Claude Code v2.1.234 Major Updates - New /design skill that can design (research preview)
This page has been translated by machine translation. View original
This is Ishikawa from the Cloud Business Division. Claude Code v2.1.234 (released 2026-08-17) has been released. This is a release with many changes, and the depth of permission-related fixes is particularly notable. I will also try out the new /design skill (research preview) that enables design creation.
The previous update article is here.
Update Summary
v2.1.234 includes 51 changes. In the categories used in this article, the breakdown is: 12 security, 15 fixes, 16 improvements, 4 new features, 2 breaking changes, and 2 performance changes. Nearly a quarter of the total are permission and trust-related changes, making this a release that focuses more on solidifying the foundation than adding features.
In addition, a new /design skill (research preview) that enables design creation has been introduced. I will explain how to actually use it.
Notable Updates
New Features
Sessions automatically continue when usage limits reset. When the claude.ai usage limit resets, sessions now automatically continue. You can turn this off in /config under "Continue automatically at usage limit." For those whose work gets interrupted by hitting usage limits, this should make operations easier since you no longer need to worry about the timing of resumption.
A GitLab merge request badge has been added. It is displayed in the footer and status line. For repositories with a GitLab remote where the glab CLI is authenticated, MR !N is displayed with draft / pending / green status. For teams using GitLab, being able to track merge request status without switching to a browser is something that quietly makes a difference.
Bug Fixes
A crash via third-party gateways has been fixed. A problem where receiving API responses containing thinking blocks missing the thinking field or text blocks missing the text field in the non-streaming fallback path would cause a crash has been fixed. For those using Claude Code via third-party gateways, this may reduce hard-to-diagnose crashes.
False rejections of sandbox commands in auto mode have been fixed. A problem where, after a conversation was compacted in a very long session, auto mode would repeatedly re-evaluate and reject network access for sandbox execution commands has been fixed. The longer the sessions you run in auto mode, the more noticeable the difference this fix makes.
Security
Windows NT namespace paths are now rejected. Remote file reading, session restoration, CLAUDE.md includes, workflow scripts, and file uploads are covered. This hardens the file access paths that remained before pre-authorization against NTLM credential leakage vectors. For those running Claude Code on Windows, updating promptly is worth prioritizing.
Secrets no longer appear in MCP diagnostic output. Scope conflict warnings are displayed in the configured ${VAR} format, and connection failure details now show only the server origin. For those passing MCP server credentials via environment variables, the barrier to sharing diagnostic output with colleagues or in issues is lowered.
Permission preview masking no longer obstructs approval decisions. Commands, paths, and destinations are no longer hidden from approvers by credential masking in relayed permission previews. Large private key blocks are redacted at full strength. For operations where permission approvals are delegated to another person, the necessary information for making approval decisions is finally all there.
Update Details
Items covered in "Notable Updates" are excluded.
New Features
- The
CLAUDE_CODE_PROJECT_DIR_NAMEenvironment variable (optional) has been added. - A
selection:clearkey binding action has been added, allowing you to assign a key to clear text selections within the app. It also works in the agent view.
Security
- Limiting handling of account email addresses: Claude is now instructed to use email addresses only for user identification and not to send them to unrelated services unless requested.
- Fixed issue where session-scoped permission answers were discarded: When answering tool permission prompts from background sub-agents, session-scoped answers including rejections were being discarded.
- Fixed marketplace allowlist host mismatch:
strictKnownMarketplaceswas accepting SCP-format git marketplace sources with hosts different from what git actually connects to. - Fixed loss of permission settings when switching to fullscreen renderer: Accepting the "Try the new fullscreen renderer?" confirmation would restart the session without carrying over permission mode, tool allow/deny rules, or model and effort flags.
- Fixed incorrect answers to permission re-confirmation prompts: If an IDE diff tab was closed during re-confirmation, the previous input could answer the new prompt.
- Additionally, fixes have been made for carrying over
--allowed-tools/--disallowed-toolsduring/tuiswitching, repository-wide scope warnings in trust confirmations, limiting the relay destinations of permission previews, and API token masking gaps, among other permission-related fixes.
Performance
- The context cost of loading the built-in
claude-apiskill has been reduced from over approximately 200k tokens to approximately 25k tokens by making reference documents load on demand. - Mantle: The admin-pin availability check at startup is now skipped if a model has already been selected for the main loop.
Improvements
/permissionscan now be opened during work: Rule changes also apply to the remainder of the current turn./add-dir <path>can also be used during work, and in the fullscreen TUI, the/autocompact,/theme,/help,/config, and/advisordialogs can be opened mid-turn.- Your own prompts are also rendered in Markdown: On the transcript, your input prompts are now displayed with highlighted code blocks, inline code, and lists, just like replies.
- Session titles are now shorter and more specific: Short names (e.g., "Login button bug") are now generated instead of rephrased request sentences (e.g., "Fix the login button on mobile").
- Esc no longer clears text selection: In fullscreen mode, interruptions and screen dismissals work as before, while mouse-selected text remains highlighted.
/goalno longer waits indefinitely: If/goalhas been waiting for more than 30 minutes due to a background task, Claude now checks on the task's status (can be disabled withCLAUDE_CODE_GOAL_CHECKIN_MINUTES=0).- Additionally, minor improvements have been made for permission mode / effort synchronization in Remote Control, session list handling for
SendMessageandListAgents, more detailed API error messages, and automatic clearing of/goalon unrecoverable errors.
Fixes
- Fixed operations around
!shell commands: Issues have been fixed where dialogs such as/permissionsopened during execution would be closed upon command completion, and where queued commands edited with the up arrow key would be sent to the model as plain text. - Fixed behavior of queued messages: Queued messages no longer reappear in the prompt history, and Esc while a selection is active no longer interrupts a turn.
- Fixed file sending in Remote Control: Files sent to users in sessions hosted by the desktop version or VS Code are now uploaded and can be opened on smartphones or the web.
- Fixed startup hang on Windows: When
~/.claude.jsonwas read-only, rename retries would loop and halt startup. - A quietly appreciated fix: A problem where copying text inside modals in fullscreen—such as OAuth URLs from
/login—would result in missing characters has been fixed. For anyone who has had to redo/loginafter failing to copy a URL, this is a quietly effective fix. - Additionally, many minor bugs have been fixed, including Markdown horizontal rule display, rendering speed for messages containing special Unicode, destination resolution for
SendMessage, misreading of git remote hosts, and dropped messages between sessions.
Breaking Changes and Deprecations
The "Default teammate model" setting has been removed from /config. Agent team members now use the leader's model if no model is specified at spawn time. claude setup-token now errors on unexpected additional arguments. Previously, they were silently ignored.
Trying Out the New /design Skill (Research Preview)
The /design skill introduces Claude Design's artboard workflow (a method of presenting multiple options side by side for selection) to the CLI and Desktop, and is built on top of artifacts. Run /design to get an editable artboard for UI — choose one, adjust it, then have Claude implement it.
How to Use (Examples)
/design Spring new menu A4 flyer
/design Internal admin dashboard, mockup is fine
/design SaaS landing page, match existing UI
Just write one line describing what you want to create, and multiple options (artboards) appear laid out on a single pannable/zoomable canvas. After it opens, you can click on elements to edit them directly, and you can also export to PNG/PDF.
What It's Good For
- Screen-related — UI mockups, screen flows, wireframes
- Web — landing pages, banners, social media graphics
- Print — posters, flyers, tri-fold brochures (1 page = 1 artboard), notes and reports (single-page layout)
What Helps Improve Accuracy If Provided Upfront
- Not needed if running inside a repository. It reads the existing design system, components, and tokens, then adds new screens matching the colors, fonts, border radius, and spacing.
- Otherwise: brand guide, reference images, screenshots of existing screens, intended use and size, target audience
- If the visual direction isn't decided yet, I can present 2–4 rough options and proceed by having you choose.
Notes
- For revisions, just say something like "make the header in B bigger" in the conversation, and I'll update the same link.
- This is an early preview of running Claude Design's canvas editor inside Claude Code, so it does not have all the same features as claude.ai/design
Trying "/design Spring new menu A4 flyer"
I'll run the sample prompt "/design Spring new menu A4 flyer" as-is. I'll answer each of the questions.
The target domain is "AWS usage fees / cost management."
The visual style is "Dark, high-density (data-first)."
I confirm my answers with the above content.
A design is generated based on the answers. As the X post says "built on artifacts," it appears to be a repurposing of artifacts.
In my environment, once generation was complete, a browser automatically launched and three design options were displayed side by side.
I was able to select design properties and immediately change, reflect, and verify them in the Claude Design canvas editor.
Once changes are complete, you can download in any of the following formats:
- PNG 1x (1360 x 940 px)
- PNG 2x (2720 × 1880 px)
- HTML (.zip) (Page + its images and runtime)
It's very convenient that the artifacts cultivated in Claude Cowork can be explicitly used in Claude Code as well, through a method called Skills.
Conclusion
v2.1.234 feels like a release that collectively plugged the gaps in permission handling. Session-scoped rejections being dropped, allow/deny rules being lost on restart, and approvers not being able to see the critical content — these are all the kind of bugs that fall into the "thought it was working but it wasn't" category. The /design skill (research preview) is also an exciting update to look forward to going forward.
References
