RSS
Tools

Claude Code 2.1.328: /code-review now reports only bugs, auto mode for Opus 4.6 and Sonnet 4.6 on Bedrock and Foundry

Claude Code v2.1.328 makes /code-review report only bugs and CLAUDE.md violations, with cleanups moved to /simplify. It also brings auto mode to Opus 4.6 and Sonnet 4.6 on Bedrock, Agent Platform and Foundry, and lets mcp_tool hooks block when they fail.

Magnifying glass on white table illustrating Claude Code 2.1.328
Photo by Markus Winkler on Unsplash

Anthropic shipped Claude Code v2.1.328 on November 19, 2026. The release notes list well over a hundred changes. We picked three that matter most for daily use. First, /code-review now reports only bugs and CLAUDE.md violations on most models, and cleanup suggestions move to /simplify. Second, auto mode now covers Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway. Third, mcp_tool hooks can now block an action when the hook itself fails. The facts come from the GitHub release notes and the Claude Code documentation. Suggestions are ours and labelled as such.

/code-review now reports only bugs

On most models, /code-review now reports only bugs and CLAUDE.md violations. Each finding is tagged Important, Pre-existing or Nit, and each summary ends with a suggested fix. A clean review now says "Code review found no issues." If you want suggestions about reusing existing helpers, simplifying code or efficiency, the release notes say to run /simplify.

The split is easier to follow with each command's current job, as the commands reference describes it:

  • /code-review [low|medium|high|xhigh|max|ultra] [--fix] [--comment] [target]: reviews the current diff, a PR number, a branch or a path for correctness bugs. --fix applies the findings, --comment posts them to the GitHub PR or GitLab merge request, and ultra runs the deep multi-agent review in the cloud. A session-scoped scheduled task never launches the cloud review. /review is an alias.
  • /simplify [target]: four review agents run in parallel, covering reuse of existing helpers, simplification, efficiency, and whether the change sits at the right level of abstraction. It then applies the fixes. It doesn't look for correctness bugs.
  • /security-review: checks the diff for security vulnerabilities.

Before this release, depending on model and effort level, a /code-review run could mix bug reports with cleanup ideas in one list. Now the two are separate. The release notes give the three labels but don't say what each one means or which models count as "most models". Don't read more into a label than its name. Our suggestion: if a script or CI step parses /code-review output, test it after updating, because the labels and the clean-review line have changed.

There's an older naming change to keep in mind too. The documentation notes that /simplify was the name of the bug-finding command before v2.1.147, when it applied fixes by default. If an old script still calls /simplify expecting bug reports, switch it to /code-review --fix.

Auto mode for Opus 4.6 and Sonnet 4.6 on Bedrock, Agent Platform and Foundry

Auto mode lets Claude Code act without routine permission prompts. A separate classifier model reviews actions first and blocks anything that goes beyond your request. Before this release, the permission-modes documentation listed only Sonnet 5 or later, Opus 4.7 or later, Haiku 5.5 and the Fable models as supported on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and signed-in Claude apps gateway sessions. Version 2.1.328 adds Claude Opus 4.6 and Sonnet 4.6 on those providers. An interactive session with no permission mode set now starts in auto mode.

The release also changes plan mode and Bash prompts for Opus 4.6 and Sonnet 4.6 on third-party providers so they use and offer auto mode, as newer models already did. Setting disableAutoMode turns both off. The people affected are those running Opus 4.6 or Sonnet 4.6 on these providers. An interactive session that used to start in Manual can now start in auto mode. Older models such as Sonnet 4.5 and Opus 4.5 still aren't supported. Non-interactive claude -p and Agent SDK runs need an explicit permission mode (--permission-mode auto for claude -p) to start in auto mode. If you want to keep Manual, set a permission mode explicitly, or use permissions.disableAutoMode in managed settings. The docs also note that defaultMode "auto" doesn't take effect from .claude/settings.json or .claude/settings.local.json; it belongs in ~/.claude/settings.json. When we fetched the permission-modes page on November 20, its provider list hadn't yet added the 4.6 models, so treat the release notes as the current word.

Two related fixes are in the same release. One: auto mode no longer fails every turn outside Claude Desktop when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS is set behind a gateway that rejects betas, on Bedrock, Agent Platform, Foundry or the Claude apps gateway. Two: CLAUDE_STREAM_IDLE_TIMEOUT_MS now does lengthen the 5-minute connection limit on gateways, custom base URLs and cloud providers.

mcp_tool hooks can now fail closed

Claude Code hooks can call a tool on a configured MCP server (type: "mcp_tool") instead of running a shell command. By default, when a hook fails, the action goes ahead. So a policy hook whose MCP server isn't connected would let everything through. Command and HTTP hooks have had "onFailure": "block" since v2.1.295. Version 2.1.328 adds it to mcp_tool hooks. Here is a PreToolUse example that blocks Bash calls when the policy server can't answer. The fields follow the hooks documentation, and the server and tool names are placeholders:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "mcp_tool",
            "server": "policy_server",
            "tool": "check_command",
            "input": { "command": "${tool_input.command}" },
            "onFailure": "block"
          }
        ]
      }
    ]
  }
}

For an mcp_tool hook, the documentation counts these as failures: the server isn't connected, the tool returns an error, the hook times out, or its output can't be parsed. With block set on events such as PreToolUse or UserPromptSubmit, a non-empty answer also counts as a failure unless it is {} or JSON that sets a hook output field other than reason. A blocked failure acts like exit code 2 on that event, except on PermissionRequest, where it denies the request. The flag has no effect on Stop, SubagentStop, TaskCompleted and TeammateIdle hooks. One more catch: SessionStart at launch (including --continue and --resume) and every Setup event fire before MCP servers are available, so Claude Code skips mcp_tool hooks there. The docs list other edge cases, including Windows without Git Bash, background command hooks and Code Mode's PreToolUse handling. Our caution: with block set, a flaky server or a timeout will stop ordinary work too.

Other changes worth knowing

  • Fixed Bash commands in zsh stopping with "not found" at a word starting with =, such as echo ==== or [ a == b ].
  • Fixed automatic compaction reusing the earlier system prompt instead of rebuilding it the way /compact does.
  • Fixed saving a scheduled task erasing the project's other saved tasks when the scheduled tasks file can't be read (on third-party providers or with telemetry or feature flags off). Separately, fixed a task with a typo in that file being deleted on the next save.
  • Code Mode: removed a 0.75-second pause after most scripts in auto and bypass modes, shortened tool descriptions to save tokens, and fixed several resume bugs. Code Mode also no longer runs UserPromptSubmit hooks (managed ones included) on scripts' output. To redact that output, use a PostToolUse hook with updatedToolOutput.
  • Managed PostToolUse hooks no longer fire for calls that were refused before they ran.
  • New CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME lets a host stop claude -p --resume and --continue from running a tool call a hook deferred.
  • Self-hosted runner: --repository-folder for the runner at --capacity 1, --hook-label for claude self-hosted-runner orchestrator, --metadata key=value for claude --environment, and new orchestrator metrics for queue wait by reason and runners given up.

A prompt to check your setup after updating

Try it with your AI. Give this to Claude Code in a repository after updating to v2.1.328. It reads your settings and reports; it changes nothing until you approve.

I just updated Claude Code to v2.1.328. Check my setup against three changes in that
release.
1. First confirm the installed Claude Code version is 2.1.328 or later and tell me how
   you checked. If it is older, stop and tell me.
2. Read every settings layer you can see: ~/.claude/settings.json, this project's
   .claude/settings.json and .claude/settings.local.json, plugin settings and visible
   managed settings. List every hook with type "mcp_tool", its event, server, tool and
   whether onFailure is set. For any PreToolUse or UserPromptSubmit mcp_tool hook that
   acts as a policy check, explain what "onFailure": "block" would block if the server
   is down or slow, and whether that is acceptable. Show me any proposed diff and wait
   for my confirmation before editing.
3. Search this repo's scripts and CI config for calls to /simplify or /code-review. Tell
   me if any expect bug reports from /simplify, or parse /code-review output in a way
   the new Important / Pre-existing / Nit labels or the line "Code review found no
   issues." would break.
4. Tell me which model and provider this setup uses. If it is Opus 4.6 or Sonnet 4.6 on
   Bedrock, Google Cloud's Agent Platform or Microsoft Foundry, explain that interactive
   sessions without a permission mode now start in auto mode, and that a defaultMode
   setting only takes effect from ~/.claude/settings.json, not from the project files.
Only read files; do not edit anything without my confirmation, do not touch credentials,
and do not delete anything.

FAQ

Why does /code-review no longer suggest cleanups? From v2.1.328, on most models it reports only bugs and CLAUDE.md violations. Run /simplify for reuse, simplification and efficiency suggestions.

Does auto mode work with Opus 4.6 on Amazon Bedrock? Yes, from v2.1.328, for interactive sessions. It also works with Sonnet 4.6, and on Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway. claude -p and Agent SDK runs need an explicit auto permission mode. disableAutoMode turns it off.

Which Claude Code version supports onFailure on mcp_tool hooks? v2.1.328 or later. Command and HTTP hooks have supported it since v2.1.295.

Source: Claude Code v2.1.328 release notes, Claude Code docs: Review locally with /code-review, Claude Code docs: Commands, Claude Code docs: Permission modes, Claude Code docs: Hooks

This story first appeared in the claudefox daily briefing for November 20, 2026.