Source: ai-research/claude-code-docs-changelog-2026-10-23.md — code.claude.com changelog, versions v2.1.303–v2.1.309 (October 18–22, 2026), fetched 2026-10-23. The official What’s New digest page code.claude.com/docs/en/whats-new/2026-w43 returned HTTP 404 (consistent with all prior sweeps since W38; Tavily OAuth unavailable non-interactive; x-search ENOENT). Continues Week 42.

Product: Claude Code · Date: 2026-10-18 to 2026-10-22

Week 43 delivers safety and security hardening alongside new tooling for Code Review subscribers and Auto Mode users. Highlights: /auto-mode-setup now available on all major cloud providers and gateway deployments (v2.1.304); Code Review levels (Quick/Standard/Deep) with per-org and per-repo control (v2.1.304); the new claude_code.safety_event OpenTelemetry event for complete audit trails of blocked tool calls (v2.1.307); @claude review stop to pause Code Review always-on subscriptions (v2.1.307); decider: "person" for mod tool.check forcing human approval (v2.1.304); claude ultrareview now requires workspace trust (v2.1.309); and sed -i behavior changed to use the standard Bash prompt — no longer restored by /rewind (v2.1.309).

Key Takeaways

  • /auto-mode-setup expands to all cloud/gateway platforms (v2.1.304): the guided auto-mode tuning command now works on Bedrock, Vertex, Foundry, Claude apps gateway, and telemetry-disabled sessions — previously terminal-only.
  • Code Review levels: Quick / Standard / Deep (v2.1.304): organization owners can now set a default review depth. Individual repositories can override the org default.
  • @claude review stop (v2.1.307): ends an @claude review always Code Review subscription on a PR without triggering a new review. Also accepts off and pause.
  • claude_code.safety_event OTel event (v2.1.307): emitted for every tool call stopped, prompted, or sent back by a permission rule, auto mode, a safety check, a hook, or the sandbox. Enables complete security audit trails in observability pipelines.
  • decider: "person" in mod tool.check (v2.1.304): a Mod can require that only a human (via permission dialog or host callback) approves a tool call — refused with the mod’s reason where no human can be asked.
  • e.isReadOnly in mod tool.check (v2.1.304): lets Mods detect and handle read-only tool calls differently from write calls.
  • Mods settings.read hooks rewritten (v2.1.304): now receive each settings layer and can rewrite merged settings — enables managed Mods to set an organizational settings ceiling (including retention via cleanupPeriodDays).
  • BREAKING: claude ultrareview requires workspace trust (v2.1.309): on a terminal, Claude asks for trust before proceeding; in scripts, pass --trust-workspace (before PR number/base branch) or set CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1.
  • BREAKING: sed -i now uses standard Bash prompt (v2.1.309): sed -i runs as written and is no longer restored by /rewind. Update any workflows that relied on /rewind to undo sed -i changes.
  • sandbox.allowNestedUserNamespaces (v2.1.305): allows sandboxed Linux commands (Chromium, rootless podman) to create user namespaces; disables the sandbox’s file-protection bypass guard.
  • allowedMcpServers no longer filters env-var servers (v2.1.305): managed-mcp.json stdio servers that use ${VAR} only in env now load even if not on the allowlist; use deniedMcpServers to block them.
  • /claude-api api-version (v2.1.307): updates bundled API code to use API version 2026-10-21.
  • /claude-api uses multiagent_20261001 with dynamic workflows (v2.1.309): Managed Agents subagent teams now default to the new API type with dynamic workflows on; replaces coordinator.
  • Dataviz skill accepts APCA contrast (v2.1.309): in addition to WCAG 2 ratios (still required where AA conformance is needed), the bundled dataviz skill now validates contrast with the APCA algorithm and advises white or near-black label text inside colored fills.
  • Tool search on by default behind custom ANTHROPIC_BASE_URL (v2.1.304): gateways that return 400/422 are retried without tool search automatically.
  • Dangerous rm/rmdir in bypass mode denied immediately (v2.1.309): Desktop and VS Code no longer prompt — the command is refused outright with a rewrite hint, so unattended sessions continue.
  • Compaction resends skill list (v2.1.304): after a compaction event Claude receives the full skill list again; /reload-skills now sends only new or changed skills, not the full list.

Release table

VersionDate (UTC)Shape
v2.1.3032026-10-18Bug fixes and reliability improvements only
v2.1.3042026-10-19/auto-mode-setup expanded; Code Review levels + org-wide toggle; decider:person + e.isReadOnly in mod tool.check; settings.read hook rewrite; /claude-api managed-agents-secure; tool names in compaction summary; 43 bug fixes
v2.1.3052026-10-20sandbox.allowNestedUserNamespaces; CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS; allowedMcpServers env-var fix; mod pane width 45% cap; 22 bug fixes
v2.1.3062026-10-21Two targeted bug fixes (claude -p Not logged in; prompt-too-long misreport with many MCP tools)
v2.1.3072026-10-21/claude-api api-version (2026-10-21); claude_code.safety_event OTel; @claude review stop; WebSearch budget notice; self-hosted runner checkout hook timeout 30min; 26 bug fixes
v2.1.3082026-10-22Bug fixes and reliability improvements only
v2.1.3092026-10-22Dataviz skill APCA; sed -i standard Bash prompt (BREAKING); ultrareview workspace trust (BREAKING); rm/rmdir bypass-mode deny-at-once; /claude-api → multiagent_20261001 + WIF OTel; sandbox write-list ask; 62 bug fixes

/auto-mode-setup on all cloud platforms (v2.1.304)

/auto-mode-setup is a guided command that reads your project and recent sessions to configure auto mode for your environment. It was previously only available in standard terminal sessions. v2.1.304 extends support to:

  • Amazon Bedrock
  • Google Vertex AI
  • Microsoft Azure Foundry
  • Claude apps gateway deployments
  • Sessions with telemetry disabled

This makes the guided setup accessible to enterprise and self-hosted deployments that route through a cloud provider or an internal API gateway. See Auto Mode.

Code Review levels: Quick / Standard / Deep (v2.1.304)

Organization admins can now configure the depth of Code Review analysis in admin settings:

  • Quick — fastest, lighter signal
  • Standard — the default (unchanged behavior)
  • Deep — most thorough analysis

The org-level default can be overridden per repository. Additionally, admins can now opt an entire GitHub organization into Code Review coverage (all current and future repositories) from a single admin toggle, without selecting repositories individually.

See Code Review GA.

@claude review stop (v2.1.307)

Code Review subscribers who have enabled @claude review always on a pull request can now end the subscription without triggering a new review:

@claude review stop
@claude review off
@claude review pause

All three variants are equivalent. The subscription ends; the next push to that PR will not trigger an automatic review. Previously, there was no way to disable review always without deleting the PR comment.

See Code Review GA.

claude_code.safety_event OpenTelemetry event (v2.1.307)

Removed in v2.1.312 (2026-10-27): Claude Code stopped sending this event. See What's New — Week 44.

A new OTEL event emitted for every tool call that a safety mechanism stopped, prompted, or sent back:

  • Triggers: permission rule match, auto mode decision, safety check, hook, or sandbox intercept
  • Useful for: security auditing, compliance audit trails, debugging unexpected tool denials, visibility into what the sandbox is blocking

This complements the existing Agent Observability OTel instrumentation (spans from Managed Agents, Messages API, SDK) with fine-grained per-tool-call safety events. Each event records what stopped the tool call and why.

Mods: decider: "person" and e.isReadOnly (v2.1.304)

Two additions to the Mods tool.check API:

decider: "person" — a Mod can return this from its tool.check handler to require that a human approve the tool call:

  • If a permission dialog is available, it is shown
  • If no human can be asked (unattended session, no host callback), the call is refused with the Mod’s reason

This is distinct from the default auto-approve/deny behavior: it guarantees the decision always goes through a human gate.

e.isReadOnly — the tool.check event now includes isReadOnly: true for tool calls that are read-only (e.g., Read, Glob, Grep). Mods can use this to apply different policies to read vs. write operations.

See Claude Mods.

Mods settings.read hooks rewritten (v2.1.304)

settings.read hooks in Mods now receive each settings layer (user, managed, org, repo, etc.) and can rewrite the merged result. This enables managed Mods to:

  • Set a ceiling on any setting, including cleanupPeriodDays (the retention sweep period)
  • Apply org-wide policy overrides before the settings reach the session
  • Enforce that certain settings cannot be overridden by per-session or per-repo config

Previously, settings.read hooks could only observe settings, not modify them per-layer. See Claude Mods.

BREAKING: claude ultrareview requires workspace trust (v2.1.309)

claude ultrareview now requires workspace trust before it can run:

  • In a terminal session: Claude prompts for workspace trust before proceeding
  • In scripts or CI: pass --trust-workspace as a flag (place it before the PR number or base branch argument), or set CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1
# Script usage:
claude ultrareview --trust-workspace main
claude ultrareview --trust-workspace 123
 
# Env var override:
CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1 claude ultrareview main

Unattended pipelines that call claude ultrareview without workspace trust will now fail. Update CI scripts accordingly.

BREAKING: sed -i uses standard Bash prompt, no longer restored by /rewind (v2.1.309)

Prior to v2.1.309, sed -i ran in a simulated file-edit preview mode that /rewind could restore. Now:

  • sed -i runs as a standard Bash command
  • /rewind will NOT restore sed -i changes, the same as other Bash edits
  • The standard Bash permission prompt applies

If you had workflows that used /rewind to undo sed -i results, those workflows need to be updated to use explicit backup-and-restore patterns instead.

sandbox.allowNestedUserNamespaces (v2.1.305)

A new sandbox configuration field allowing sandboxed Linux commands to create user namespaces:

{
  "sandbox": {
    "allowNestedUserNamespaces": true
  }
}

Required for tools that internally use user namespaces, such as Chromium (headless browser) and rootless Podman. Note: enabling this allows the sandboxed process to work around the sandbox’s file-system protections — enable only when the subprocess is trusted.

allowedMcpServers behavior change for env-var servers (v2.1.305)

managed-mcp.json stdio MCP servers that specify ${VAR} substitutions only in the env field — not in the command or args — now load unconditionally, even if not listed in allowedMcpServers. Previously they were filtered out by the allowlist check.

Migration: if you have an env-var-only server that should remain blocked, add it explicitly to deniedMcpServers instead of relying on its absence from allowedMcpServers.

See MCP.

Other additions and changes

  • /claude-api managed-agents-secure (v2.1.304): a new diagnostic that checks a Managed Agents agent’s settings against Anthropic’s published security guide, without modifying anything.
  • Tool names in compaction summary (v2.1.304): Ctrl+O transcript view now shows which tools were called in the section being compacted.
  • Pro/Max plan resumption warning (v2.1.304): a warning appears when resuming an inactive conversation would consume a significant fraction of the 5-hour usage limit.
  • CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS (v2.1.305): for unattended sessions using CLAUDE_CODE_RETRY_WATCHDOG, this env var controls how frequently the watchdog rechecks whether a usage limit has been lifted (default: check only at scheduled reset).
  • /claude-api api-version (v2.1.307): updates the bundled API code examples to use the new API version 2026-10-21 (the multiagent_20261001 API type, which enables dynamic workflows by default).
  • Skill hook limits (v2.1.304): a single skill source can register at most 1,024 hooks and 16,384 matcher characters per session.
  • Dataviz skill APCA contrast (v2.1.309): the bundled dataviz skill now validates accessible-perceptual-contrast-algorithm contrast in addition to WCAG 2 — includes an apca(text, background) function in the palette validator and a white/near-black label recommendation for chart fills.
  • claude -p cloud session fix (v2.1.306): fixed “Not logged in” errors in claude -p when started without the session’s environment variables.
  • Self-hosted runner: checkout hook 30-minute timeout (v2.1.307): checkout hooks are now stopped after 30 minutes by default (use --checkout-hook-timeout-min or 0 for no limit).
  • Mod docked pane width cap (v2.1.305): a mod’s docked pane stays within 45% of the terminal width.
  • Dangerous rm/rmdir in bypass mode (v2.1.309): in Desktop and VS Code, bypass-permissions rm and rmdir are now denied immediately with a rewrite hint; the session continues uninterrupted.

Breaking and behavior changes

  • claude ultrareview requires workspace trust (v2.1.309): add --trust-workspace to CI scripts. See above.
  • sed -i no longer restored by /rewind (v2.1.309): update workflows relying on rewind for sed undo. See above.
  • allowedMcpServers no longer filters env-var servers (v2.1.305): use deniedMcpServers instead. See above.
  • Skill hook limits enforced (v2.1.304): 1,024 hooks / 16,384 matcher chars per source per session.
  • /artifacts delete confirmation changed to numbered choices (v2.1.304/v2.1.305): Enter/1 to keep, 2 to delete; y/n no longer work as of v2.1.305.
  • axScreenReader / CLAUDE_AX_SCREEN_READER (v2.1.307): now only applies when set from user, managed, or --settings settings sources; ignored from repo/project settings.
  • Tool search on by default behind custom ANTHROPIC_BASE_URL (v2.1.304): gateways that don’t support tool search return 400/422 and Claude retries without it automatically.

Try It

  • Run /auto-mode-setup if you’re on Bedrock, Vertex, or Foundry and haven’t tuned auto mode for your project yet.
  • Check claude ultrareview CI scripts: any automated pipeline calling claude ultrareview needs --trust-workspace added.
  • Audit sed -i usage: if your workflow uses /rewind to undo sed -i, replace with explicit backup-restore patterns.
  • Add deniedMcpServers for any env-var-only MCP servers that should remain blocked (no longer filtered by allowedMcpServers).
  • For Code Review admins: visit admin settings to configure review depth (Quick/Standard/Deep) and optionally enable org-wide coverage.
  • Use claude_code.safety_event in your OTel pipeline to build a complete audit trail of tool call decisions. No longer sent as of v2.1.312 (Week 44).

Open Questions

  • v2.1.297 and v2.1.298 (unresolved from W42): these fall between v2.1.296 (Oct 9, W41) and v2.1.299 (Oct 14, W42). Still not detailed in available changelog summaries.
  • sandbox.allowNestedUserNamespaces and file protection scope: the changelog notes this “lets them get around the sandbox’s file protections” but does not specify which protections or what the safe-use boundary is.
  • /auto-mode-setup and existing auto-mode config: does it merge with or overwrite existing settings.json auto-mode configuration?
  • multiagent_20261001 API type: what is in the 2026-10-21 API version that requires a new API type? The changelog points to a new /claude-api api-version skill command but does not detail the API changes.
    • Partly answered 2026-10-27: the cookbook’s plan-big-execute-small recipe now sets multiagent_20261001, lists the worker under subagents.predefined_agents, disables subagents.inline_agents and workflows, and uses create_agent / send_to_agent / list_agents / send_to_parent; the SDKs took the API version as a breaking change (Python v2.0.0, TypeScript sdk-v1.0.0). See Anthropic SDK Releases — October 2026.
  • What’s New — Week 42 — the preceding digest (idle compaction, BREAKING MCP allow-rule scope fix, fork subagents outside terminal).
  • Auto Mode — /auto-mode-setup expanded to all cloud platforms.
  • Code Review GA — review levels (Quick/Standard/Deep); @claude review stop.
  • Agent Observability — claude_code.safety_event OTel event for tool call auditing.
  • Claude Mods — decider:person, e.isReadOnly, settings.read hook rewrite.
  • MCP — allowedMcpServers env-var behavior change.
  • Hooks — skill hook limits (1,024 hooks / 16,384 matchers per source).
  • CLI Reference — /auto-mode-setup, /claude-api api-version, /claude-api managed-agents-secure, sandbox.allowNestedUserNamespaces, CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS.
  • Scheduled Tasks — CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS for unattended retry mode.