Claude Code v2.1.232: Subagent Forking, GitLab Support, and Security Fixes

Claude Code v2.1.232 is out, and it brings a few significant changes: subagent forking is now default, GitLab support lands in plugin marketplaces, and several security patches address permission bypasses. If you're using Claude Code for multi-session workflows or in enterprise environments, this update is worth a look.
Subagent Forking On by Default
The headline feature: subagent_type: "fork" is now enabled by default. Forked subagents inherit the full conversation and prompt cache, which means they start with all the context you've built. Also, non-teammate agent spawns in interactive sessions now run in the background by default — no more blocking the main session.
Cross-Session Messaging via @
You can now type @ in the prompt to mention another Claude session by name. Claude uses SendMessage to reach that session directly. Bare names that exactly match a live session are delivered immediately (no confirmation prompt with a ref). Session names are also enforced to be unique on a machine; if you try to start or rename to a name already in use, you'll get a name-word-word variant and a notification.
New /config rows let you control "Dialog expiry" and "Messages from your other sessions" — the latter offers accept/hold/refuse for cross-session inbound requests.
GitLab Everywhere
Secret redaction now covers a whole family of GitLab tokens: glrt-, gloas-, glptt-, glagent-, glimt-, glsoat-, glcbt-, glft-, glffct-, plus full redaction of routable glpat-/gldt- tokens. The glab CLI config store gets the same sandbox and credential-path protection as gh.
Plugin marketplaces now support GitLab: bare gitlab.com repo URLs (including nested subgroups) clone just like GitHub URLs, and clone auth-failure hints name your actual Git host.
Settings and Enterprise Policy
additionalMarketplacesandallowedMarketplacesare accepted as friendlier aliases forextraKnownMarketplacesandstrictKnownMarketplaces.- A URL-typed
blockedMarketplacesentry for a bare repo URL continues to block that URL when the CLI classifies it as a Git clone. - Gateway: the
desktop:overlay now accepts every released Desktop setting (was 11 hand-listed keys), validated at boot against Desktop's own schema. Unknown or invalid keys fail boot. - Gateway: empty
managed.policies[].match.groups/admin.admin_groupsentries and malformedemail_domainvalues (empty, or containing@, whitespace, or commas) now fail at boot instead of silently matching no one or granting admin access.
Security Fixes
- Fixed a PowerShell permission bypass where variable-writing parameters could silently overwrite
$PSDefaultParameterValuesand redirect later commands' file access. - Fixed a Windows permission bypass where Git Bash followed Cygwin-style symlinks that path validation saw as regular files; writes through them now require permission approval.
- Fixed nested git repositories inheriting trust from a parent directory; each repository now requires its own trust confirmation.
Other Notable Fixes
- MCP connections no longer hang for the full 30-second connect timeout when a server fails to answer or sends a malformed reply to the protocol-version probe.
- Remote Control: sessions hosted by a bridge inside a cloud session no longer inherit that session's transcript or credentials; sessions started from Desktop/IDE reattach to the existing claude.ai session; idle sessions are reachable; bridge sessions restore conversation history on worker restart; resuming a deleted session starts a replacement instead of failing.
- Cloud gateway
/loginshows the reason when managed settings fail to load, instead of exiting silently or leaving an unresponsive terminal. - Voice mode on native builds shows rejection immediately when the voice service rejects the connection.
- mTLS client certificate rotation no longer requires a restart; Claude Code reloads the rotated cert and key automatically on connection errors.
- Malformed AWS or Vertex region values fall back to the default region instead of being used to build request URLs.
- Stream idle timeout errors recover properly on Bedrock, Vertex, and gateway deployments.
- Overlay rendering fixed for content-sized overlays with truncated text.
Also, Fable 5 is back as an advisor in /advisor for organizations with Fable access, with usage-credits consent set up through /model fable.
If you're running Claude Code in enterprise or multi-session setups, this release tightens security and improves cross-session interactions. Update and check your configurations.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Why AI Is Still Hard to Fully Deploy Across Enterprise Domains
A Reddit discussion highlights that probabilistic AI models struggle in high-accuracy fields like scientific research and report generation, where basic errors are unacceptable.

Exploring Clawra's Architecture and Social Autonomy Framework
David Im's Clawra experiments with a parallel world framework for AI companions, focusing on autonomy and local-first data privacy.

Nano‑Native Marketplace Paves the Way for Autonomous Agent Collaboration with NanoBazaar
NanoBazaar, the new nano-native marketplace, revolutionizes agent-to-agent work by allowing AI coding agents to collaborate autonomously and efficiently. Discover how this innovative platform empowers machine-driven transactions.

MLX Inference Performance Update: April 2026 Benchmarks and Features
MLX inference performance has improved significantly, with Qwen3.5-35B-A3B reaching 71.8 tokens/second at 4K context and new features like Multi-Token Prediction and SpecPrefill providing 2.3x-5.5x speedups for large models.