Claude docs changes for November 13th, 2026 [diff]

Executive Summary

  • Claude Code 2.1.323 adds Code Mode, an experimental opt-in where Claude writes JavaScript that makes several tool calls in one step. It has a new dedicated docs page and touches the hooks, auto mode, permissions and settings docs.
  • Plugins gain an engines["claude-code"] version range in plugin.json, plus TypeScript type checking of hooks modules in claude plugin validate.
  • A large security hardening pass: stricter Bash permission checks, sandbox changes (worktree git commands, symlinked writes), MCP OAuth redirect blocking, and tighter hook and permission-check failure behavior.
  • On Amazon Bedrock and Vertex AI, a session whose model is rejected as unavailable now switches to another model automatically, and CLAUDE_CODE_DISABLE_MODEL_ACCESS_FALLBACK=1 turns that off.
  • API docs add message threads (server-side conversation state), deprecate many old beta header values, document hosting options for Claude in Microsoft Foundry, and expand the Managed Agents workflow-run docs.

New Claude Code versions

2.1.323

New features

  • Added Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools. It can be turned off for an organization.
  • Added engines["claude-code"] to plugin.json. Plugins outside their stated Claude Code version range are not installed, updated or loaded, and install and update pick the newest version that fits.
  • claude plugin validate now type-checks a plugin's TypeScript hooks module and reports the $ nouns a mod's engine.create hook adds, withholds or replaces.
  • Support for CLAUDE_CODE_AUTO_MODE_MODEL in managed settings for Claude Desktop sessions on third-party providers.
  • Options menu (stop, switch to usage credits, upgrade) when a model needs usage credits that are off or used up.
  • Progress heartbeat every 30 seconds while hooks run long in claude -p and SDK sessions.
  • .docx support for Artifact tool asset uploads.
  • New env vars: CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES and CLAUDE_CODE_RESUME_HINT_COMMAND.
  • Start-up warnings for Bash allow rules with a * in a variable's value, and for plugin language servers that can't start.
  • unsetOptionCount in the claude plugin install --json result, and isPromptCovered in the props a mod's pane receives.
  • VS Code: trust prompt before the first session in an untrusted folder, and live output of running Bash/PowerShell commands in the chat.

Existing feature improvements

  • Clearer sign-in errors in /login and claude setup-token.
  • More consistent Bash permission checks; some commands that could run differently from how they were checked now ask for permission.
  • /mcp shows how many of a server's tools were left out because of invalid input schemas.
  • Better memory use with many subagents, faster WebFetch on deeply nested pages, and the install script no longer downloads the latest version twice.
  • The Claude apps gateway /v1/models states max_input_tokens where the context window is certain.
  • The context window assumed for an unrecognized model ID changed from 200K to 1M. Use CLAUDE_CODE_DISABLE_1M_CONTEXT=1 to keep 200K.
  • Messages sessions can send each other without a prompt rose from about 20 to 100,000.
  • /security-review now asks before Claude runs it.
  • Sandbox: git worktree add/move/remove and commits in a worktree the session wasn't started in now fail read-only. Symlinked writes that lead elsewhere are ignored. An invalid --settings file with a sandbox key stops start-up.
  • MCP: project (.mcp.json) server approvals match the exact server name, and finishing an MCP sign-in always asks you to confirm.
  • "Don't ask again" rules that look like they hold a secret, or that would be saved to a git-tracked settings.local.json, last for the session only.
  • Permission check failures after a PreToolUse "allow" hook, or on input rewritten by a PermissionRequest hook, are now decided as if no hook had answered, or asked again.
  • Worktree creation refuses checkouts whose .git folders are links or hold hard links.
  • Marketplace and plugin installs fail when a git submodule isn't on the recorded commit.
  • /logout in a background session now signs out every session on the computer.
  • Removed the subagentStatusLine setting. A plugin's agent setting is deprecated.

Major bug fixes

  • Fixed /compact summarizing away a message Claude hadn't answered yet.
  • Fixed a resumed session joining your next message to an old interrupted message.
  • Fixed memory spikes after Claude changed a very large file, and Edit/Bash running out of memory on certain /proc files, devices and named pipes.
  • Fixed a settings save in a running session undoing a hand edit of settings.json.
  • Fixed claude plugin commands and claude mcp list signing you out of a Console profile.
  • Fixed Windows file-reading Bash commands running without a permission prompt when given a wildcard or list of files.
  • Fixed non-streaming requests failing on errors that streamed requests recover from.

Claude Code changes

New Documents

Code Mode [Source]

New page for Code Mode, an experimental, off-by-default feature in which Claude writes JavaScript that makes several tool calls (parallel commands, chained results, filtered output) in one step. It explains how to view the code (Ctrl+O transcript viewer), when to use it, and its limitations: resuming only in the terminal, auto mode classifying the code rather than each call, and a ten-minute limit unless code runs in the background. It also covers requirements, permission checks, hook behavior, and how to turn it on or off for an organization.

Changed documents

amazon-bedrock [Source]

  • When Bedrock rejects the session's model as unavailable, Claude Code switches to another model (or a configured fallback chain). The page lists when it doesn't switch and how to disable it with CLAUDE_CODE_DISABLE_MODEL_ACCESS_FALLBACK=1. [lines 348-358] [Source]
  • Mantle: set both CLAUDE_CODE_USE_BEDROCK and CLAUDE_CODE_USE_MANTLE, and a new 400 section covers model IDs Mantle doesn't serve. [lines 516-522] [lines 622-627] [Source]

auto-mode-config [Source]

  • Notes how routing all shell commands through the classifier works in a Code Mode session.

claude-apps-gateway-config [Source]

  • Per-developer attribution on Bedrock: set assume_role with session_name on every bedrock and mantle upstream. [line 518] [Source]
  • Foundry RBAC now names the Foundry User or Cognitive Services User role. [line 657] [Source]
  • To use the fable alias with auto_include_builtin_models: false, add a claude-fable-5 entry to models. [line 1002] [Source]
  • Policy merging: deny-lists and hook arrays (permissions.deny/ask, deniedMcpServers, deniedModels, hooks, and others) take the union of base and policy. [lines 1028-1032] [Source]
  • Claude Code doesn't read an alias's upstream_model to size its context window. [line 1284] [Source]

env-vars [Source]

  • New CLAUDE_CODE_EXPERIMENTAL_CODE_MODE variable (1 to start sessions in Code Mode, 0 to keep it off). [line 297] [Source]

errors [Source]

  • New "Stopped an OAuth request" section: Claude Code stops MCP OAuth requests that an endpoint redirects elsewhere or more than 5 times. [lines 1824-1840] [Source]
  • New Remote Control errors about the --settings file being skipped or changing after start. [lines 251-252] [Source]
  • New auto-mode denial reason for a call changed after review. [line 827] [Source]
  • Cloud sessions: a 403 or DNS error can mean the host isn't on the environment's allowlist. Image-size errors can name a different pixel limit. [line 2222] [Source]
  • PreModelSwitch refusal when managed-plugin hooks can't be checked now applies only before v2.1.323. [line 2890] [Source]

feature-availability [Source]

  • Added Code Mode to the availability table (experimental, local CLI only). [lines 116-125] [Source]
  • Updated the per-provider "Not available" lists. [line 240] [Source]

google-vertex-ai [Source]

  • Same automatic model switch when the model is unavailable in your project or region, with /model s to keep working and CLAUDE_CODE_DISABLE_MODEL_ACCESS_FALLBACK=1 to opt out. [lines 227-237] [Source]

hooks [Source]

  • subagentStatusLine removed from the list of settings narrowed to managed settings. [line 256] [Source]
  • Clarified the HTTP hook onFailure field. [line 504] [Source]
  • Notes that PreToolUse timeouts and failures work differently in Code Mode. [line 900] [line 984] [Source]
  • Code Mode output can be withheld by hooks, and the block reason is added to context in that case. [line 1519] [line 1568] [Source]
  • ConfigChange hooks also run for managed policy and skill file changes. [line 3008] [Source]

interactive-mode [Source]

  • Alt+M cycles permission modes on Windows when VT input mode isn't enabled; clarified spellcheck behavior. [line 35] [line 595] [Source]

keybindings [Source]

  • New default ctrl+x chords in Chat (ctrl+x ctrl+=, ctrl+x ctrl+-, ctrl+x ctrl+_, ctrl+x =, ctrl+x -) and the matching unbind example. [line 580] [lines 630-634] [Source]

microsoft-foundry [Source]

  • Simplified permissions guidance to the Foundry User or Cognitive Services User role, replacing the custom role definition. [line 146] [Source]

model-config [Source]

  • If a managed plugin or its hooks fail to load, model switches are still allowed but its PreModelSwitch hooks don't run. [line 158] [Source]
  • Unrecognized model IDs now warn that Claude Code assumes a different context window. [line 904] [Source]

permission-modes [Source]

  • How to leave plan mode with Shift+Tab, and a note that with Code Mode the classifier reviews the code once, not each inner tool call. [line 303] [line 592] [Source]

plugins-reference [Source]

  • Restructured plugin.json reference with a clearer intro and fields table (mirrors the manifest reference page). [line 116] [Source]

plugins/components [Source]

  • Rewrote the intro and example-plugin description, and removed some embedded example content. [line 21] [Source]

plugins/manifest-reference [Source]

  • Clarified the top-level keys table (only name is required) and the default-location layout example. [line 116] [line 621] [Source]

sandboxing [Source]

  • On Linux and WSL2, a filesystem entry for a path that doesn't exist yet has no effect, so create it first. [line 275] [Source]
  • Tighter rules on who can set filesystem.disabled: not project settings; managed-only when managed settings configure sandbox.filesystem; ignored when CLAUDE_CODE_SUBPROCESS_ENV_SCRUB is set. [lines 377-379] [Source]

settings-reference [Source]

  • Auto mode shell-classifier key behaves differently in Code Mode. [line 883] [Source]
  • Sandbox path entries that don't exist yet are ignored on Linux/WSL2. [line 1531] [Source]
  • The env setting overwrites the same shell variable, and the highest-precedence settings file wins. [line 2453] [Source]
  • Code Mode can be turned on via CLAUDE_CODE_EXPERIMENTAL_CODE_MODE from the shell, user settings or --settings. [line 2494] [Source]
  • Removed the subagentStatusLine example. [line 3173] [Source]

setup [Source]

  • Alpine/musl: install requires bash and curl, and Claude Code requires bash or zsh. [line 140] [Source]

statusline [Source]

  • Removed the subagent status line section, matching the removal of the subagentStatusLine setting. [line 1083] [Source]

API changes

Changed documents

agents-and-tools/tool-use/programmatic-tool-calling [Source]

  • The pending call's ~4 minute timeout starts when Claude's code makes the call, so don't use expires_at as your deadline. Respond as soon as you can. [line 456] [line 1468] [Source]

agents-and-tools/tool-use/tool-runner [Source]

  • No longer described as beta. [line 48] [Source]
  • New caveat: when you take over an iteration, you must keep the conversation valid by appending the assistant message and tool results. [line 836] [Source]
  • New "Compacting long conversations" section covering threshold compaction and on-demand compaction (compact_before_next_turn()). [line 836] [Source]

api/beta [Source]

  • Many AnthropicBeta values (such as message-batches-2024-09-24, prompt-caching-2024-07-31, pdfs-2024-09-25, token-counting-2024-11-01, extended-cache-ttl-2025-04-11, cache-diagnosis-2026-04-07) are marked deprecated and no longer needed. files-api-2025-04-14 and skills-2025-10-02 are deprecated with migration steps, since sending them selects the old beta response shapes. The same change repeats across the per-language SDK beta pages (Python, TypeScript, Go, Java, Ruby, C#, CLI).

build-with-claude/claude-in-microsoft-foundry [Source]

  • New hosting options comparison: hosted on Azure (US Data Zone Standard available, prompts stay within Azure) vs hosted on Anthropic (all Claude models, Global Standard).
  • New RBAC prerequisite (Foundry User or Cognitive Services User), deployment setup steps (model version selects hosting, region scope), and the ANTHROPIC_FOUNDRY_BASE_URL variable.
  • Notes on private networking, 1M context availability, and updated troubleshooting.

build-with-claude/message-threads [Source]

  • Expanded guide to server-side message threads: create, continue (send only new messages plus previous_message_id) and replay, with expiry after 5 minutes idle and deletion within 1 hour.
  • New SDK thread helper section and examples with top-level cache_control for caching each turn.
  • Recovery, forking, params, fallbacks and sticky routing behavior, the helper's refusal handling, and new error codes thread_fingerprint_mismatch and thread_unsupported_request. Threads are append-only, so compaction means starting a new thread from a summary.

manage-claude/admin-api-keys [Source]

  • Reorganized around a comparison table of the key for Claude Console vs Claude Enterprise (sign-in address, where to create, prefix, who can create, scopes, which APIs it works with), with a separate Console creation section.

manage-claude/analytics-api [Source]

  • Key types for the two analytics APIs presented as a table; points to the in-product Analytics dashboard.

manage-claude/compliance-api-access [Source]

  • Key types table, enablement for standalone vs parent-linked Console organizations, a warning that turning the Compliance API off stops activity recording, and who can create keys.

manage-claude/wif-reference [Source]

  • Clarified token exchange fields and environment variables. The env-var path activates only when the rule ID, organization ID and a token source are set, and the minted token is scoped to the named workspace.

managed-agents/migration [Source]

  • Added "What you control" list and guidance for agents built with client.messages.create(); mapped the SDK permission_mode/can_use_tool to per-tool permission policies.

managed-agents/reference [Source]

  • Documented the redacted content block, the new session.deleted event, and clarified workflow_run.status_running.

managed-agents/scheduled-deployments [Source]

  • On Claude Platform on AWS, set execution_identity to an aws_role so each run's session uses an IAM role.
  • Archiving or deleting an agent archives its deployments. Added a recovery procedure for a deployment paused by an archived subagent.

managed-agents/self-hosted [Source]

  • Added a License section for the server software.

managed-agents/workflow-runs [Source]

  • Clarified run lifecycle: runs start running or idle, pause at the budget and resume when it's raised, and are "open" from workflow_run.created until workflow_run.status_ended.
  • Run events arrive on the primary thread's stream and don't trigger webhooks. New timeout_error for the lifetime limit.
  • Agent threads in a run: agent field, thread events on the primary stream, no phase field, and how to ask the agent to print the workflow it wrote.

Other API pages

  • Smaller updates to Managed Agents (events and streaming, multiagent orchestration, permission policies, secure deployment pages, self-hosted sandboxes, session threads), text-watermark, ip-allowlist, and generated SDK reference pages (sessions, memory stores, environments, vaults, organization) across all languages.