No More "Let Me Explain That Again"—Claude Code v2.1.232 Forks Subagents With the Whole Conversation and Lets You @-Mention Sessions Already Running

Claude Code v2.1.232, released by Anthropic on August 13, quietly changes how you work when several agents run at once. Both how you hand a task off to a subagent and how you call out to another session that's already running have been reworked as default behavior. It's not a flashy new model, but it's a change that hits hardest for people who spin up agents every day.

No more rewriting the background on every handoff

Until now, whenever you handed a task to a subagent, you had to rewrite which files you'd touched and the decisions you'd made so far, every single time. Because the child agent had no knowledge of the parent's conversation, a person had to prepare a "handoff document" of the context. v2.1.232 changes that assumption.

Three things are different:

• fork is enabled by default: A subagent with subagent_type: "fork" inherits the parent's conversation history and prompt cache as-is. There's no need to re-enter the context.
• Background execution by default: Non-teammate agents launched from an interactive session now run in the background by default. Their tool output doesn't flow into your local context, so it's less likely to clutter the main conversation.
• You can revert it: When this behavior gets in the way, export CLAUDE_CODE_FORK_SUBAGENT=0 returns things to how they were.

In short, you no longer have to re-narrate "how far along we are" to your stand-in. Because it inherits the entire conversation and cache, you can get straight to the point the moment you hand off.

Calling another running session by name

The other pillar is a mechanism for talking directly to another session that's running independently. Type @ in a prompt and you can name another Claude session by its name, and Claude reaches out to that counterpart via SendMessage. You can now drop a word into work you're running in another window, straight from your current conversation.

The surrounding conventions have been tidied up too.

• The name alone is enough: SendMessage delivers directly, without a reference-ID check, as long as the name uniquely matches a live session.
• Preventing name collisions: Session names on the same machine are kept unique, and if you launch or rename to a name that already exists, an alias in the form name-word-word is assigned and you're notified of it.

This SendMessage is the same tool already used to reach subagents and teammates within a single session. It's not that a new tool was added—rather, the existing plumbing has been made usable across sessions.

You decide whether to accept

Since others (other sessions) can now call out to you, controlling what you receive is essential. In v2.1.232, /config gains a setting for handling messages from other sessions. You can choose from three receiving behaviors:

• accept: Receive messages from other sessions.
• hold: Pause them and check later.
• refuse: Don't receive them.

You can also configure a dialog expiry. The design assumes you'll switch to hold or refuse while you want to stay focused on your work.

As parallel work gets lighter, watch for blind spots and permissions

With context handoff now automatic and other sessions callable by name, running multiple agents in parallel gets a lot lighter. The chore of re-narrating the plan to your stand-in disappears, and you can inject into work that's still running. From small personal parallel tasks to coordinating multiple sessions with divided roles, the range of ways to structure work widens.

That said, there are things worth keeping in mind. Once background execution becomes the default, a child agent's progress no longer surfaces in your local context, which may make it harder for a person to grasp what it's doing. And if forking becomes effortless, there's a real concern that the number of agents and token consumption balloon without you noticing. It's probably wise not to leave things running unattended, and to spin up a fork to check its behavior when needed.

The breaking changes on the configuration side shouldn't be overlooked either. In releases from the same period, sandbox.ripgrep can no longer be overridden from project settings and must be moved to user settings or managed settings. Overriding binaries in a managed sandbox now requires approval, and alongside hardening of Windows permissions and the Git sandbox, some operations may revert to asking for permission confirmation again. It's safest to give your own config files a once-over before upgrading.

References: Claude Code changelog (official) / Releasebot: Claude Code Updates / DevelopersIO: Claude Code v2.1.230–v2.1.232 / anthropics/claude-code Releases