Claude Code v2.1.218 transforms /code-review into a background subagent. When you launch a code review, the work runs in a separate session and doesn’t fill your main conversation with results. It remains available as a target for stacked slash commands, but no longer pollutes the thread where you’re working.

The release also fixes a string of concrete bugs: Windows paths with \u segments that were corrupted into CJK characters making files inaccessible, the left arrow key that discarded the conversation with no undo option, the /ultrareview command that failed on descriptive arguments like “review my auth changes”, and incorrect counting of MCP servers requiring authentication on startup. It adds HTTP status and error text to claude mcp list when a server won’t connect, plus a warning for hidden spaces in MCP config values.

Why it matters to you. If you use Claude Code for long sessions, code review cluttering the conversation was a real problem: you’d lose track of your current work amid dozens of analysis lines. Having it in the background means you can request a review and keep working, then return to the results when you need them. It’s a small shift in daily use but a big one in experience. The release series from v2.1.211 to v2.1.218, over ten patches in three weeks, tells the story of a tool stabilizing for production use. As we covered on July 22 with v2.1.217, the signal is clear: beta phase is over, now the edges are being polished.

To update: claude update or npm install -g @anthropic-ai/claude-code.

In detail

The most interesting architectural change in this release is /code-review moving to a background subagent. Until v2.1.215, Anthropic had made code review launch explicit: the initiative passed from the agent to you, invoking it with a command. Now that command runs outside the main conversation, in a separate session. It’s the same logic applied to /fork in v2.1.212, where delegation became an independent background session.

The pattern is consistent: Anthropic is separating long, noisy tasks from the conversation where you work, keeping context clean. In practice, if you request a code review on a branch with 40 modified files, you used to see the entire analysis scroll through the thread, line by line, until you couldn’t find your last question anymore amid the output. Now the subagent works in the background and delivers results when done. The main thread stays readable, and you can keep iterating on code while the review is running.

This separation between work session and analysis session has a concrete secondary benefit: your main conversation context doesn’t get saturated by review output. In long sessions, where token counts accumulate, having dozens of analysis lines in the middle of dialogue reduces useful space for subsequent interactions. Moving the review to background frees that window.

The Windows path fix deserves mention: segments like C:\Users\unicorn were being interpreted as Unicode escapes and transformed into CJK characters, rendering files literally inaccessible. A silent bug for Mac or Linux users, but blocking on Windows, where a path with \u can pop up anytime in user profiles or environment variables.

The MCP diagnostics improvement is useful for those connecting external tools: previously, when an MCP server wouldn’t connect, you didn’t see why. Now claude mcp list and /mcp show HTTP status and error text, plus a warning for hidden whitespace in config values, which are hard to spot manually and can fail connection with no clear message. For those integrating Claude Code with custom MCP servers, the difference between “doesn’t work” and “doesn’t work because there’s invisible space on line 42 of your config” is the difference between a lost afternoon and a two-minute fix.

Another silent but important fix: /code-review ultra in non-interactive sessions was launching a local review instead of the cloud one. Users of Claude Code in CI pipelines could get unexpected results without knowing why, and the issue went unnoticed because the command didn’t error, just delivered shallower analysis than intended.

What we don’t know: the changelog is detailed but doesn’t include performance or adoption metrics. The rapid patch series (10 releases in three weeks) signals development intensity, but also that the tool had accumulated bugs in non-trivial areas. No public data exists on real production stability at scale.