Claude Code 2.1.218 moves /code-review to a background subagent
A 37-change drop: auto mode clears its own dangerous-command prompts, Hugging Face's MCP server gains code sandboxes, and a data-loss key gets fixed.
Copy markdown
/code-review now runs in the background
Claude Code 2.1.218 turns /code-review into a background subagent, so review output no longer floods your main conversation and stacked slash-commands can target it. Reviewing and coding finally sit in separate lanes.
Auto mode stops interrupting for risky commands
The dangerous-rm, background-& and suspicious-Windows-path checks no longer pop permission dialogs; the auto-mode classifier adjudicates them instead, so unattended runs pause far less. `claude mcp list` and `/mcp` now also print HTTP status and error text when a server won't connect.
A silent data-loss key and Windows Unicode paths, fixed
The left-arrow key can no longer discard your entire conversation with no undo, and Windows paths with \u-prefixed segments (like C:\Users\unicorn) no longer corrupt into CJK characters in tool inputs.
Hugging Face MCP server gets code sandboxes
HF's MCP server adds Sandboxes — secure execution environments attached to repos and buckets — plus a unified hf_fs tool that folds repo, storage, docs and paper search into roughly 1,000 tokens. An agent can now run code against a dataset without leaving the hub.
Claude's Managed Agents gain webhooks and session seeding
The Claude Developer Platform now fires webhooks on environment and memory-store lifecycle events and lets you seed a session with up to 50 user.message / user.define_outcome events, with an optional version field for optimistic concurrency on agent updates.