Version 2.1.308 brings several improvements to API handling, subagent workflows, and error recovery, along with better control over settings and a smoother sign-in process.
Commands and interface
You can now move Claude API code to version 2026-10-21 with the new /claude-api api-version command. The /usage command has been improved—each warning now provides a one-line fix with its source, and you can apply fixes with f. These same fixes also appear in VS Code and /doctor.
The /artifacts delete confirmation has changed: press 1 or Enter to keep, 2 to delete, and Esc to cancel—y and n no longer work.
Tip
If you work with API integrations, updating to the latest version ensures compatibility with newer features.
Hooks and automation
The CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS setting has been added, so unattended retry mode (CLAUDE_CODE_RETRY_WATCHDOG) can detect if a usage limit is lifted before its reset time.
The axScreenReader setting and CLAUDE_AX_SCREEN_READER in an env block now only apply from user, managed, or --settings configurations. This prevents unintended overrides from other sources.
Subagents and workflows
Forked subagents launched together now receive information about their own launch, reducing cases where they mistakenly try to start each other’s work. This makes parallel workflows more reliable.
Settings and security
The sandbox.allowNestedUserNamespaces setting now allows sandboxed commands on Linux to create user namespaces, which helps tools like Chromium and rootless podman bypass the sandbox’s file protections.
The allowedMcpServers setting no longer filters managed-mcp.json stdio servers that use ${VAR} only in env. If a server was previously blocked due to this, it will now load on upgrade. Use deniedMcpServers to block it if needed.
Error handling and recovery
Claude Code now retries unparseable tool calls up to twice before ending the turn with an error, improving reliability when API responses are malformed.
The notice Claude receives when the WebSearch budget is exhausted now also instructs it not to attempt searches through browsers or computer use, preventing wasted attempts.
User experience
Browser sign-in is faster—the success page appears immediately after sign-in completes, without waiting for your account profile to load. Remote Control connections also improve for sessions containing large pasted images or PDF reads.
MCP consent dialogs are less likely to misinterpret an Enter meant for the prompt as an “Accept” action, reducing accidental confirmations.
Everything else, briefly
- Improved recovery from unparseable tool calls
- Better coordination between forked subagents
- Faster browser sign-in
- Smoother Remote Control connections with large files
- WebSearch budget notice now prevents redundant search attempts
Where the course covers this
- Unattended runs with limits
- Choosing an orchestration pattern
- Model, effort and the cost of an answer
- /goal, /loop and Agent view: work that runs without you
Read next
- The Claude Code practicum on GitHub: 14 hands-on tasks with automatic checks
- Claude Code models, effort and fast mode: which knob to turn, and when
For the full list of changes, see the official CHANGELOG.