Claude Code v2.1.205: Session Security, Bug Fixes, and 400 MB RAM Savings

Claude Code v2.1.205 was released on July 8, 2026 at 21:22 UTC with a series of bug fixes, security improvements, and one optimization that will not go unnoticed by those working in memory-constrained environments.

New Security Rule: Session Transcript Protection

The most notable change in this release is a new rule in auto-mode that blocks any attempt to manipulate the transcript files of the active session. Transcripts record the full course of a conversation and commands — protecting these files prevents scenarios in which malicious content or a compromised external dependency might attempt to retroactively alter the record of previous actions or inject false instructions into the session context.

What About Stuck Background Agents?

Background agents no longer get stuck in “failed” or “completed” states after a session resumes. This was a particularly frustrating problem in long-running development workflows where background agents appeared “dead” but the system would not release them. In addition, background task notifications now clearly indicate that there was no human input — giving users a more accurate picture of what the agent was actually doing while unmonitored.

Fixed Bugs That Affected Day-to-Day Work

Several fixes address scenarios that could occur without any visible warning and therefore go unnoticed for a long time.

JSON schemas with the format keyword had been silently returning unstructured output instead of properly applying the specified schema. The tool would appear to work, but was actually bypassing the structure — with no warning or error. Version 2.1.205 now correctly accepts and fully processes schemas with the format keyword.

Messages sent to Claude while it was actively working were lost if the agent reached its --max-turns limit at that moment. Messages are now properly preserved and processed even when the agent completes its turn.

Windows users receive a fix for specific NTFS junction point behavior: removing a worktree no longer deletes files outside the target directory when junctions pointing to the same paths are present. This could cause accidental file loss in more complex development setups on Windows.

Auto-Update Optimization: ~400 MB Less Peak RAM

Until now, the automatic update mechanism downloaded the entire binary into working memory before writing it to disk. The new implementation streams the binary directly to disk as soon as the download begins, reducing the updater’s peak RAM consumption by approximately 400 MB. On systems with limited memory resources or in environments where Claude Code runs alongside other tools, this is a concrete and measurable benefit.

Improvements to Agent Display and the /doctor Command

The agent list display received color-coded status labels alongside auto-generated titles — each agent in the list now carries a short descriptive text that makes it easier to navigate in environments with multiple parallel agents.

The /doctor command has been expanded into a full-featured diagnostic tool for checking environment settings. Instead of just a handful of basic checks, it now provides a comprehensive overview of configuration, MCP servers, authentication, and other components that affect the tool’s operation.

In addition, a header truncation issue in the agent display was fixed, a crash on startup when the working directory was deleted or unmounted on Windows was resolved, and a problem with invalid LSP plugin server initialization that was blocking other valid servers from starting was corrected.

Summary

Version 2.1.205 is a technical rather than cosmetic update: it patches silent bugs that affected tool reliability in production workflows, strengthens session transcript security, and reduces the memory footprint during updates. Users working on Windows with complex worktree setups, those using JSON schemas with the format keyword, or those relying on background agents in long-running sessions will see direct benefit from this release.