Why it mattersA review level per repository lets a team run a cheap pass on a sandbox and a careful one on a core service, with the choice in admin settings.
A pull request review careful enough for the payroll service is too slow for the sandbox a developer uses to try out a merge conflict. Claude Code 2.1.304 lets an admin set the two differently.
Anthropic released the version to subscribers and lists three new Code Review levels: Quick, Standard (the default), and Deep. An owner sets a level for the whole GitHub organisation and may override it per repository in the admin settings. The same page carries a new option that enables review on every repository the organisation holds, and on any repository added later, so a team no longer has to pick them one by one. Pull requests opened in the first minute or two after an admin turns Code Review on, which used to be silently skipped, are now retried and reviewed.
What else moved
The mod API answer for a tool permission check gets a new value: decider: "person" means only the permission dialog or a host callback can approve the action, and when nobody can be asked, the ask is refused with the mod's own reason. A matching event field e.isReadOnly on tool.check tells a mod whether the call it is screening will only read.
Managed Agents gets a read-only audit at /claude-api managed-agents-secure, which compares an agent's settings against Anthropic's published security guide without changing anything. Mod settings.read hooks now receive each settings layer and rewrite the merged settings, which lets a managed mod set an organisation ceiling on options such as the cleanupPeriodDays retention sweep.
The fixes a working team will see
Two fixes land on the stalls teams have reported. Model switches, messages and interrupts no longer wait behind a context usage check in SDK and desktop sessions, so a session that is close to the limit can still change models or stop a turn. Sandboxed commands on Linux no longer error when a sandbox.filesystem.denyRead glob contains brackets that are not a character class, which was hitting any team with a [user-id] route folder.
Three safety fixes close holes a developer would only notice once. rm -r of a mounted drive, or of a folder at the top level of /Volumes, /media or WSL's /mnt/c, now triggers the same confirmation prompt that rm -rf /Applications already triggered. claude --worktree no longer removes the worktree and its branch without asking at exit when HEAD was detached or on another branch, a shape that could lose commits. Mod sharing in the built-in plugin-authoring skill now gets past the secret scan on machines with managed settings and on Team and Enterprise plans.
Transcripts, hooks and skills all tighten. A skill's once hook no longer removes another skill's identical hook after it runs. Skills from one source (a repository, your own folder, or a plugin) are now capped at 1,024 hooks and 16,384 matcher characters per session. CLAUDE_CODE_AUTO_COMPACT_WINDOW=200k used to silently set a 100K window because the suffix was dropped; the variable now accepts k and M the same way --autocompact does, and warns on out-of-range values.
Code Review's Deep level is the one setting worth setting carefully. A review that takes more time on every pull request slows the whole team if the admin turns it on everywhere, so the point of the per-repository override is to run Deep where the extra time pays back and Quick where it does not.
Source
Release v2.1.304, Anthropic.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
