Claude Code v2.1.228~v2.1.229 Major Updates - Concentrated Crash Fixes and /commit-push-pr Safety Improvements

This is Ishikawa from the Cloud Business Division. Claude Code v2.1.228 and v2.1.229 have been released. I had been watching v2.1.228 quietly, but with two consecutive days of related updates, including breaking changes and changes that could affect automation flows, I feel this is content worth reviewing before updating.

The previous update article is here.

https://dev.classmethod.jp/articles/20260811-cc-updates-v2-1-227/

Update Summary

This covers 2 versions (v2.1.228–v2.1.229, 2026-08-11–2026-08-12) with a total of 50 changes. The breakdown is: fixes account for the majority at 29, improvements 9, new features 3, security 3, breaking changes 2, documentation 2, performance 1, and developer experience 1. Crash and display corruption fixes, along with adjustments around self-hosted runners, are particularly notable.

Notable Updates

Added command source to plugin marketplace (v2.1.229)

command has been added as a source for the plugin marketplace. A local command (such as an IDE) outputs a plugin directory, and the result is re-resolved per session and applied without restarting. When mode: "link" is specified, the output location is used as-is.

I feel this opens up room for environments where the plugin location changes on the IDE side to follow along without rewriting the marketplace definition.

Fixed display disappearing and double output during streaming (v2.1.229)

An issue where long responses would partially disappear during streaming and then be output twice to the terminal has been fixed.

This is a fix that comes into play when working while reading long responses, and I think it's easy to understand as a motivation for updating.

Fixed issue where project memory folder was being deleted (v2.1.228)

An issue where session cleanup processing was deleting the contents of the project's memory folder has been fixed.

I feel the impact is greater for those who continuously cultivate their project memory.

Fixed auto mode failing on all tool calls (v2.1.229)

An issue where auto mode was failing on all tool calls for users who had disabled the attribution header with CLAUDE_CODE_ATTRIBUTION_HEADER (direct connection to Anthropic API) has been fixed.

Organizations that had disabled the attribution header with the environment variable must have been in a state where auto mode was completely unusable, so if this sounds familiar, I feel this is a version you'll want to upgrade to first.

Enhanced handling of skills synced from claude.ai (v2.1.228)

For skills synced from claude.ai, they no longer overwrite local commands or MCP prompts, and description text is now sanitized and labeled. Additionally, ! command execution and @ file expansion in skill bodies no longer occur in local environments.

This looks like a change that becomes more impactful as the practice of bringing skills created on the claude.ai side into local environments spreads.

IPv6 notation and fail-closed behavior for sandbox (v2.1.229)

In the sandbox network domain list, IPv6 literals are now expressed with brackets ([::1]:443). Ambiguous notation is handled as fail-closed (blocking side), and /doctor will flag this.

Since situations where allowlist entries could pass through with ambiguous notation will be reduced, I feel it's worth checking for environments running sandboxes in production.

Target Versions and Period

New Features

• Added command source to plugin marketplace. A local command outputs a plugin directory and is re-resolved per session (v2.1.229)
• Support for server-provided Claude Code hooks in self-hosted runner sessions, same as managed environments (v2.1.229)
• [VSCode] Added session groups to sidebar. Can be created, renamed, and deleted via right-click, and multiple sessions can be moved together with Cmd/Ctrl click or Shift click (v2.1.229)

Improvements

• Added SSE keepalive pings to gateway streaming responses to prevent idle timeout disconnections during long thinking periods (targets Vertex / Bedrock upstream, v2.1.229)
• Changed workflow fan-out to stagger startup times of sibling agents with the same prefix. Subsequent agents read the cached prompt prefix (CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS=0 to disable, v2.1.229)
• Changed Vertex AI to fail in seconds rather than retrying for minutes when credentials are expired or not set (v2.1.228)
• During compaction, now displays retry countdown and hints when stalled, not just a progress bar (v2.1.228)
• Changed cross-session messages to display sender and body inline rather than as collapsed lines. For Remote Control sessions on other machines, your own Remote Control session name is displayed as the sender (v2.1.228)
• Changed Write tool behavior so that new models can overwrite existing files not yet read in that session (same rules as the Edit tool. Existing models still require prior reading, v2.1.228)
• Changed "prompt is too long" error to explain why auto-compaction could not recover, rather than just recommending /compact (v2.1.229)
• ListAgents now labels disconnected Remote Control sessions as offline and your own cloud sessions as cloud (v2.1.229)
• Added claude remote-control --continue to documentation for resuming the most recent Remote Control session (v2.1.229)
• Warning about overriding with CLAUDE_CODE_OAUTH_TOKEN is now shown again even after a successful /login (v2.1.229)
• [VSCode] /btw side question panel can now be resized by dragging the border (supports both side-docked and stacked layouts, v2.1.229)
• Other adjustments include: reduced tab bar flickering from character changes in terminal title busy spinner (v2.1.228), changed VSCode "Report a problem" and /bug destination to the built-in feedback dialog (v2.1.229), and removed outdated notes from first-use notifications for Pro / Max / Team plans (v2.1.228).

Fixes (Major)

Excerpting fixes primarily related to stability and usability.

• Fixed interactive session screen redraw stopping: An issue where, after a rare internal layout error, the process continued running but screen redrawing stopped completely (v2.1.228)
• Fixed erroneous deletion of memory folder: An issue where session cleanup processing was deleting the contents of the project's memory folder (v2.1.228)
• Fixed streaming display disappearing and double output: An issue where long responses would partially disappear mid-stream and be output twice to the terminal (v2.1.229)
• Fixed crash from non-string parameters: An issue where glob / file_path / command values in tool calls being non-string would crash to the error screen (including when --resume-ing the affected session, v2.1.229)
• Fixed RangeError crash in narrow terminals: An issue where rendering progress bars or Markdown tables in very narrow windows would crash (could also cause startup crashes with claude --continue / --resume, v2.1.229)
• Fixed auto mode failing on all tool calls: An issue where auto mode failed on all tool calls when the attribution header was disabled with CLAUDE_CODE_ATTRIBUTION_HEADER (v2.1.229)
• Fixed crash with special Windows paths: An issue where tool calls or messages referencing files with extended-length paths (\\?\) or UNC paths would crash (v2.1.229)
• Fixed git / Git Bash detection failure on Windows: An issue where launching Claude Code from the parent folder of git's installation directory would fail to find git / Git Bash (v2.1.228)
• Fixed MCP OAuth authorization failure: Changed redirect URI to use 127.0.0.1 instead of localhost, enabling authentication with strict authorization servers (v2.1.229)
• Fixed Claude Code Review workflow not posting: An issue where the workflow generated by /install-github-app would complete without posting a review to the pull request (v2.1.229)
• Fixed compaction retry on exceeding 32MB: An issue where conversations where messages alone exceeded the API's 32MB request limit would keep retrying even when there were no images or documents to strip. Now fails once with a clear message (v2.1.229)
• A quietly welcome fix: Fixed an issue where the UI would freeze for several seconds after editing a file with thousands of IDE diagnostics while an IDE extension was connected (v2.1.229). I feel the perceptible difference will be greater for projects dealing with large files with many diagnostics.
• Additionally, numerous minor bugs have been fixed including self-hosted runner startup and shutdown processing, plugin cache, file watching, OpenTelemetry export, and SDK input validation.

Breaking Changes and Deprecations

--base-dir is now required for self-hosted runner startup on Windows (v2.1.229)

Explicit specification of --base-dir is now required when starting the self-hosted runner on Windows. There is no default checkout directory on Windows.

Below is a constructed example showing the impact of the change.

Before the change (up to v2.1.228):

# Windows (up to v2.1.228): Could start without specifying --base-dir
claude self-hosted-runner

After the change (v2.1.229 and later):

# Windows (v2.1.229 and later): --base-dir specification is required
claude self-hosted-runner --base-dir C:\runner\workspace

Dangerous flags no longer auto-approved in /commit-push-pr (v2.1.229)

/commit-push-pr has been changed so that git and gh commands containing dangerous flags such as --force / --amend / --no-verify are no longer auto-approved.

Below is a constructed example showing the difference in behavior.

Before the change (up to v2.1.228):

# Commands containing dangerous flags were also subject to auto-approval
/commit-push-pr
  → git push --force ... would execute as-is

After the change (v2.1.229 and later):

# Commands containing --force / --amend / --no-verify etc. are not auto-approved
/commit-push-pr
  → git push --force ... will prompt for approval

Closing Thoughts

I feel these 2 versions have their center of gravity on foundational stability rather than new features. The fact that commonly encountered bugs like streaming display corruption and crashes in narrow terminals have been resolved together is significant. On the other hand, the revision of /commit-push-pr auto-approval and the mandatory --base-dir can affect automation flows, so I feel it's reassuring to check these beforehand.

Why not update and give it a try? If there are changes that interest you, start by checking them out in your local environment.

References

https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

https://code.claude.com/docs/en/changelog

https://dev.classmethod.jp/articles/20260811-cc-updates-v2-1-227/