Claude Code Now Runs Code Reviews in the Background and Gets Opus 5

Claude Code shipped two releases back-to-back — v2.1.218 on July 22 and v2.1.219 on July 24 — that change how you use the tool day to day. The headline features: code reviews no longer fill your conversation transcript, and Claude Opus 5 is now available with a 1-million-token context window.

If you use Claude Code for anything more than quick edits, these changes matter. Here is what changed, why it matters, and how to use it.

Code reviews run in the background now

Before v2.1.218, running /code-review in Claude Code would stream the review directly into your conversation. That sounds fine until you have asked Claude to review a pull request with 30 changed files. The review text fills your context window, pushes out earlier conversation, and makes the rest of your session harder to work with.

Version 2.1.218 changes this: /code-review now runs as a background subagent. When you type the command, Claude Code spins up a separate worker that handles the review while your main conversation stays clean. You can keep chatting, asking questions, or working on other tasks while the review runs.

The review still uses whatever context you have set up — your stacked slash commands, your recent conversation, the files Claude Code already knows about. The difference is that the review output goes into the background agent’s transcript, not yours. When the review finishes, you see a summary and can open the full results.

For /code-review ultra — the version that runs a cloud-side review instead of a local one — the fix also matters. Previously, running /code-review ultra in a non-interactive session (like a CI pipeline) would silently fall back to a local review. Now it correctly launches the cloud review even in headless mode.

This is a workflow change that experienced Claude Code users will feel immediately. Background reviews mean you can fire off a review request and keep working without your context getting cluttered.

Claude Opus 5: 1M context, fast mode, new default

Version 2.1.219 adds Claude Opus 5 (claude-opus-5) as the new default Opus model. This is Anthropic’s most capable model, and in Claude Code it ships with a 1-million-token context window by default.

What does 1M context mean for you? In practical terms, you can feed Claude Code an entire large codebase — thousands of files, tens of thousands of lines — without hitting the context limit. Earlier models would start compacting (summarizing and dropping) parts of your conversation as context filled up. With 1M tokens, compaction is rare even for the largest refactoring tasks.

The model runs at $10 per million input tokens and $50 per million output tokens in fast mode. The /model picker now shows “Opus (1M context)” so you can confirm you are on the right variant.

If you were already using Opus in Claude Code, the upgrade is automatic. The model row in the picker has been updated to reflect the new default. If you are picking a model for the first time, Opus 5 with 1M context is now the Opus option.

Strict sandbox networking for locked-down environments

A new sandbox.network.strictAllowlist setting lets you deny connections to non-allowlisted hosts without prompting. In earlier versions, Claude Code’s sandbox would ask for permission when a command tried to reach a host not on the allowlist. In strict mode, it simply blocks the connection.

This matters for two audiences. First, enterprise teams that need to enforce network policies — strict allowlist mode means the agent cannot accidentally phone home to an unknown endpoint. Second, CI/CD pipelines where interactive prompts are not possible and you need deterministic behavior.

When sandbox.network.strictAllowlist is enabled, any network request to a host not explicitly allowed in your configuration is denied immediately. No prompt, no approval button. This is the setting you want when compliance requires that no unvetted traffic leaves your machine.

Directory-added hook for mid-session workspace changes

Claude Code has always supported hooks — scripts that run automatically when certain events happen. Version 2.1.219 adds a new hook: DirectoryAdded. This fires whenever you add a working directory to your session, either through the /add-dir command or through the SDK’s register_repo_root control request.

Why is this useful? If you have setup scripts that need to run when a new directory enters the session — for example, generating type definitions, installing dependencies, or running linters — the DirectoryAdded hook gives you a clean trigger point. Before this, you would have to manually run setup after each /add-dir, or use a more generic hook that fired on every session event.

Dozens of fixes that matter

The two releases together include fixes for issues that have been frustrating Claude Code users for weeks. Here are the ones most likely to affect your daily work:

Windows path corruption fixed. Windows paths with segments starting with \u (like C:\Users\unicorn) were being corrupted into CJK characters in tool inputs. If you use Claude Code on Windows and have a username or path segment that starts with \u, this was making files inaccessible. Now fixed.

MCP error visibility. When an MCP server fails to connect, Claude Code now shows the HTTP status code and error text in the output of claude mcp list and the /mcp slash command. Before, you would just see that a server was not connected, with no indication of why. Additionally, MCP config values with hidden leading or trailing whitespace now trigger a warning — a common source of silent failures.

Background session isolation. Sessions sent to the background (via /background or the left arrow key) could escape their workspace folder if the working directory was a symlink. The fix canonicalizes symlinked directories so background sessions stay contained.

Long session performance. Sessions with many turns would develop multi-second stalls as message normalization cost grew quadratically. This is fixed, so long-running sessions should feel responsive again.

Stale auto-compact fix. /compact would fail once over the context limit on Claude Opus 4.8 on Bedrock, and auto-compact would never trigger for that model. Both are fixed.

Permission prompt on restart. If you approved a permission while a self-hosted runner was restarting, the approval would be silently dropped. The approved action now runs after the session resumes.

What this means for Claude Code users

These two releases represent a maturation of Claude Code’s workflow. Background code reviews keep your context clean. Opus 5 with 1M context removes the practical ceiling on how much code Claude can reason about at once. The strict sandbox setting makes Claude Code viable in environments where network control is mandatory.

If you have been waiting for Claude Code to handle large-scale code review without polluting your session, this is the release that delivers it. And if you have been hitting context limits on big refactoring tasks, 1M tokens is the answer.

Update with:

# If installed via npm
npm update -g @anthropic-ai/claude-code

# If installed via Homebrew
brew upgrade claude

The latest stable is v2.1.220, which includes the fixes from both v2.1.218 and v2.1.219 plus additional reliability improvements.