Claude Code Just Made Subagents Free — And They Remember Everything

Claude Code v2.1.232 landed 6 hours ago (August 13, 23:29 UTC) and it’s the kind of release that changes how you work — not just what the tool can do in theory.

The headline: subagent forking is now on by default. When you spawn a subagent with subagent_type: "fork", it inherits the full conversation history and prompt cache. Non-teammate agents in interactive sessions now run in the background automatically. Your agents don’t just delegate — they remember.

Add @ mentions to reach other sessions directly, GitLab support (finally), and Fable 5 advisor access for orgs that have it, and this is the most collaborative Claude Code has ever been.

The Feature That Changes Everything: Subagent Forking by Default

Until now, subagents started fresh. They knew the task you gave them, but not the context you built — the files you read, the decisions you made, the bugs you already found. That changes in v2.1.232.

subagent_type: "fork" now means: the subagent gets the entire parent conversation and prompt cache. It picks up exactly where you left off.

// In your agent config or skill
{
  "subagent_type": "fork"
}

What this means in practice:

• Debugging session → handoff to fixer subagent: The fixer sees every stack trace, every log, every hypothesis you already tested
• Research → implementation handoff: The implementer inherits all the API docs, code patterns, and architectural decisions you gathered
• Code review → refactor subagent: The refactorer knows the exact style violations, the test expectations, the “why” behind every comment

Background by default for non-teammate agents means you spawn a subagent and keep working. It runs in parallel, reports back when done. No more blocking your main session.

@ Mentions: Your Sessions Can Finally Talk to Each Other

Type @ in the prompt and you’ll see a list of your other live Claude sessions — by name. Pick one, and Claude uses SendMessage to reach it directly.

@backend-session "The API contract changed: /users now returns {id, email, created_at}. Update your tests."

What’s new in v2.1.232:

• SendMessage delivers to a bare name that exactly matches a live session — no more “confirm with a ref” prompt
• Sessions on one machine keep unique names automatically: if you start/renames to a name another live session uses, you get name-word-word and a notice
• New /config rows: “Dialog expiry” and “Messages from your other sessions” (accept/hold/refuse inbound cross-session messages)

This isn’t “agents talking” in the abstract. It’s you orchestrating multiple parallel sessions — frontend, backend, infra, docs — each with full context, able to sync when interfaces change.

GitLab Support: No More “It Only Works with GitHub”

Bare gitlab.com URLs (including nested subgroups like gitlab.com/group/subgroup/project) now clone like GitHub URLs. Clone auth-failure hints name your actual git host.

Plugin marketplace support for GitLab:

• additionalMarketplaces / allowedMarketplaces — friendlier aliases for extraKnownMarketplaces / strictKnownMarketplaces
• Enterprise policy: blockedMarketplaces with bare repo URLs keeps blocking even when CLI classifies it as a git clone

If your company uses GitLab, Claude Code just became a first-class citizen in your workflow.

Fable 5 Advisor Access (For Orgs With Fable Access)

Fable 5 (Anthropic’s Mythos-family model with extra safeguards) is back in /advisor for organizations with Fable access. Consent flows through /model fable.

This matters if you’re in a regulated environment or need the strongest guardrails — Fable 5 is built for high-stakes coding where hallucination risk must be near zero.

The Fixes You’ll Actually Notice

Windows & PowerShell (Finally Taken Seriously)

• PowerShell permission bypass fixed: Variable-writing parameters could silently overwrite $PSDefaultParameterValues and redirect later commands’ file access
• Windows Git Bash symlink bypass fixed: Cygwin-style symlinks that path validation saw as regular files now require permission approval
• Nested git repos no longer inherit parent trust: Each repo requires its own trust confirmation

Remote Control (Actually Reliable Now)

• Sessions hosted by a bridge inside a cloud session no longer inherit that session’s transcript or credentials
• Resuming a conversation whose session was deleted from claude.ai now starts a replacement instead of failing with a login error
• Remote Control keeps reconnecting for ~30 minutes after network blips (was: drops after a few blips across an hour)
• Terminal now tells you why a session ended: taken over by another device, ended from another app, or deleted

MCP & Gateway

• MCP connections no longer hang for the full 30-second timeout when a server fails to answer or sends a malformed protocol-version probe
• Stream idle timeout errors now recover on Bedrock, Vertex, and gateway deployments instead of failing the request
• Gateway: empty managed.policies[].match.groups and malformed email_domain values now fail at boot instead of silently granting admin access
• mTLS client certificate rotation no longer requires a restart — Claude Code reloads rotated cert/key automatically on connection errors

Secret Redaction (Expanded)

• GitLab token families now redacted: glrt-, gloas-, glptt-, glagent-, glimt-, glsoat-, glcbt-, glft-, glffct-
• Full redaction of routable glpat-/gldt- tokens
• glab CLI config store gets same sandbox and credential-path protection as gh

Agent Panel & UX

• Completed subagents hide immediately with a /tasks footer hint
• “��� N more” overflow indicator moved left for visibility
• /feedback and /bug now open immediately while Claude is responding (no waiting for turn to finish)
• /plugin install plugin@marketplace refreshes marketplace first — newly published plugins install without manual update
• /code-review at high/xhigh/max effort now runs in a background agent like other levels

What Breaks (Almost Nothing)

This is a patch release (2.1.228 → 2.1.232), so breaking changes are minimal:

• sandbox.ripgrep now honored only from user, managed, and --settings — project settings can no longer override the sandbox’s ripgrep binary (security hardening)
• Startup tip suggesting custom subagents removed (the feature is now mainstream)
• Hardened auto-generated cross-session socket directory on shared /tmp — pre-planted symlinks or another user’s directory now refused
• Hardened Linux filesystem sandbox against protected-path bypass

If you had project-level sandbox.ripgrep config, move it to user or managed settings.

How to Update

npm update -g @anthropic-ai/claude-code
# or
claude /update

Verify:

claude --version
# Should show 2.1.232

Test the new features:

1. Spawn a subagent with subagent_type: "fork" — watch it inherit your full context
2. Type @ in any prompt — see your other live sessions listed
3. Clone a GitLab repo with a bare gitlab.com/... URL — it just works
4. Check /config for the new “Dialog expiry” and cross-session message rows

The Bottom Line

v2.1.228 fixed Windows crashes. v2.1.232 makes Claude Code collaborative.

Subagent forking by default means delegation doesn’t mean amnesia. @ mentions mean your parallel sessions can sync without you playing telephone. GitLab support means your tool fits your infra, not the other way around.

If you’ve been running multiple Claude sessions and manually copying context between them — stop. Update to 2.1.232 and let the agents talk.

Related Reading

• Claude Code v2.1.228: Windows Git Fix, Interactive Session Crash, Cross-Session Messaging Leaks
• Claude Code v2.1.224: Self-Hosted Runners, Cross-Session Messaging, and Tighter Secret Handling
• Beware: Claude Code Cross-Session Content Bleed — GitHub Issue #77147
• Claude Code Skills vs Subagents vs MCP: Which One Actually Does What?

Related articles

• Claude Code Alternatives in 2026: 12 Options Compared
• Oh My Pi v17.3.3 Fixes Gemini Reasoning Loops, Hashline Edge Cases, and TUI Rendering — What Beginners Need to Know
• Claude Code: Skills vs Subagents vs MCP — The 2026 Decision Guide

Updated August 14, 2026 — covers v2.1.232 released August 13, 2026 23:29 UTC. Full changelog: github.com/anthropics/claude-code/releases/tag/v2.1.232