
Claude Code v2.1.233 Major Update - Todo Tool Disabled by Default
This page has been translated by machine translation. View original
This is Ishikawa from the Cloud Business Division. Claude Code v2.1.233 (released 2026-08-14) has been released. Since there is a breaking change involving the default disabling of the Todo tool depending on the model, we recommend checking the impact.
The previous update article is here.
Update Summary
v2.1.233 includes 20 changes. The breakdown is 4 new features, 6 improvements/developer experience changes, 8 bug fixes, 1 security fix, and 1 breaking change. Bug fixes specific to Windows and Linux are notable, and it also includes regression fixes and reversions of changes introduced in 2.1.232.
Notable Updates
GitLab Merge Request URLs Are Now Supported
The --worktree flag and claude agents view now support GitLab merge request URLs. In the claude agents view, merge requests are displayed in the !N format.
For organizations using GitLab, the workflow of branching a worktree for each merge request and working in parallel is now directly usable, and I feel the gap in navigation that previously assumed GitHub has narrowed.
Memory Limits Can Now Be Set for Bash Execution on Linux
An opt-in memory cgroup limit setting CLAUDE_CODE_TOOL_MEMORY_LIMIT has been added for Bash tool execution on Linux. This prevents runaway builds from stopping sessions.
For those running Claude Code on CI, self-hosted runners, or shared development servers, I feel this is a setting worth keeping in mind.
CPU Pinning During Idle on Linux Has Been Resolved
A bug where idle sessions with sandboxing enabled on Linux would continuously use 100% of one CPU core has been fixed.
For those who keep sessions open with sandboxing enabled on Linux, if you've noticed fans running continuously even while idle, I feel it's worth updating.
The Bug Where Auto Mode Repeatedly Requested Approval on Windows Has Been Resolved
A bug (regression from 2.1.232) where auto mode repeatedly requested manual approval for normal Bash commands like cd <dir> && <command> > file on Windows has been fixed.
For those who regularly use auto mode on Windows and felt that approval requests increased since 2.1.232, I expect this fix resolves that.
Path Validation Bypass Issue on Windows Has Been Fixed
An issue where Windows paths written with the NT device prefix \??\ could bypass UNC path validation has been fixed, closing a pathway that could lead to NTLM credential leakage.
For those who have Claude Code handle externally obtained repositories or files on Windows, I feel this is worth prioritizing for an update.
Update Details
New Features
- The
--worktreeflag andclaude agentsview now support GitLab merge request URLs (MRs are displayed in!Nformat in theclaude agentsview) - An opt-in setting
forward_user_identityhas been added to the apps gateway for Anthropic upstream, which sends the signed-in user's identity as a header. Proxies under the gateway can now allocate usage on a per-user basis - An opt-in memory cgroup limit setting
CLAUDE_CODE_TOOL_MEMORY_LIMIThas been added for Bash tool execution on Linux - An environment variable
CLAUDE_CODE_WEBFETCH_CACHE_TTL_MShas been added to configure the WebFetch session URL cache TTL (the default value remains 15 minutes)
Improvements
- Session startup for
claude self-hosted-runneris now faster. Session branches are now created without rewriting the working tree, and 2 server round-trips no longer block agent startup - Error forwarding for apps gateway has been improved, so 400/413 errors returned by Vertex, Foundry, and Claude Platform on AWS upstreams now retain the upstream's own messages. A bug with auto-compact via apps gateway has also been fixed
claude plugin validatenow also inspects plain.claude/skillsdirectories and reports SKILL.md files that fail frontmatter parsing- Screen reader mode has been improved: the
/effortselector is now displayed as a numbered list with a numbered input prompt, and hint and dialog text no longer gets cut off - Print mode diagnostics have been improved: when a request is sent with a model ID that Claude Code does not recognize, a
[claude-code:unrecognized_model]line is now output to stderr. This can be suppressed by mapping it withmodelOverrides - For repositories where the origin remote is gitlab.com or bitbucket.org, the GitHub app setup prompt is no longer displayed. The enterprise marketplace prompt still covers non-GitHub internal git hosts
Bug Fixes
- Fixed CPU pinning during idle on Linux: A bug where idle sessions with sandboxing enabled would continuously use 100% of one CPU core has been fixed
- Fixed unnecessary approval requests in auto mode on Windows: A bug (regression from 2.1.232) where auto mode repeatedly requested manual approval for normal Bash commands like
cd <dir> && <command> > filehas been fixed - Fixed repeated stream reopening in MCP v2: A bug where the subscriptions/listen stream would endlessly reopen for servers that disconnect long-held streams with a fixed timeout (such as serverless hosts) has been fixed
- Fixed false "lost" determination for cloud sessions: A bug where cloud sessions could be marked as lost when the environment shuts down while waiting for a permission confirmation prompt has been fixed
- Fixed Unknown command for built-in skill aliases: A bug where aliases such as
/checkupand/reviewwould result in "Unknown command" in-pmode or when loading plugins/MCP, when user or project skills were shadowing built-in skills, has been fixed - Fixed argument expansion in skills/commands: Argument values are no longer re-expanded as template markers
- Partially reverted 2.1.232 changes around Bash permissions: Changes related to Windows Cygwin-style symbolic links and input redirection (
< file) have been reverted. A more narrowly scoped version is planned for reintroduction in a later release - A quietly welcome fix: A bug where Notification hooks would not fire on permission confirmation prompts when running on Claude Desktop or VS Code has been fixed. For those using hooks to send notifications at work breakpoints, I feel this is a fix that quietly makes a difference
Security
- An issue where Windows paths written with the NT device prefix
\??\could bypass UNC path validation has been fixed, closing a pathway that could lead to NTLM credential leakage
Breaking Changes / Deprecations
-
Todo/task tracking tools (TaskCreate/Get/Update/List, TodoWrite) are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and later models. They can be re-enabled by setting
CLAUDE_CODE_ENABLE_TODO_TOOLS=1.# Not available by default on target models. Set the environment variable to use them as before CLAUDE_CODE_ENABLE_TODO_TOOLS=1 claude-
Claude will not create task lists
Since TaskCreate / TaskGet / TaskList / TaskUpdate cannot be called, a progress checklist will not be presented even for lengthy tasks. Instead, Claude will explain the plan in prose and proceed. In situations where you want to track "how far along we are" in a list for multi-step tasks, the presentation will change.
-
Reduction in round-trips and cost
In actual measurements, the same task went from 7 turns → 5 turns, output 2,286 → 1,373 tokens, $0.405 → $0.323 (approximately 25% reduction). This is because the 10 tool calls previously spent on task registration and updates are eliminated.
-
Sub-agent related features are not affected
Task (sub-agent launch), TaskOutput, and TaskStop remain available by default. Despite the similar names, these are different things, and /workflows and background agent operations remain unchanged.
-
In Closing
v2.1.233 has one breaking change, but overall I feel it is a release centered on fixes and foundational improvements. If your workflow does not depend on the Todo tool, it should be straightforward to update.
For those using Claude Code on Windows or Linux, and for teams doing development with GitLab, I feel this is particularly worth reviewing. Why not update and give it a try?
References
