto-mode-setup, a guided setup that teaches auto mode about your environment. It also adds a warning before resuming an inactive conversation that would use a noticeable share of your 5-hour usage limit.
  • claude -p and Agent SDK runs with no permission mode set now start in Manual mode (v2.1.302+). Calls that need approval are denied unless you add allow rules or pass --permission-mode auto. Auto mode also gains a configurable classifier model (CLAUDE_CODE_AUTO_MODE_MODEL).
  • Routines were renamed "scheduled tasks" across the Claude Code docs. The Desktop app now has one Scheduled tasks page covering both local and cloud tasks, with step-by-step setup for local tasks.
  • A recursive rm on a mounted drive (/Volumes, /media, WSL /mnt/c) now gets a safety prompt. CLAUDE_CODE_DISABLE_MOUNTED_DRIVE_RM_PROMPT=1 turns it off.
  • The API web fetch tool gets a url_sources parameter for limiting which URLs Claude may fetch. The Compliance API gains a user preferences endpoint and about 8 new activity types.

New Claude Code versions

2.1.304

New features

  • Added /auto-mode-setup, a guided setup that reads your project and recent sessions to teach auto mode about your environment. It is now also available on Bedrock, Vertex, Foundry, Claude apps gateway and telemetry-disabled sessions.
  • Added /claude-api managed-agents-secure, which compares a Managed Agents agent's settings with Anthropic's published security guide without changing anything
  • Added a warning on Pro and Max plans when resuming an inactive conversation would use a noticeable share of your 5-hour usage limit
  • Mods: added decider: "person" to a tool.check answer, added e.isReadOnly to the tool.check event, and added a transcript line when an earlier-running plugin skips one of your mod's hooks
  • Added the names of earlier-called tools to the compaction summary in the transcript view (Ctrl+O)
  • Code Review: owners can choose Quick, Standard (the default) or Deep review levels per organization or per repository, and can review every repository in a GitHub organization, including ones added later
  • [VSCode] Added mod drawing in the chat panel for sessions on Bedrock, Vertex AI and Foundry, behind an LLM gateway, or with telemetry turned off

Existing feature improvements

  • Tool search is now on by default behind a custom ANTHROPIC_BASE_URL; a 400 or 422 from the gateway is retried without it
  • CLAUDE_CODE_AUTO_COMPACT_WINDOW now accepts k and M suffixes, and warns when a value is out of range or ignored
  • Compaction now sends Claude the skill list again afterwards, and /reload-skills sends only new or changed skills
  • Skills from one source can register at most 1,024 hooks and 16,384 matcher characters per session
  • Mods' settings.read hooks now receive each settings layer and can rewrite the merged settings, so a managed mod can make an organization's setting a ceiling
  • /usage now says when a rate limit on usage data ends
  • In cloud sessions, a new conversation's first turn no longer waits for local (stdio) MCP servers of plugins synced from claude.ai, unless the server sets alwaysLoad: true
  • Better responsiveness with many concurrent subagents, lower memory use in long conversations, and faster first request when resuming large compacted sessions in SDK and headless mode
  • Smaller session files in auto mode, because the safety classifier's context is no longer stored with every request
  • The /artifacts delete confirmation now shows numbered choices
  • The /claude-api skill uses the ant CLI on Claude Platform on AWS
  • [VSCode] Background sub-agents now show as one status line under their Agent card instead of streaming every step into the chat

Major bug fixes

  • Fixed rm -r of a mounted drive or top-level mount folder (/Volumes, /media, /mnt/c) not getting the safety prompt
  • Fixed claude --worktree removing the worktree and branch without asking at exit when HEAD was detached or on another branch, which could lose commits
  • Fixed sandboxed commands failing on Linux when a sandbox.filesystem.denyRead glob contains brackets that aren't a character class
  • Fixed CLAUDE_CODE_TRANSCRIPT_LOCAL_GC dropping messages from the saved transcript when the disk filled up
  • Fixed a plugin keeping a headless run going past --max-turns, Stop, or ended input
  • Fixed context_management: Extra inputs are not permitted errors behind LLM gateways
  • Fixed OpenTelemetry settings from server-managed settings being ignored for the session after first-run sign-in or /logout
  • Fixed a PreToolUse hook's deny reason being replaced by a timeout error when an Agent SDK callback hook on the same call timed out
  • Fixed "Context limit reached" / "Prompt is too long" errors after stopping a turn during compaction
  • Fixed subagents sometimes forgetting their most recent work when messaged or resumed right after finishing
  • Fixed Remote Control failing to start from Claude Desktop with DISABLE_TELEMETRY or DO_NOT_TRACK set on a paid plan
  • Fixed unsandboxed Bash commands in Claude Desktop Code tab sessions not seeing environment variables exported in shell startup files
  • Fixed agent teams: approving a shutdown request from the main session no longer exits the whole CLI
  • [VSCode] Fixed sessions failing to start with "Subprocess initialization did not complete within 60000ms" because of slow SessionStart hooks
  • [Code Review] Fixed pull requests opened right after Code Review was turned on being silently skipped

Claude Code changes

New Documents

None.

Changed documents

Automate work with scheduled tasks (routines) [Source]

  • Routines were renamed "scheduled tasks" (formerly routines). The page keeps the same cloud schedule, API and GitHub triggers. [lines 1-7] [Source]
  • web-scheduled-tasks.md received the same changes.

Desktop scheduled tasks [Source]

  • The Scheduled tasks page now lists both local and cloud tasks. [lines 7-12] [Source]
  • Added a step-by-step guide to creating a local task. It covers the Only on this computer switch, picking a local folder, the schedule, and the permission mode and model. [lines 36-95] [Source]
  • Added a Manual schedule option, which runs only when you click Run now. [line 102] [Source]

Desktop and Desktop changelog [Source]

  • Roughly 270 lines were added to the Desktop changelog (mirrored in desktop.md) with new Claude Desktop releases.

Env vars [Source]

  • API_TIMEOUT_MS now describes a per-attempt timeout for response headers, and how it applies to non-streaming requests. [line 179] [Source]
  • CLAUDE_CODE_AUTO_COMPACT_WINDOW now accepts suffixed values such as 200k (v2.1.304+). [line 211] [Source]
  • Added CLAUDE_CODE_AUTO_MODE_MODEL and CLAUDE_CODE_DISABLE_MOUNTED_DRIVE_RM_PROMPT.

Monitoring usage [Source]

  • Added CLAUDE_CODE_OTEL_CONTENT_MAX_LENGTH to limit content-bearing attributes. [line 124] [Source]
  • Added message.uuid and client_request_id attributes to events, with correlation details. [lines 745-747] [Source]
  • Tool decisions can now be cancelled. Tool decision events carry a tool_source attribute listing the decision sources. [lines 305-308] [lines 964-980] [Source]

Mods API [Source]

  • Added a section on calls and background work after the user interrupts a turn. It covers which calls are refused, how to retry them from a turn.complete hook, and what happens to queued prompts, commands, scheduled jobs and monitors (v2.1.304+). [lines 217-299] [Source]

Permission modes [Source]

  • claude -p and Agent SDK runs with no mode set now start in Manual mode (v2.1.302+). Calls needing approval are denied unless allowed by a rule, hook, callback or --permission-prompt-tool. The page gives CLI, TypeScript and Python examples for auto mode. [lines 81-130] [Source]
  • Added a "Default classifier model" section: Sonnet 5 by default, with a table of overrides. Also added how to set CLAUDE_CODE_AUTO_MODE_MODEL, when your model is not used, and how to check the model in the debug log. [lines 383-441] [Source]
  • Added "The mounted drive prompt": which paths trigger it, how to turn it off, and which removals still prompt. [lines 824-870] [Source]
  • Clarified that a mod can still approve calls in dontAsk mode, and that mod approval doesn't bypass your rules and hooks. [line 19] [Source]

Sub-agents [Source]

  • Updated for the new background sub-agent display and related behavior (about 84 lines changed).

Tools reference [Source]

  • Updated tool behavior details (about 52 lines changed).

Other Claude Code pages

  • About 70 other pages received smaller edits: routines renaming links, errors, ide-integrations, vs-code, corporate-launcher, claude-projects, cloud-environments and settings-reference. A new /auto-mode-setup command is covered in commands and slash-commands.

API changes

New Documents

List organization members' personal preferences [Source]

Compliance API endpoint GET /v1/compliance/organizations/{org_uuid}/user_preferences. It returns each current member's "Instructions for Claude" and Cowork "Global instructions", oldest member first. It is paginated with limit, page and next_page, and a page may be short or empty without being the last. It has an include_unset option. The parent page user_preferences.md is also new, and organizations.md links to it.

Changed documents

Build a tool-using agent [Source]

  • Large rewrite: roughly 800 lines of cURL shell scripting were removed.

Compliance API (activities) [Source]

  • Eight new activity types (525 to 533), also in activities.md and activities/list.md: claude_plugin_served_version_updated, claude_plugin_version_created, org_hipaa_claude_code_cowork_restrictions_deferred, org_hipaa_claude_code_cowork_restrictions_enabled, org_library_submission_created, org_library_submission_resolved, org_library_submission_withdrawn and org_member_avatar_removed.

MCP connector [Source]

  • The pinned-toolset example was simplified. Instead of shell scripts that capture and re-inject the mcp_tool_listing, it now shows two explicit requests, with a pinned echo tool in the second. [lines 575-700] [Source]

Multiagent orchestration [Source]

  • Examples and text were reworked (about 70 lines changed). The Managed Agents quickstart and github pages were also updated.

Vault credentials (all SDK and CLI variants) [Source]

  • Reference text was rewritten (same changes in the TypeScript, Python, Ruby, Java, Go, C#, PHP and CLI variants).
  • Titles now read "Create a Credential", "Retrieve a Credential", and so on.
  • expires_at must be in the future. When you send a new access_token, send its expires_at too, or the token keeps the old expiry.
  • mcp_oauth client secret updates must match the credential's current token_endpoint_auth type.
  • An empty scope string clears the stored scope.
  • Secret host lists accept up to 16 entries. A *.example.com entry matches subdomains only, not example.com itself.

Web fetch tool [Source]

  • Added an optional url_sources parameter to turn off any of three allowed URL sources: user_input, client_tool_results and server_tool_results. Each takes all, none, or an only/except tool filter, and it works with web_fetch_20250910 and later versions. It is offered as an anti-exfiltration measure. [lines 25-34] [lines 418-422] [lines 641-685] [Source]
  • The url_not_in_prior_context error also covers URLs that appear only in sources url_sources excludes. [line 622] [Source]

CLI models [Source]

  • About 55 lines of new CLI reference content (mirrored in cli/beta/models.md).