Claude Code 2.1.227–2.1.229 — safer skills and steadier remote sessions
Claude Code hardens synced skills, improves Remote Control and self-hosted runners, fixes several crashes, and adds dynamic plugin marketplace sources.
Claude Code 2.1.227 through 2.1.229 focus on reliability and tighter boundaries around synced skills, remote execution, and automated Git workflows. Version 2.1.227 is a small fix release; 2.1.228 and 2.1.229 contain the changes most users will notice.
What changed
- Synced-skill hardening (2.1.228): skills synced from claude.ai can no longer shadow local
commands or MCP prompts. Their descriptions are sanitized and labeled, and their local bodies do
not execute
!commands or expand@file references. - Write behavior aligned (2.1.228): newer models can overwrite an existing file they have not read during the current session, matching the Edit tool's rules. Older models still require a read first.
- Remote Control improvements (2.1.228–2.1.229): cross-session messages are clearer, remote
session states are labeled
offlineorcloud, and several resume, spinner, and message-delivery failures are fixed. - Dynamic marketplace sources (2.1.229): a marketplace source can now be a local command that resolves a plugin directory each session, including a link mode that avoids copying it.
- Safer generated PR workflows (2.1.229):
/commit-push-prno longer auto-approves dangerous git and GitHub CLI flags such as force, amend, and no-verify. - Streaming and crash fixes (2.1.229): long responses no longer partly disappear and print twice, narrow terminals avoid progress/table crashes, and Windows UNC paths no longer crash the session.
- Gateway stability (2.1.229): SSE keepalive pings reduce idle-timeout disconnects during long thinking pauses on Vertex AI and Bedrock upstreams.
Why it matters
The synced-skill restrictions reduce the chance that cloud-sourced content can impersonate a local command or trigger local expansion features unexpectedly. Teams using Remote Control or self-hosted runners get a more predictable session lifecycle, while gateway users should see fewer connections die during long reasoning intervals.
The /commit-push-pr change is a sensible boundary: convenience automation should not silently
approve history-rewriting or safeguard-skipping flags. Existing workflows can still run those
operations when explicitly authorized, but they no longer receive automatic approval from the
generated flow.
What to do
claude update
Upgrade if you sync skills from claude.ai, use Remote Control, run self-hosted agents, connect
through Vertex or Bedrock, or rely on /commit-push-pr. Plugin authors using dynamic local
marketplaces should review the new command-source and link-mode behavior before adopting it.