Anthropic Dispatch: Claude Code 2.1.147 and the Rise of Pinned Background Sessions

Background Agents That Actually Stay Alive

May twenty-first, twenty twenty-six. Claude Code two point one point one forty-seven shipped what might be the most important infrastructure change for anyone running background agents. Pinned background sessions, the ones you create with control-T in the claude agents view, now stay alive when idle. They get restarted in place to apply Claude Code updates. And they only get shed under memory pressure after all non-pinned sessions have been shed first.

This is a fundamental shift in how background sessions work. Before this, your background agents would die when they went idle. You would come back to check on a long-running task and find the session gone. The agent had finished thinking but the session evaporated before you could read the results. Now, if you pin a session, it stays. It survives idle periods. It survives Claude Code updates. It only disappears if your machine is genuinely running out of memory, and even then it is the last to go.

For anyone building workflows around background agents, this is the release that makes them reliable enough to depend on. You can kick off a code review agent, a test runner, a documentation updater, pin all three, and come back hours later knowing they will still be there with their results.

Slash Simplify Dies, Slash Code-Review Is Born

The second headline is a rename that is actually a reimagining. Slash simplify is now slash code-review. But it is not just a name change. The command now reports correctness bugs at a chosen effort level. You can run slash code-review high for a thorough analysis. And you can pass dash dash comment to post the findings as inline GitHub PR comments directly.

The old cleanup-and-fix behavior has been removed. This is Anthropic making a choice about what the command should be. It is not a code simplifier anymore. It is a code reviewer. The distinction matters. Simplification is about aesthetics and conciseness. Review is about correctness, bugs, and structural issues. By renaming and refocusing the command, they are signaling that this tool is for catching problems, not for making code prettier.

The effort level flag is particularly interesting. Slash code-review low gives you a quick scan. Slash code-review high gives you the deep analysis. You pick how much of your context budget you want to spend on the review. For a quick check before pushing, low is fine. For a pre-merge review on a critical change, high is worth the tokens.

The Auto-Updater Gets Serious

The auto-updater got a meaningful improvement. It now retries transient network failures instead of just failing. It reports specific error categories and OS error codes on failure. And it shows you the current version when an update fails, so you know where you stand.

This sounds like a minor change until you remember that Claude Code updates almost daily. At the pace they are shipping, a flaky updater means you fall behind quickly. If your update failed silently on a coffee shop wifi connection, you might not notice for days. Now it retries, and when it truly fails, it tells you why with enough detail to actually diagnose the problem.

The Permission Bypass That Should Worry You

Buried in the bug fixes is a security fix worth calling out. Enterprise login restrictions, the forceLoginOrgUUID and forceLoginMethod managed settings, were not being enforced against third-party-provider and API-key sessions. If your organization was using these settings to ensure people logged in through your identity provider, someone could bypass that by using an API key directly.

This is the kind of security bug that only matters in enterprise environments, but in those environments it matters a lot. If you are an admin who set up forced login through your SSO provider and assumed that meant everyone was going through SSO, they were not. Fix it by updating.

The Paste Bug

One bug fix deserves special mention for how absurd it is. Pasted text was being delivered to background agents as an unreadable placeholder that said "pasted text number N" instead of the actual content. If you pasted code into a background agent session, the agent received a useless reference string instead of your code. It had no idea what you pasted. It was working blind.

Imagine debugging a background agent that keeps getting things wrong, not realizing that every time you pasted context into it, the context was replaced with a meaningless token. That bug must have caused hours of confused troubleshooting for people who never thought to check whether their paste actually arrived.

The Diff Rendering Fix

Large file edit diffs now render faster. This is a performance fix, not a feature, but if you have ever watched Claude Code freeze for several seconds while rendering a diff on a big file change, you know why it matters. The rendering was recalculating layout on every line. Now it does not. Your terminal stays responsive during large edits.

Putting It Together

Two point one point one forty-seven is the background agents release. Pinned sessions that survive. A code review command that posts to GitHub. An updater that actually works on bad networks. These are not flashy features. They are the infrastructure that makes the flashy features reliable. If you are running Claude Code as a daily tool with multiple agents, background sessions, and automated reviews, this is the release where those workflows stopped being fragile experiments and started being things you can actually count on.