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

Persistent Data Loss in Claude Projects: Conversations Disappearing Without Recovery
A long-form writer reports losing entire days of work in Claude Projects due to conversations disappearing from the project chat list, unsearchable and unrecoverable, with no response from Anthropic support after three incidents.

Tensions Escalate Between The Pentagon and AI Company Anthropic
The Pentagon's use of Anthropic's AI in classified operations, such as a raid in Venezuela, has created tension over the company's AI safety policies.

Claude 4.6 Adaptive Thinking: Reddit User Reports Token Waste and Provides Disable Commands
A Reddit user reports that Claude 4.6's new adaptive thinking feature can waste tokens and add latency in Claude Code, providing shell commands to disable it or cap thinking tokens.
The Mundane Risk: Why AI Safety's Biggest Threats Are Boring, Not Dramatic
An essay argues that mundane AI failures are already causing damage at scale, current alignment approaches depend too heavily on sandboxed environments, and capability convergence makes accidental open-world exposure increasingly plausible.