Claude Code v2.1.204 to v2.1.205 Major Updates

This is Ishikawa from the Cloud Business Division. I'd like to share a summary of the Claude Code updates for v2.1.204 through v2.1.205 (both released on 2026-07-08).

Update Summary

This covers 2 versions (v2.1.204 through v2.1.205, 2026-07-08). While the majority of changes are bug fixes, the release also includes usability improvements such as the transformation of /doctor into a full setup inspection tool and a refreshed agent view display, security and safety enhancements around auto mode, and a fix for a data loss risk related to worktree deletion on Windows.

Highlights

• New Features & Improvements: /doctor has become a full setup inspection tool that not only diagnoses issues but also fixes them (/checkup is an alias). Additionally, the row display in the agent view (claude agents) has been refreshed (v2.1.205)
• Bug Fix: Fixed an issue on Windows where deleting a worktree that contained NTFS junctions or directory symbolic links would also delete files outside the worktree (v2.1.205)
• Security: Auto mode rules have been added to block tampering with session transcript (conversation history) files, and fabricated approvals within transcripts are now prevented from being executed (v2.1.205)

Target Versions and Period

New Features & Improvements

• /doctor becomes a full setup inspection tool: /doctor is now a setup inspection command that can both diagnose and fix issues. /checkup is its alias (v2.1.205)
• Refreshed agent view row display: Each row in claude agents now displays colored status words and classifier-generated headings instead of raw tool call text. For blocked sessions, opening the peek (preview) shows the full status including the exact request details (v2.1.205)
• Agent view links existing PRs: Sessions that have edited, merged, commented on, or pushed to an existing PR will now link that PR in claude agents (v2.1.205)

Security & Safety

Safety enhancements related to auto mode and background execution have been introduced.

• Blocking transcript tampering: Auto mode rules have been added to block tampering with session transcript (conversation history) files (v2.1.205)
• Preventing fabricated approvals from executing: Background task notifications now explicitly state that "no human input has been provided," preventing fabricated approvals within transcripts from being executed (v2.1.205)
• Confirmation before rm -rf on unresolvable variables: Auto mode now prompts for confirmation before executing rm -rf on variables that cannot be resolved from context (v2.1.205)

Performance

• Reduced memory usage during auto-updates: The auto-update binary download has been changed from buffering in memory to streaming to disk, reducing peak memory usage during updates by approximately 400MB. This is an improvement during the update process, not during normal usage (v2.1.205)

Fixes (Major)

A selection of fixes related to stability and usability:

• Fixed issue where external files were deleted when removing a worktree on Windows: Files outside the worktree were being deleted when the worktree contained NTFS junctions or directory symbolic links (v2.1.205)
• Fixed issue where messages sent upon reaching --max-turns were silently lost: Messages sent while Claude was running were being silently dropped when a turn ended at the --max-turns limit (v2.1.205)
• Fixed issue where --json-schema silently produced unstructured output when invalid: Fixed an issue where specifying an invalid schema would silently produce unstructured output, and an issue where schemas using the format keyword were rejected (v2.1.205)
• Fixed two types of crashes on Windows: Fixed a crash when the launch directory was deleted, locked, or unmounted during command execution, and a crash when the file watcher was closed while a directory scan was in progress (v2.1.205)
• Fixed poor streaming in SessionStart hooks: Events were not being streamed in the SessionStart hook for headless sessions, potentially causing remote workers to be idle-reaped during hook execution (v2.1.204)
• Fixed login failure (regression) in Cowork VM mode: Local agent sessions in VM mode were failing to start with "Not logged in · Please run /login" since CLI 2.1.203 (v2.1.205)
• Fixed issues with claude attach and claude mcp add-from-claude-desktop: Fixed an issue where claude attach would error during a background agent upgrade restart (changed to wait for recovery), and an issue where claude mcp add-from-claude-desktop would stall when a server name contained unsupported characters (now reports invalid names and continues importing the rest) (v2.1.205)
• In addition, numerous minor bugs have been fixed in areas including agent view and agent list status display, Web/mobile Remote Control panel status display, session and PR association, plugin LSP server initialization, and project verify skill regeneration (v2.1.205)

Breaking Changes & Deprecations

• Reservation of MCP server names "Claude Browser" and "Claude Preview": In anticipation of a Claude Desktop pane rename, the MCP server name "Claude Browser" has been reserved (along with "Claude Preview"). As a result, user-configured MCP servers can no longer be registered with these names. If you have an MCP server registered under one of these names, you may need to rename it (v2.1.205)

Closing

v2.1.204 through v2.1.205 are primarily bug fix releases, but they include changes that shouldn't be overlooked, such as the fix for the data loss risk associated with worktree deletion on Windows and the safety enhancements around auto mode. Additionally, the transformation of /doctor into a full setup inspection tool and the refreshed agent view display are improvements that will make a difference in day-to-day operations. Those affected by the breaking change regarding reserved MCP server names should take note, and if any of the changes catch your attention, please update and check them out.

References

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

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