Claude Code ships /doctor and safety rules
Claude Code v2.1.205 adds /doctor for setup diagnostics, transcript tampering protection, a ~400 MB auto-update memory cut, and a Windows NTFS junction fix.
Claude Code v2.1.205, tagged July 8, ships a cluster of improvements to auto mode safety, agent view readability, PR linkage, memory use, and Windows reliability.
/doctor is now a real diagnostic
/doctor is a full setup checkup that diagnoses and fixes configuration problems. /checkup is its alias. The command audits your Claude Code install, catches broken setups, and confirms every proposed fix before applying it.
Two new auto mode safety rules
Auto mode now blocks any tool call that would modify a session transcript file. The target is a class of prompt-injection attacks that tried to forge approved actions inside the session log. Separately, auto mode now stops and asks before executing rm -rf on a shell variable it cannot resolve from context, so an unset or mistyped variable does not silently delete a path.
Agent view and PR linking
Session rows in claude agents now show a colored state label and a classifier-written headline instead of raw tool call text. Sessions that edit, merge, comment on, or push to an existing GitHub PR now link that PR in the view automatically.
Auto-update memory cut
Binary downloads for auto-updates now stream to disk rather than buffering in memory, cutting peak memory use by roughly 400 MB.
Windows fix
A bug that deleted files outside the worktree during worktree removal when an NTFS junction or directory symlink was present inside it has been fixed.
For how the permission model that governs auto mode works, see Claude Code makes manual mode the default. Teams running agents in cloud sandboxes should read background agents and the cloud execution model for how these new rules interact with remote environments.