Claude Code v2.1.304 adds guided auto-mode setup, managed agent security checks, and transcript improvements

Anthropic released Claude Code v2.1.304, a tool-focused update that adds a guided setup for auto mode and a new security command for managed agents. The /auto-mode-setup command reads a project’s files and recent sessions to teach auto mode about the environment, and is now available across multiple deployment surfaces including Bedrock, Vertex, Foundry, and the Claude apps gateway, as well as telemetry-disabled sessions.

The update also introduces /claude-api managed-agents-secure, which compares a managed agent’s configuration against Anthropic’s published security guide without altering any settings. This provides a non-intrusive way to audit compliance with documented best practices for managed agents.

Transcript handling and session management receive several improvements. The compaction summary in the transcript view now includes the names of tools called earlier, and fixes address transcript message loss when disks fill during trimming. Linux sandbox path handling is corrected for glob patterns that are not character classes, such as routes with brackets like [user-id].

Session reliability is improved with fixes for Remote Control history loss when switching accounts, context limit errors after stopping turns during compaction, and subagent resume issues where recent work could be forgotten. Terminal display issues are addressed, including truncation of notes in multi-choice prompts in narrow terminals and correct path display for files outside the working directory.

Additional fixes target mod management, effort indicators, and command-line behavior. A failed mod reload no longer removes tools and commands from the previously loaded version, and effort status commands now reflect actual settings rather than guessed values. Command-line help text is updated to match actual behavior, and worktree handling is corrected to prevent data loss when HEAD is detached.