Claude Code 2.1.307-308: stop automatic reviews, more reliable MCP

Claude Code's 2.1.307 and 2.1.308 follow 2.1.306 closely. 2.1.307 brings a real new feature for teams running Code Review on GitHub, plus six improvements and twenty-six fixes, about ten of them touching MCP servers and agents. 2.1.308 is a single line, "bug fixes and reliability improvements," with no public detail. Nothing here breaks a habit, except on self-hosted runners whose checkout hooks run longer than 30 minutes, and several annoying MCP and cloud-session bugs are gone.

@claude review stop ends the subscription without touching a review in progress

Code Review is the managed service, available on Team and Enterprise plans, that reviews every GitHub pull request and posts its findings as inline comments. Since a July 2026 update, commenting @claude review always on a pull request subscribes it to push-triggered reviews, even in a repository set to Manual or Once after PR creation. 2.1.307 adds the mirror command: @claude review stop (also off or pause) ends that subscription without starting a new review and without canceling one already running, and Claude replies with a plain comment instead of a check run. In a repository set to After every push, the command has no effect: every push there starts a review, subscription or not, and converting the pull request to a draft is still the way to stop reviews. If you run Code Review on a repository set to Manual, this is the command worth remembering for a pull request you subscribed for a while and no longer want re-reviewed on every push.

Everything else: sturdier MCP and agent sessions

On the MCP side, Claude Code already reconnects a dropped server and retries tool discovery requests automatically after a transient error. 2.1.307 fixes a case where an HTTP MCP tool call was sent twice when the connection dropped mid-call, which could duplicate an action on the server side. Only read-only or idempotent tools are now retried that way. Two other fixes concern the connection itself: an MCP server is no longer reported as failed just because your Mac or Linux machine went to sleep while it was connecting, and on macOS, claude -p no longer shows "needs authentication" when the real problem is a locked keychain. On the agent side, fork subagents launched together now know which launch is their own, so a fork is far less likely to start another fork's work, and a malformed tool call is now retried up to twice before the turn ends in an error. A conversation cleared with /clear no longer comes back when a cloud or self-hosted runner session restarts, and some "Autocompact is thrashing" errors are gone in turns that read many PDF pages, images or other large tool results. For self-hosted runner admins, one default-behavior change worth noting: checkout hooks are now stopped after 30 minutes by default, which fails the session's setup; raise the limit with --checkout-hook-timeout-min, or set it to 0 to remove it. 2.1.307 also adds /claude-api api-version, which moves your Claude API code to API version 2026-10-21, and the claude_code.safety_event OpenTelemetry event, emitted for each tool call that a permission rule, auto mode, a safety check, a hook or the sandbox stopped, prompted or sent back. Finally, it fixes a bug that made a model switch through the SDK's setModel() fail after waiting more than 5 seconds. 2.1.308, for its part, says nothing beyond "bug fixes and reliability improvements."