Claude Code v2.1.136 ships fixes for MCP server configuration, extended thinking, and terminal integration

Anthropic released version 2.1.136 of Claude Code on May 8, addressing a series of bugs that affected authentication, tool execution, and terminal integration. The release notes indicate that the team prioritized stability across multiple development environments rather than introducing new capabilities.

A significant fix targets OAuth token management in multi-server Model Context Protocol setups. Previously, concurrent refresh operations would cause some servers' tokens to be lost, forcing users to re-authenticate daily. The updated version now preserves tokens during parallel refresh cycles, reducing authentication friction for teams deploying multiple remote MCP servers.

Extended thinking integration received attention after an earlier issue where the feature would emit API 400 errors when redacted thinking blocks appeared immediately after tool calls. This fix ensures that extended thinking reasoning blocks don't interfere with downstream tool execution.

The update also resolves a login loop vulnerability where concurrent credential writes could overwrite freshly-rotated OAuth tokens, forcing unintended re-authentication cycles. A related fix ensures that the --resume and --continue CLI flags now correctly locate sessions when project paths contain underscores, a common pattern in many codebases.

UI consistency improvements touched slash command dialogs, file picker behavior in both small non-git directories and larger 100+ entry directories, and keyboard shortcut reflection from VS Code keybindings.json files. Terminal rendering fixes addressed bash output colors, markdown code block formatting, and image paste behavior in WSL2 via PowerShell fallback when xclip/wl-paste are unavailable.