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

Executive Summary

  • Self-hosted Claude Managed Agents launched: you can now run the whole Managed Agents service (API, agent loop, storage, sandboxes) on your own infrastructure, with a new 10-page documentation section covering quickstart, Kubernetes, configuration, upgrades and more.
  • The preserved thinking guide was split into focused pages (prefix checks, changing tools or instructions, fixing "bound to a different conversation" errors), and Fable 5.2 / Mythos 5.2 now return a 400 error when the prefix check fails.
  • New disableRemotePreviewScripts managed setting for the desktop app lets admins stop scripts in HTML previews for SSH and WSL sessions without disabling external browsing.
  • The /claude-api skill gained a latency-optimize workflow, and the mods test kit documents a much larger set of expect assertions.

New Claude Code versions

No new Claude Code versions were added today.


Claude Code changes

Changed documents

chrome [Source]

  • Plan mode permission prompt description no longer lists running a shortcut. [line 127] [Source]

commands [Source]

  • /claude-api gained a latency-optimize subcommand. [line 64] [Source]

desktop [Source]

  • New disableRemotePreviewScripts managed setting turns off only scripts in HTML previews for SSH and WSL sessions, without blocking external browsing (requires Desktop v5.1231.0+). Earlier versions ignore it; a note explains the disableBrowserExternalNavigation fallback. [lines 147-153] [Source]
  • New section on checking a computer for scripts that still run, using the notice text above the preview. [lines 167-170] [Source]
  • Managed settings table and MDM precedence text now list disableRemotePreviewScripts with per-key minimum Desktop versions. [lines 2044-2057] [Source]
  • Guidance on stopping users from signing in with their own API key or Console account: also deploy forceLoginOrgUUID in the Desktop policy, with cautions for Console organizations and the Enterprise Admin Console. [lines 2122-2129] [Source]

desktop-changelog [Source]

  • Mirrors the same disableRemotePreviewScripts, script-check, and forceLoginOrgUUID changes as the desktop page. [line 147] [Source]

ide-integrations [Source]

  • New note: if a reload interrupts Claude in plan mode, plan mode turns off and Claude doesn't continue on its own; select Plan to keep planning. [line 398] [Source]
  • New claude-vscode.toggleDictation shortcut (Cmd/Ctrl+D) for voice dictation. [line 652] [Source]

model-config [Source]

  • "Fix a mistyped model name" section moved earlier in the page, and the intro now points to the "did you mean" warning. [lines 167-206] [Source]

plugins/mods/overview [Source]

  • The built-in mods table is now partial; for others, open the plugin's details in /plugin to read what it does or disable it. [lines 217-232] [Source]

plugins/mods/test [Source]

  • New "Check a result with expect" section listing many more assertions (toStrictEqual, toStartWith, numeric comparisons, toHaveLength, toHaveProperty, toBeNull, toBeInstanceOf, etc.) plus .resolves / .rejects. [lines 193-204] [Source]

self-hosted-environments-reference [Source]

  • New CLAUDE_RUNNER_DISABLE_RENEWAL_WARNING runner variable suppresses the session access renewal refusal message (Claude Code v2.1.319+). [line 124] [Source]

settings-reference [Source]

  • New managed-only disableRemotePreviewScripts setting, with type, default, parentSettingsBehavior guidance and an example. [lines 5005-5028] [Source]

skills [Source]

  • The /claude-api skill gained a latency-optimize workflow that finds where time goes in existing Claude API code and proposes changes (v2.1.319+). [line 90] [Source]

slash-commands [Source]

  • Same latency-optimize addition as in skills. [line 90] [Source]

vs-code [Source]

  • Same plan-mode reload note and toggleDictation shortcut as in ide-integrations. [lines 398-652] [Source]

API changes

New Documents

build-with-claude/preserved-thinking/change-tools-or-instructions [Source]

Split out of the preserved thinking guide. Shows how to change tools or system instructions without failing the prefix check, and how to append assistant turns exactly as returned.

build-with-claude/preserved-thinking/fix-bound-to-a-different-conversation [Source]

Troubleshooting page for the "The block is bound to a different conversation" 400 error, including the drop_block stopgap and its output-token cost.

build-with-claude/preserved-thinking/prefix-checks [Source]

Explains what the prefix check covers (system prompt, tools, earlier messages), what counts as an edit, and what to check in your integration.

managed-agents/self-hosted [Source]

Landing page for self-hosted Claude Managed Agents. Compares cloud sandboxes, self-hosted sandboxes, and a fully self-hosted service, and explains when to choose each. Access is by contacting your Anthropic account team.

managed-agents/self-hosted-data-and-access [Source]

Describes where data lives and who can access it with self-hosted Managed Agents.

Self-hosted Managed Agents section (managed-agents/self-hosted/)

Changed documents

about-claude/pricing [Source]

  • Self-hosted Managed Agents pricing is available through your Anthropic account team. [line 441] [Source]

api/errors [Source]

  • Thinking block errors are reworded: a 400 occurs when thinking blocks don't match what the API returned, and the prefix check failure now links to the new prefix-checks page. [lines 439-447] [Source]
  • Two recovery options for a conversation whose history already holds an edit: use drop_block on every later request, or strip the named block and all later thinking blocks. [lines 543-549] [Source]
  • Removed the paragraph about 429 acceleration limits. [line 38] [Source]

build-with-claude/effort [Source]

  • Expanded per-model effort guidance for Fable 5.2, Opus 5.5 and Sonnet 5.5: where to start, run your own effort sweep, and set a large max_tokens. [lines 267-310] [Source]
  • Changing top-level effort between requests can invalidate all or part of the prompt cache (previously stated as fully invalidating). [line 432] [Source]

build-with-claude/mid-conversation-system-messages [Source]

  • Microsoft Foundry is now listed as stable, and the beta header guidance for mid-conversation tool changes was restructured around inline-tools-2026-09-15. [lines 33-49] [Source]

build-with-claude/overview [Source]

  • Added self-hosted Claude Managed Agents to the capabilities table. [line 121] [Source]

build-with-claude/preserved-thinking [Source]

  • Substantially shortened; the prefix-check, tool/instruction-change and error-fix content moved into three new sub-pages.

build-with-claude/structured-outputs [Source]

  • Large revision with significant additions (about 400 lines changed).

managed-agents/overview [Source]

  • Updated to reference self-hosted Managed Agents.

managed-agents/permission-policies [Source]

  • Substantial rework of the permission policies documentation.

managed-agents/self-hosted-sandboxes [Source]

  • Now cross-references the full self-hosted Managed Agents option.

managed-agents/tools-web-restrictions [Source]

  • Updated web restriction documentation.

models/fable-5-2/migration-guide [Source]

  • Prefix check section revised to match the new preserved thinking pages.

models/fable-5-2/whats-new-fable-5-2 [Source]

  • Breaking change reframed as "Failed prefix checks return a 400 error", with links to the new sub-pages and the drop_block stopgap moved under it. [lines 13-57] [Source]
  • Microsoft Foundry added to platforms that use the inline-tools-2026-09-15 header. [line 91] [Source]

models/opus-5-5/migration-guide [Source]

  • Migration guidance updated for prefix checks and thinking blocks.

release-notes/overview [Source]

  • November 12, 2026 entry announcing the launch of self-hosted Claude Managed Agents. [lines 15-18] [Source]