Claude Code v2.1.187: Structured Output Fixes, Sandbox Security, and Org Model Restrictions

Anthropic released Claude Code v2.1.187 with a batch of practical fixes and additions. The headline items: a sandbox.credentials setting to block sandboxed commands from reading credential files and secret environment variables, and org-configured model restrictions now enforced across the model picker, --model flag, /model command, and ANTHROPIC_MODEL env var. Users see a "restricted by your organization's settings" message when selecting a blocked model.

Key Changes

• Structured output fix: --json-schema and workflow agent({schema}) no longer allow the model to re-call StructuredOutput indefinitely after a successful call. Follow-up turns now reliably return structured output instead of falling back to free text.
• Remote MCP timeout: Tool calls that hang for 5 minutes now abort with an error instead of blocking indefinitely. Override with CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT.
• Startup performance: Remote sessions no longer add ~2.7s overhead from the agent proxy CA system-trust install.
• Subagent depth tracking: Resumed subagents restore their original spawn depth; forked subagents now count toward the depth cap instead of possibly bypassing limits.
• Worktree cleanup: Locked .git/worktrees/ entries from killed agents are now cleaned up automatically, preventing leaked registrations.
• Mojibake fix: Pasted Korean/CJK text in terminals that deliver paste as per-byte extended-key events no longer turns into garbage.
• Agent stop notifications: Now correctly attribute who stopped the agent and use "finished"/"stopped" wording instead of "came to rest".
• /install-github-app improvement: GitHub Actions workflow setup is now optional — you can install just the GitHub App and skip the workflow/secret steps.
• /btw navigation: Use ←/→ arrows to step through earlier answers.
• /plugin improvement: Surfaces plugins you haven't used recently so you can clean them up.
• Mouse click in fullscreen: Select menus (permission prompts, /model, /config) now support mouse clicks in fullscreen mode.
• VS Code: Fixed extension becoming unresponsive when resuming a large session.
• --help fix: claude --help now lists the --bg/--background flag.
• Other fixes: --resume failing with "No conversation found" after -p runs with no turns; background jobs stuck in "working"; channel connection drops after navigating to agents view or after /bg, /tui, /update; Esc/Ctrl-C/Ctrl-D not working during /share upload; /update hanging when a startup trust dialog would show; Cmd+click not opening URLs in fullscreen mode in Ghostty on macOS.

This release is primarily for teams using structured output workflows, remote MCP servers, or running sandboxed commands that need credential isolation. The performance regression fix (2.7s startup delay) is significant for anyone using Claude Code Remote.