Claude Code 2.1.229: plugins load from a directory with no restart

Anthropic on August 12 changed how Claude Code finds plugins. In version 2.1.229, a plugin marketplace can take its source from a local command, such as an IDE, that prints the path to a plugin directory. Claude Code resolves that path again at the start of every session and applies what it finds without a restart.
The release notes on GitHub list 32 items, and 18 of them are bug fixes. Most of the rest deal with Remote Control, self-hosted runners and traffic that goes through gateways.
Plugins get one more option. With mode: "link", Claude Code uses the directory in place, where it already sits. The new source lets another tool decide which directory is current, and that answer can change from one session to the next. A separate fix covers one-shot claude plugin commands, which could leave a stray liveness file behind and block cleanup of outdated plugin versions.
claude remote-control --continue, which resumes your most recent Remote Control session, is now documented. ListAgents now marks disconnected Remote Control sessions as offline and labels your cloud sessions as cloud. Anthropic’s Remote Control documentation describes the web and mobile clients as a window into a session running on your own machine, so the computer has to stay on and the claude process has to keep running. The same docs say a session reconnects on its own once a sleeping laptop or a dropped network comes back. An offline label describes the state of your machine, while a cloud session runs on Anthropic’s infrastructure. A smaller fix: Remote Control clients no longer show a stuck working spinner after a slash command is typed in the laptop terminal.
Streaming through gateways now includes SSE keepalive pings, small messages sent on an open connection so it doesn’t look idle, during long thinking pauses. Without them, Vertex and Bedrock upstreams were dropping the connection on idle timeout. Three related fixes:
/modelwas rejecting Sonnet and Opus 1M for claude.ai subscribers who route through a customANTHROPIC_BASE_URLgateway.- Auto mode failed on every tool call for users on direct Anthropic API connections who disable the attribution header with
CLAUDE_CODE_ATTRIBUTION_HEADER. - OpenTelemetry export from Claude Desktop sessions was rejected when the Desktop-managed gateway was also the telemetry endpoint.
The self-hosted runner mode added in 2.1.224 gets several changes. Its sessions now receive Claude Code hooks from the server, the same way managed environments do. Runner sessions and other remote sessions used to exit at startup when managed-mcp.json was deployed and the server also delivered MCP servers; those servers are now skipped with a warning. Repository preparation could hang on a Git Credential Manager prompt, and git now fails fast when credentials are missing. On Windows, runner startup requires an explicit --base-dir, since there is no default checkout directory there.
One change hits the bill directly. When a workflow fans out into sibling agents that share a prompt prefix, Claude Code now staggers their start, so later agents read the cached prefix instead of each paying for it again. Setting CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS=0 turns the delay off. Dynamic workflows inside CPU-limited containers also stop reading the host’s core count and use the container’s CPU limit.
Most of the remaining fixes are about crashes and the terminal:
- Long responses partly disappeared during streaming and then printed twice.
- A tool call with a non-string
glob,file_pathorcommandvalue crashed to the error screen, including on--resumeof that session. - A progress bar or markdown table in a very narrow terminal threw a RangeError, which could also break
claude --continueand--resumeat startup. - On Windows, extended-length (
\\?\) and UNC paths caused a crash. - Editing a file with thousands of IDE diagnostics while the IDE extension was connected caused multi-second UI stalls.
- A whitespace-only message in SDK and
--input-format stream-jsonsessions returned a 400 error.
Conversations whose messages alone exceed the API’s 32 MB request limit kept retrying compaction when there were no images or documents to strip. They now fail once with a clear message, and “prompt is too long” errors explain why automatic compaction couldn’t recover instead of only suggesting /compact.
The VS Code extension adds session groups in the sidebar, a resizable /btw panel, and a built-in feedback dialog for /bug that replaces a retired survey link. /login now repeats the CLAUDE_CODE_OAUTH_TOKEN override warning after a successful login.
The sandbox now brackets IPv6 literals in network domain lists ([::1]:443), enforces ambiguous spellings fail-closed and flags them in /doctor. MCP OAuth puts 127.0.0.1 in the redirect URI instead of localhost, which fixes sign-in against strict authorization servers. The Claude Code Review workflow generated by /install-github-app had been completing its run without posting a review on the pull request. And in /commit-push-pr, git and gh commands with dangerous flags such as --force, --amend and --no-verify are no longer auto-approved.
Up next
Auto mode let Claude Code run downloaded code in 6 of 10 tests
Claude Code downloaded a Python script from an external server and ran it in 6 of 10 prompt injection tests, with auto mode’s classifier approving it. Same call: 78 under a Python audit hook, blocked. 15 inside a homemade automation script, passed.
Related stories
- Claude Code 2.1.232: forked subagents are on by default
- Claude Code 2.1.228 patches a conversation-history leak in Remote Control
- OAuth login to MCP servers is fixed
- Claude Code sessions can now run on your own machines
- Claude Code on desktop folds an iPhone Duo to test layouts
- Claude Code 2.1.318: login expiry warning, runner log change
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

