Reviews Move Out of the Conversation, Deep Dives Wait to Be Called — How Claude Code v2.1.218 Draws the Line at "No Running Off on Its Own"

On July 22, Anthropic released v2.1.218 of Claude Code. It's a bundle of small bug fixes and, at first glance, an unremarkable update. But follow the changes and a single thread emerges: pushing heavy automated work out of the conversation (into the background) and keeping Claude from kicking off heavy processes on its own—a tightening of the reins. It follows directly from the previous version, which capped the number of concurrently running subagents (a default of 20).

Reviews stop filling up the conversation

The change with the most practical impact is that /code-review now runs as a background subagent. Until now, review output flowed into the main conversation and tended to bury the exchange. From this release, reviews proceed behind the scenes, leaving your working conversation free for the task at hand. Any slash commands you queued up just beforehand are carried over as the review target.

Alongside that, gaps in unattended, non-interactive use were closed as well.

• Headless misrouting: Fixed a bug where /code-review ultra silently fell back to a local review in non-interactive sessions; it now correctly launches a cloud review.
• Descriptive arguments: Fixed a case where /ultrareview failed on prose arguments like "review my auth changes." It now reviews the current branch and treats that text as a note.

The less anyone is watching—CI or automated operation—the harder it is to notice "a lighter process running quietly, contrary to intent." These are fixes that correct that on the defaults' side.

Don't start a heavy investigation until asked

The other axis is a shift toward "don't launch on your own." /deep-research was changed to run only when invoked manually, so Claude no longer decides to spin it up on its own. In addition, skills that carry context: fork now run in the background by default; if you need the old behavior, you can write background: false in the individual skill to restore it.

Both changes move in the direction of reducing the "a heavy task was running before I noticed" situation. Because deep dives and fork runs eat a lot of tokens and time, simply not starting on their own makes cost far more predictable in unattended operation.

Hooks from untrusted folders won't run

The safe-side defaults were strengthened too. Hooks written in an agent definition (frontmatter) now execute only when the folder containing that file has earned the workspace's trust. The aim is to close the path by which a definition pulled in from an untrusted location could run arbitrary commands on its own. The auto-mode area was also cleaned up: dangerous rm, background execution with a trailing &, and suspicious Windows paths are now adjudicated by the auto-mode classifier without surfacing a permission dialog. Under plan + auto, "Bash that can't be proven read-only" is likewise ruled on by the classifier.

An update that pays off the more you run agents unattended

None of this is a flashy new feature. But it pays off the more you run Claude Code unattended on CI or a server. Reviews and investigations don't fill the conversation and won't start running unless you ask, so both the predictability of execution and the visibility of cost improve.

That said, it's worth keeping in mind that the scope of decisions delegated to the classifier is expanding. Fewer permission dialogs is convenient, but there's a flip side: "why something was allowed or denied" becomes harder for a human to see. The concern that automated judgment turns into a black box behind the convenience remains, just as with other automation. It's safest to delegate only once the operator understands where the trusted-folder line is drawn and how auto mode behaves. On the whole, this version leans on "tightening the defaults" rather than adding features—a set of measures aimed at unattended operation.

References: Claude Code official changelog / GitHub: anthropics/claude-code CHANGELOG / Releasebot: Claude Code Updates