Claude Code 2.1.232: forked subagents are on by default

Claude Code 2.1.232, released on August 13, switches subagent forking on by default. A subagent of type fork inherits the full conversation and the prompt cache, and in interactive sessions agent spawns other than teammates now run in the background by default, the release notes on GitHub say. Versions before it leave two permission bypasses unfixed, one in PowerShell and one in Git Bash on Windows, and lack this build’s hardening of the Linux filesystem sandbox against a protected-path bypass.
Fork or named subagent?
A fork always runs the parent session’s model, so exploration work meant for a cheaper model such as Haiku needs a named subagent with model: haiku in its definition file, Start Debugging explains. Yet the same release removes the startup tip that suggested creating custom subagents, along with the matching nudge in the /powerup tour.
The other difference is the cache, Start Debugging notes: a fork’s system prompt and tool definitions are byte-identical to the parent’s, so its first request reads the parent’s prompt cache, while a named subagent gets no cache hits on its first call and uses the five-minute TTL even on a subscription where the main conversation gets the one-hour TTL. Cached reads bill at roughly 10% of the standard input rate. The fork’s tool calls also stay out of the main conversation, and only its final result comes back.
Messages between sessions
Typing @ in the prompt now mentions another Claude session by name, and Claude reaches it directly through SendMessage, which delivers to a bare name that exactly matches one live session without first asking to confirm with a ref. Interactive sessions on one machine also keep unique names: starting or renaming a session to a name another live session already uses gives it a name-word-word variant, and Claude Code tells you.
/config gains a “Messages from your other sessions” row, which sets whether inbound messages are accepted, held or refused, and a “Dialog expiry” row. Cross-session messaging arrived in v2.1.224 for macOS and Linux, holding messages to a session that runs with bypassed permissions for approval and delivering the rest automatically. A message sent this way cannot approve a permission request or change configuration: commands such as /compact arrive as plain text, and if acting on a message needs a permission, the receiving session still prompts the user.
Permissions, sandbox and GitLab
Nested git repositories no longer inherit trust from a parent directory, so each one now requires its own trust confirmation. On Windows, Git Bash followed Cygwin-style symlinks that path validation saw as regular files; writes through them now need permission approval. The PowerShell hole let variable-writing parameters silently overwrite $PSDefaultParameterValues and redirect later commands’ file access. Bash input redirections (< file) are now permission-checked like their argument spellings on all platforms.
sandbox.ripgrep is now honored only from user, managed and --settings settings, so a project can no longer swap the sandbox’s ripgrep binary through its project settings. Server-managed overrides of the sandbox binaries now need approval in the managed settings dialog. On gateways, empty policy or admin group entries and malformed email_domain values, which used to silently match no one or grant admin access, now fail at boot.
Secret redaction now covers GitLab token families such as glrt-, gloas- and glptt- and fully redacts routable glpat- and gldt- tokens, while the glab CLI config store gets the same sandbox and credential-path protection as gh. Plugin marketplaces also clone bare gitlab.com repo URLs, nested subgroups included, the way they clone github.com ones.
What changes in Remote Control
Remote Control sessions hosted by a bridge inside a cloud session no longer inherit that session’s transcript or credentials. Version 2.1.228 fixed a related problem, where running /resume during connection could bleed a restored conversation’s title or history into the current session. Sessions started from Claude Desktop or an IDE now reattach to their existing claude.ai session on resume, and Remote Control keeps reconnecting for about 30 minutes after a network blip.
The terminal now says whether a session was taken over by another device, ended from another app or deleted, and stops suggesting a reconnect that would undo it. Resuming a conversation also no longer silently takes Remote Control away from another Claude Code on the same machine that still holds it: running /remote-control there moves it.
Up next
Codex desktop app appears to be dropping Fast mode on API keys
Codex CLI hits 90.2 tokens/sec while the desktop app gets 54.7 on the same API key: per one developer’s benchmark, the desktop app strips service_tier=fast from requests and sends service_tier=None instead. There’s no Fast toggle in the desktop UI.
Related stories
- Claude Code 2.1.229: plugins load from a directory with no restart
- Claude Code 2.1.233 splits proxy spend by user
- Claude Code on desktop folds an iPhone Duo to test layouts
- Claude Code 2.1.318: login expiry warning, runner log change
- Claude Code 2.1.312 stops overwriting big CLAUDE.md files
- Claude Code 2.1.311 compacts past 200K instead of stopping
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

