Claude Code v2.1.228: Fixed Git Detection, TUI Model Reverts, and Skill Security

Claude Code's latest release, v2.1.228, is out with a batch of stability fixes and a few behavior changes worth noting before you update.
Key Fixes
- Windows Git detection: Fixed an issue where
git/ Git Bash wasn't found when launching from a parent directory of the git installation — if you kept hitting that, this should resolve it. - /tui model reversion: The TUI no longer reverts to an earlier model when
/modelwas changed mid-session. That's a nasty one to lose work to. - Session cleanup: Fixed a bug where cleanup could delete contents inside a project's
memoryfolder. Critical if you rely on that directory. - Plugin cache cleanup: A symlinked development checkout of a plugin is no longer removed when it's the only version present.
- Settings merge: Marketplace entries defined in a higher-precedence tier now merge as whole entries; they no longer accidentally pick up another tier's custom headers.
- Remote Control resume: No more leaking the resumed conversation's title/history into the session when you
/resumewhile connected. - Self-hosted runners: Fixed crashes when a checkout hook fails for a repo you don't push to, and a gap where sessions could end between a background task and the follow-up turn.
Security & Skills Hardening
Skills synced from claude.ai are now more safely sandboxed:
- They no longer shadow local commands or MCP prompts.
- Their descriptions are sanitized.
- Their bodies don't run
!commands or expand@files on your machine.
Behavior Changes
- Write tool: Newer models can now overwrite an existing file they haven't read this session (same as the Edit tool). Older models still require a read first.
- Vertex AI: Expired Google Cloud credentials now fail in seconds, not minutes.
- Compaction progress: Shows a retry countdown and stall hint — not just a bare progress bar.
- Terminal title spinner: Updated busy-spinner glyphs to reduce tab-bar jitter.
- Cross-session messages: Sender and body display inline rather than collapsed.
If you're on an older build, this is a solid maintenance release to upgrade to — especially if you've hit any of the Windows Git, TUI model, or session cleanup bugs.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Claude Code v2.1.205: Transcript Tampering Blocked, JSON Schema Fixes, and Auto-Mode Improvements
Claude Code v2.1.205 adds auto-mode rule blocking session transcript tampering, fixes large JSON schema issues, improves background agents handling, and streamlines binary updates.

Anthropic separates Claude subscriptions from third-party tool usage
Anthropic is ending Claude Pro/Team subscription coverage for OpenClaw usage starting April 4, requiring separate pay-as-you-go billing for third-party harnesses. Users must enable 'extra usage' in account settings to continue using Claude through OpenClaw.

Anthropic source code leaks reveal unannounced Claude features and internal model
Anthropic accidentally leaked 500,000 lines of source code containing details about unannounced Claude features including KAIROS background execution, dream mode, undercover mode, and an internal model called capybara. This is the second such leak in 2025.

Scoring Show HN Submissions for AI Design Patterns
A developer analyzed 500 Show HN landing pages to detect common AI-generated design patterns like Inter fonts, colored left borders, and glassmorphism. The scoring system identified 21% of sites as 'heavy slop' with 5+ patterns.