Claude docs changes for October 12th, 2026 [diff]

Executive Summary

  • Claude Desktop admins get a documented recipe to turn off local sessions and distribute SSH connections via disableDesktopLocalSessions, sshConfigs and sshHostAllowlist.
  • WorktreeCreate/WorktreeRemove hook docs now list every case that triggers them (EnterWorktree, ExitWorktree, workflows, claude remote-control) and the cleanup caveats.
  • Scheduled task timing is rewritten: */5 * * * * (what /loop 5m creates) is no longer delayed, and other delays are explained per schedule shape.
  • Agent SDK adds a run_id on task messages to tell apart multiple runs of the same background task.
  • The sandbox docs explain that commands like FOO=bar npm test, make CC=clang and eval "ls" still go through the permission flow, and which allow rules stop the prompts.

New Claude Code versions

No new Claude Code versions were added today.


Claude Code changes

Changed documents

agent-sdk/typescript [Source]

  • New runId on the task-notification message origin, matching the run it belongs to, so a late notification can be attributed to the right run. [line 2211] [Source]
  • New optional run_id field on SDKTaskStartedMessage and SDKTaskNotificationMessage. It identifies one run of a task that runs again under the same task_id (such as a resumed subagent). Requires Agent SDK v0.3.292 or later. [line 5212] [Source]

debug-your-config [Source]

  • The "clean session" procedure is now step by step. It uses an empty working directory, a fresh CLAUDE_CONFIG_DIR and --setting-sources user, with separate Bash/Zsh and PowerShell instructions. [lines 85-131] [Source]
  • Added guidance on what to do whether the problem disappears or persists in the clean session (copy config files back one at a time, or check /status for managed settings and env vars). [line 131] [Source]

desktop [Source]

  • New section "Turn off local sessions and distribute SSH connections". It shows a managed-settings example combining disableDesktopLocalSessions (requires Desktop v1.37937.0+), sshConfigs and sshHostAllowlist. It also covers what the keys don't affect (CLI, Cowork, cloud sessions) and how to verify. [line 810] [Source]
  • If user@ is omitted from sshHost, Desktop uses the User from the developer's ~/.ssh/config, or their local username. [line 788] [Source]

desktop-changelog [Source]

  • Updated with the same Desktop SSH and local-session administration content (see desktop).

errors [Source]

  • New Login expired ยท Run /login to sign in again, or re-authenticate your Anthropic profile message in the index. The Anthropic profile login-expired section is reorganized around how the profile was found (ANTHROPIC_PROFILE versus discovered with user_oauth). [lines 1394-1410] [Source]
  • New steps for renaming a marketplace declared inline with "source": "settings" before removing it, so its settings entry isn't lost (Claude Code v2.1.292+). [lines 3827-3832] [Source]

hooks [Source]

  • WorktreeCreate now lists all triggers: --worktree, EnterWorktree, subagent isolation, workflow agents, background sessions and claude remote-control in worktree mode. A git worktree add run through Bash does not trigger it; use a PreToolUse hook instead. [line 3022] [Source]
  • WorktreeRemove also runs for ExitWorktree and claude remote-control cleanup. Worktrees created for subagents or workflow agents are kept, and ExitWorktree needs discard_changes: true to remove a hook-created worktree. [lines 3096-3102] [Source]
  • Without a WorktreeRemove hook, behavior differs by route. Exit and ExitWorktree fall back to git, while remote-control keeps the worktree and prints worktree removal failed, kept: <path>. [lines 3108-3112] [Source]

plugins/mods/reference [Source]

  • New prompt.autocomplete mod event. It fires when a word ends at the cursor, and the hook can return { suggestions } rows shown below Claude Code's own. [line 73] [Source]

permissions [Source]

  • Notes that some sandboxed commands still go through the regular permission flow, linking to auto-allow mode. [line 661] [Source]

sandboxing [Source]

  • Explains that commands like FOO=bar npm test, make CC=clang and eval "ls" still go through the permission flow even when sandboxed. It gives allow-rule patterns to stop recurring prompts, and notes that a matching allow rule also approves the unsandboxed retry. [line 180] [Source]

scheduled-tasks [Source]

  • Rewrote the run-time jitter explanation. Recurring tasks run late by up to half the gap to the next run (max 30 minutes), and unevenly spaced schedules can vary per run. [line 168] [Source]
  • */5 * * * * (what /loop 5m creates) is not delayed. Each run starts 4m45s after the previous one so it lands inside the 5-minute prompt cache TTL. [lines 168-178] [Source]

workflows [Source]

  • New "Which agents run again" section on resuming a stopped or paused run. Completed agents return cached results until one differs or fails, and then it and every later agent reruns. Stopping the whole run doesn't count as a failure, but stopping a single agent does. [line 370] [Source]

API changes

Changed documents

api/go/* (Go SDK reference) [Source]

  • Large regenerated diffs across the Go SDK reference pages (beta, beta/agents, beta/deployments, beta/environments, beta/messages, beta/sessions, beta/vaults, messages, and subpages). These are mostly mechanical regeneration changes rather than new concepts.

manage-claude/api-and-data-retention [Source]

  • The feature eligibility table's Endpoint column is reformatted, for example /v1/messages with \bash` tool` replaces the parenthesized form. The ZDR/HIPAA eligibility values are unchanged. [lines 160-197] [Source]

managed-agents/budgets [Source]

  • Field names are now annotated per SDK language (Java/PHP camelCase, C#/Go PascalCase). Removing a budget is documented per SDK, such as budget=None in Python and param.NullStruct in Go, and the PHP example sends an explicit null as an extra body parameter. [line 146] [Source]

managed-agents/scheduled-deployments [Source]

  • New examples for updating a deployment's budget (CLI ant apply deployment.md with front matter, plus SDK examples), and per-SDK field-name annotations for schedule, expression, timezone and budget. [lines 263-395] [Source]

Other Managed Agents pages

  • Smaller updates, mostly per-SDK field-name annotations and example adjustments, across agent-setup, define-outcomes, dreams, environments, events-and-streaming, files, github, mcp-connector, memory, migration, multiagent-orchestration, permission-policies, reference, sessions, session-threads, skills, tools-web-restrictions, vaults, webhooks and workflow-runs.