Claude docs changes for October 22nd, 2026 [diff]

Executive Summary

  • New Claude API version 2026-10-21 with a full migration guide. It removes temperature, top_p and top_k on every model, rejects document blocks in assistant messages and citations in user messages, and renames several Admin and Analytics API fields. It ships with new major SDK versions (Python 2.0.0, TypeScript 1.0.0, Go v2, Java 3.0.0, and others) and ant CLI 2.0.0.
  • Claude Code 2.1.307 adds /claude-api api-version to move API code to the new version. It also adds the claude_code.safety_event OpenTelemetry event and retries unparseable tool calls up to twice.
  • New sandbox setting sandbox.allowNestedUserNamespaces lets Chromium, Electron and rootless Podman run inside the Linux/WSL2 sandbox. The docs warn that it turns off a security boundary.
  • The docs say more about managed-settings controls. strictPluginOnlyCustomization and strictKnownMarketplaces now block skills-directory plugins, and rm/rmdir critical-path protection is described as covering only commands Claude Code recognizes.
  • Scheduled tasks gain custom cron intervals, loop-stop behavior and resume semantics. A new keybinding (select:cancel) cancels a question with one key press.

New Claude Code versions

2.1.307

New features

  • Added /claude-api api-version to move Claude API code to API version 2026-10-21
  • Added the claude_code.safety_event OpenTelemetry event for each tool call that a permission rule, auto mode, a safety check, a hook or the sandbox stopped, prompted or sent back
  • [Code Review] Added @claude review stop (also off and pause), which ends an @claude review always subscription on a pull request
  • [VSCode] Added a notice on the microphone button when voice dictation stops by itself, saying why and that the text so far is kept

Existing feature improvements

  • Claude Code now retries unparseable tool calls up to twice before ending the turn with an error
  • Fork subagents launched together are now told which launch is their own, so they are much less likely to start each other's work
  • Browser sign-in shows the success page as soon as sign-in completes
  • Web search errors in cloud and self-hosted runner sessions now show the reason when a 403 gives one
  • Faster first reply in cloud sessions when one of the desktop app's MCP servers can't connect
  • Faster Remote Control connection for sessions with large pasted images or PDF reads
  • The axScreenReader setting and CLAUDE_AX_SCREEN_READER in a settings env block now apply only from user, managed or --settings settings
  • Self-hosted runner: checkout hooks are stopped after 30 minutes by default (--checkout-hook-timeout-min, 0 to disable)
  • Self-hosted runner: the certificate file built for Anthropic-managed git filters its trusted certificates as the CLI does
  • [VSCode] Dictation and the extension's own requests filter trusted certificates as the CLI does, and trust the NODE_EXTRA_CA_CERTS file from the environment or user settings

Major bug fixes

  • Fixed model-not-found errors when CLAUDE_CODE_DISABLE_1M_CONTEXT is set and the model is an alias with the [1m] suffix
  • Fixed MCP problems: "needs authentication" in claude -p on macOS with a locked keychain, false timeouts after sleep, and HTTP MCP tool calls sent twice when the connection closed mid-call (only read-only or idempotent tools are now retried)
  • Fixed some "Autocompact is thrashing" errors in turns that read many PDF pages, images or other large tool results
  • Fixed /autocompact and /clear state being lost when cloud or self-hosted runner sessions restarted
  • Fixed /permissions > Recently denied not showing the denied shell command
  • Fixed model switches through the Agent SDK's setModel() being refused when the API took more than 5 seconds to confirm; the wait is now up to 15 seconds
  • Self-hosted runner: fixed checkout hooks leaving processes running and the runner's slot held when a session was archived, requeued or stopped

2.1.308

Major bug fixes

  • Bug fixes and reliability improvements

Claude Code changes

Changed documents

errors [Source]

  • New error entries for API Error: Stream idle timeout - no chunks received and The model's tool call could not be parsed (retry also failed). [lines 24-33] [Source]
  • Automatic retries now cover a response that announces a tool call but contains none; Claude Code retries up to two times. [line 445] [Source]
  • A "Prompt is too long" error behind a custom ANTHROPIC_BASE_URL is fixed by setting the model's context window or updating to v2.1.299+. On v2.1.305 with many MCP tools, the error can hide an unrelated 400/422, which v2.1.306+ fixes. [lines 2216-2218] [Source]
  • New section "Managed settings need your approval" for claude plugin commands that can't show the approval dialog. [lines 5469-5480] [Source]

hooks [Source]

  • New "Limits on skill hooks" section. Skill frontmatter hooks may not fire because of a plugin-only lock (strictPluginOnlyCustomization) or the v2.1.304+ per-source limits (1,024 hooks, 16,384 matcher characters). It explains how to diagnose with claude --debug and fix. [lines 680-695] [Source]

interactive-mode [Source]

  • New "Cancel a question in vim mode" section: press Esc twice in the Other row, or bind a single key. [lines 261-266] [Source]
  • New "Background a session with messages queued" section: queued messages move with the session, but commands, ! shell commands and messages with an IDE selection can't. [lines 398-406] [Source]

keybindings [Source]

  • New "Cancel a question with one key press" section showing how to bind a Ctrl/Alt key to select:cancel in the Select context. [lines 688-708] [Source]

mcp [Source]

  • Reconnect behavior in claude -p and Agent SDK sessions is split: transient errors keep retrying for about half an hour, while a server that needs sign-in again stops retrying and reports needs-auth. [lines 372-374] [Source]
  • New section on signing in to an MCP server behind a gateway that requires a header (v2.1.299+): configured headers or headersHelper are sent on sign-in, refresh and revocation requests, except Authorization. [lines 1046-1057] [Source]
  • New "Read a background command's output as an MCP client" section covering backgroundOutputPath from claude mcp serve, including how to detect completion and read the tail of the file (v2.1.295+). [lines 1281-1294] [Source]

memory [Source]

  • Path-scoped rule patterns are relative to the rule's base directory (the project's .claude parent, or the primary working directory for user-level rules). For rules in a sibling worktree, start in that worktree or use /cd. [lines 207-213] [Source]
  • A managed CLAUDE.md can't be excluded with claudeMdExcludes. [lines 285-345] [Source]

permission-modes [Source]

  • The critical-path rm/rmdir protection is reworded to cover only removals Claude Code recognizes. A warning says a missed command gets no special handling and can be approved by allow rules, bypassPermissions or sandbox auto-allow. It lists mitigations: deny rules, sandboxing and strict sandbox mode. [lines 772-786] [Source]
  • New troubleshooting entry: claude agents shows the bypass permissions warning when skipDangerousModePermissionPrompt is set in local settings or a --settings file. [lines 712-715] [Source]

plugins/loading [Source]

  • New section "Skills-directory plugins blocked by managed settings": strictPluginOnlyCustomization (any locked kind) or a strictKnownMarketplaces/blockedMarketplaces list without or with skills-dir silently skips plugin folders in ~/.claude/skills/ and .claude/skills/. [lines 82-105] [Source]
  • Repository-plugin checks are reorganized into a "Check why a repository plugin didn't load" section. [lines 70-82] [Source]
  • Clarified sync timing in cloud Cowork sessions and for plugins toggled on claude.ai during a terminal session. [lines 118-126] [Source]

plugins/mods/troubleshoot [Source]

  • After a load error ending in "the version loaded before is unloaded", run /reload-plugins once fixed. [line 87] [Source]
  • New entry for "the parser stopped" / "the engine ran out of stack loading the file" errors caused by deeply nested hooks-module code. [lines 101-109] [Source]

prompt-caching [Source]

  • Compaction caching is detailed layer by layer: the system prompt layer is read from cache (rebuilt once for old resumed conversations), and the project context layer is read from cache only if CLAUDE.md and memory are unchanged. [lines 170-177] [Source]

routines [Source]

  • Enterprise organizations with custom roles need a role that includes the scheduled tasks permission. [lines 32-33] [Source]
  • Scheduled tasks run autonomously without a permission mode. [line 49] [Source]
  • Custom cron intervals can be set from the edit dialog. One-off tasks turn themselves off and show Completed after running. [lines 149-160] [Source]
  • A scheduled task with one repository can declare MCP servers in a committed .mcp.json. [line 355] [Source]

sandboxing [Source]

  • New troubleshooting section "Chromium or rootless Podman fails inside the sandbox on Linux". It covers --no-sandbox, excludedCommands, and the new sandbox.allowNestedUserNamespaces setting, with a warning that it removes a security boundary. [lines 840-861] [Source]
  • When managed settings configure sandbox.filesystem or deny-mode credentials.files, only managed settings can set filesystem.disabled. [lines 362-362] [Source]
  • Protected paths stay protected despite allowWrite or Edit allow rules. Only filesystem.disabled turns them off. [lines 541-548] [Source]

scheduled-tasks [Source]

  • Comparison table gains customizable schedule and minimum interval rows. [lines 29-30] [Source]
  • How to stop loops: Esc or Ctrl+C for self-paced loops, but ask Claude to cancel fixed-interval loops. Pressing Ctrl+C twice detaches from a background session without stopping a self-paced loop. [lines 123-131] [Source]
  • Resume semantics: --resume/--continue restore CronCreate tasks without replaying missed runs, and background Bash commands and monitors aren't restarted. [lines 234-238] [Source]

self-hosted-environments-deploy [Source]

  • New troubleshooting entry for no longer exists on the remote / couldn't find remote ref, with causes (deleted branch, checkout hook, git host rewrite, runner older than v2.1.301). [lines 670-680] [Source]
  • The runner's clone and fetches check Anthropic-managed git against a per-session certificate file. [line 300] [Source]

server-managed-settings [Source]

  • Sandbox network and isolation settings are added to the settings that need security approval. [line 250] [Source]
  • claude plugin commands that can't read the keyboard don't apply delivered settings without cached settings. [lines 281-283] [Source]
  • Delivered env variables for context window and compaction apply without approval, except CLAUDE_CODE_DISABLE_1M_CONTEXT. Proxy/base-URL variables, OTEL_EXPORTER_OTLP_ENDPOINT and CLAUDE_CODE_DISABLE_1M_CONTEXT always need approval in interactive sessions. [lines 293-304] [Source]

settings-reference [Source]

  • New sandbox.allowNestedUserNamespaces setting (Boolean, default false, user or managed scope only). [lines 1594-1616] [Source]
  • New { "source": "skills-dir" } marketplace source. Any strictKnownMarketplaces allowlist, even an empty one, stops @skills-dir plugins loading unless it is included. [lines 3904-3918] [Source]
  • Locking any strictPluginOnlyCustomization surface also stops skills-directory plugins from loading. [lines 4024-4026] [Source]

fullscreen [Source]

  • A --chrome or --no-chrome launch flag is carried into the relaunched fullscreen session. [line 31] [Source]

API changes

New Documents

api/migration-2026-10-21 [Source]

Guide for moving from API version 2023-06-01 to 2026-10-21. It lists the minimum SDK versions (Python 2.0.0, TypeScript 1.0.0, C# 13.0.0, Go v2 with a new /v2 module path, Java 3.0.0, PHP 1.0.0, Ruby 2.0.0, ant CLI 2.0.0) and how to send the anthropic-version header with cURL. It notes that on Amazon Bedrock the new version is only available for Claude 4.7 and later models. Breaking changes covered:

  • Messages API: document blocks in assistant messages and non-empty citations on user text blocks now return 400. temperature, top_p and top_k are removed on every model.
  • Admin API and Claude Enterprise Analytics API: renamed response fields (see the versioning entry below).
  • SDK-specific code changes.

Changed documents

about-claude/model-deprecations [Source]

  • The deprecated-parameters table now lists temperature, top_p and top_k as removed in API version 2026-10-21 (400 on every model). Earlier versions reject them only on Claude 4.7+ and Claude Mythos Preview. [lines 209-213] [Source]

api/versioning [Source]

  • Added 2026-10-21 to the version history, with platform availability and its breaking changes: the Messages API changes above and renamed Admin API fields (RBAC groups roles → role_ids, members group_id → rbac_group_id, roles name → display_name, rate-limit group_type → group.type, API keys workspace_id → scope.workspace_id). In the Analytics API, actor.email → actor.email_address and summaries → data. [lines 25-52] [Source]
  • The older versions are no longer described as deprecated. [line 25] [Source]

get-started [Source]

Many other API pages (SDK samples and wording only)

  • Across most API guides, and the SDK pages for C#, Go, Java, PHP, Python, Ruby and TypeScript, code samples were updated for the new SDK major versions. For example, C# uses new UserMessageParam { Content = ... } and Ruby uses Anthropic::Model::CLAUDE_* constants. Example JSON output fences were relabeled, and the prose was lightly reworded. These are not listed individually.