Claude Code v2.1.153 Major Updates

This is Ishikawa from the Cloud Business Division. Claude Code v2.1.153 (released 2026-05-27) has been released. This update is centered on stabilization around background sessions and claude agents (background agents), and includes security fixes for MCP and authentication, as well as a behavior change for /model.

Update Summary

v2.1.153 lists 36 changes in the CHANGELOG. New features are relatively few, with the majority being bug fixes around background sessions and claude agents (background agents). It also includes security fixes for MCP server and API gateway authentication, as well as a behavior change regarding the default saving behavior of /model.

Notable Updates

• New Features
  • A skipLfs option has been added to the github/git plugin marketplace sources, allowing Git LFS downloads to be skipped.
  • COLUMNS and LINES environment variables are now passed to status line commands, enabling output adjustment to match terminal width.
• Bug Fixes
  • Fixed an issue where memory usage could reach several GB when resuming sessions on machines with a large number of saved sessions.
  • Fixed an issue where the Agent tool with subagent_type: 'claude' operated in a temporary worktree, causing output to gitignored paths to be silently discarded.
• Security
  • Fixed a regression where the user's Anthropic OAuth credentials were incorrectly sent to custom API gateways.
  • Fixed an issue where MCP servers defined in subagent frontmatter were ignoring --strict-mcp-config and enterprise management policies.

Update Details

New Features

• A skipLfs option has been added to the github/git plugin marketplace sources. Git LFS downloads can be skipped during clone and update operations.
• COLUMNS and LINES environment variables are now passed to status line commands. Scripts can adjust output size to match terminal width.

Improvements

• When npm global installation cannot be automatically updated, a one-time notification is now displayed. Remediation steps are listed in /doctor.
• Autocomplete for claude agents dispatch input now suggests native slash commands and bundled skills in addition to project skills.
• The PR column in claude agents now displays PR #N for a single PR and N PRs for multiple PRs.
• claude doctor now displays the results of the most recent update attempt.
• The startup notifications for "authentication required" that were previously displayed separately for MCP servers and connectors have been consolidated into a single message.
• On macOS, background agents now appear as "Claude Code" in "Privacy & Security," and previously granted permissions are retained after upgrades.

Security

• Fixed a regression where custom API gateways were receiving the user's Anthropic OAuth credentials instead of the gateway's own token.
• Fixed an issue where MCP servers defined in subagent (Agent tool) frontmatter were ignoring --strict-mcp-config, --bare, remote mode, enterprise managed MCP configuration, and managed-settings MCP server allow/deny policies. This affects enterprise environments using managed policies.
• --strict-mcp-config no longer removes inline mcpServers from explicitly passed agent definitions (--agents / SDK's agents). Additionally, blocked subagent MCP servers are now displayed as warnings.

Fixes

• Fixed stateful MCP server reconnection loop: Fixed an issue (a regression in v2.1.147) where stateful MCP servers without an optional GET SSE stream would enter a reconnection loop on tools/list.
• Fixed excessive memory usage on session resume: Fixed an issue where memory usage could reach several GB when resuming with a transcript file path on machines with a large number of saved sessions.
• Fixed subagent output being discarded: Fixed an issue where the Agent tool with subagent_type: 'claude' was running inside an undocumented temporary worktree, causing output to gitignored paths to be silently discarded. If you have subagents generating files to gitignored paths, you may have been affected, so we recommend checking.
• Fixed claude update ignoring release channel: Fixed an issue where claude update on npm installs would install the latest version instead of the version for the configured release channel.
• Fixed false success display in Windows PowerShell installer: Fixed an issue where "Installation complete!" was displayed even when installation had actually failed.
• Fixed exit hang in stream-json mode: Fixed an issue where the CLI would fail to exit and hang when stdin was closed without EOF, leaving stale session markers.
• Fixed response being lost when using /bg during a response: When /bg is executed while Claude is responding, the response is now continued in the background session instead of being discarded.
• Improved rollback on Windows update failure: When a Windows update fails, the original executable is now restored from a copy, and recovery instructions are provided.
• Fixed process lingering on VS Code exit on Windows: Fixed an issue where Claude Code processes would not cleanly exit when VS Code was closed, causing false "unclean exit" reports and leftover MCP servers.
• Additionally, numerous minor bug fixes have been made, including file:// link clicks, claude --help wrapping, MCP tool progress display, background session operations such as /btw, /copy, and /rename, Windows IME candidate window positioning, and terminal redrawing and background colors.

Breaking Changes / Behavior Changes

The behavior of /model has changed. If you are using the default keybindings, the new behavior will be applied automatically, but if you have customized your keybindings, manual action is required.

• Model selection via /model is now saved as the default for new sessions (to align with IDE behavior). To switch the model only for the current session, press s in the picker.
• Accordingly, if you have customized the keybinding for modelPicker:setAsDefault, you will need to rename it to modelPicker:thisSessionOnly in keybindings.json (the d action has been replaced with s).

Closing

v2.1.153 is a release centered on stabilization of background sessions and claude agents, as well as security fixes for MCP and authentication. Especially in enterprise environments using managed MCP policies, the fix to MCP server behavior via subagents is important. Since fixes related to credential handling are also included, we recommend updating promptly for configurations using custom API gateways or subagent MCP. Also, since the behavior of /model has changed, users who have customized their keybindings should check what action is required.

Recent Updates

https://dev.classmethod.jp/articles/20260524-claude-code-updates-v2-1-150/

https://dev.classmethod.jp/articles/20260524-claude-code-updates-v2-1-152/

References

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

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