Claude Code 2.1.91: MCP Persistence, Shell Execution Control

TL;DR

• MCP tool results can now persist up to 500K chars via annotation override, fixing truncation on large schemas
• New setting to disable inline shell execution in skills and plugins
• Fixed transcript chain breaks on resume that could lose conversation history
• Plan mode in remote sessions no longer loses track after container restart

New

• MCP result persistence override — Annotate with _meta["anthropic/maxResultSizeChars"] to pass results up to 500K without truncation, critical for large DB schemas and similar payloads.
• Disable shell execution setting — New disableSkillShellExecution option prevents inline bash execution in skills, custom slash commands, and plugin commands.
• Multi-line deep link prompts — claude-cli://open?q= now accepts encoded newlines (%0A) for multi-line queries.
• Plugin executables — Plugins can ship binaries under bin/ and invoke them as bare commands from the Bash tool.

Fixed

• Transcript chain breaks on --resume that silently lost conversation history when async writes failed.
• cmd+delete not deleting to start of line on iTerm2, kitty, WezTerm, Ghostty, and Windows Terminal.
• Plan mode in remote sessions losing track of plan file after container restart, causing permission prompts and empty approval modals.
• JSON schema validation for permissions.defaultMode: "auto" in settings.json.
• Windows version cleanup not protecting the active version's rollback copy.
• /feedback now explains unavailability instead of vanishing from slash menu.

Improved

• /claude-api skill guidance expanded for agent design patterns, tool surface decisions, context management, and caching strategy.
• Performance: faster stripAnsi on Bun via Bun.stripANSI routing.
• Edit tool now uses shorter old_string anchors, reducing output tokens.

Update via: brew upgrade claude-code or download from code.claude.com

Source: Claude Code