Configurable Worktree Base Branch

Highlight

Worktree Base Branch Is Now Configurable

A new worktree.baseRef setting lets you choose whether --worktree, EnterWorktree, and agent-isolation worktrees branch from origin/<default> or your local HEAD. The default flips back to fresh (origin), which reverses the local-HEAD behavior introduced in 2.1.128 — if you rely on unpushed commits being carried into new worktrees, set worktree.baseRef: "head" explicitly.

Worth Knowing

Hooks and Bash Commands Can Read the Effort Level

Your hooks now get the active effort level through an effort.level field in the JSON input and a $CLAUDE_EFFORT env var, and Bash tool subprocesses see the same variable. This lets you branch hook logic or shell commands on whether the session is running at low, medium, or high effort.

Load Plugins Directly From a URL

The new --plugin-url <url> flag fetches a plugin .zip archive from a URL and loads it for the current session, complementing --plugin-dir's new ability to accept .zip archives directly. Useful for trying out a plugin without committing it to a marketplace or local directory.

Opt Out of the Fullscreen Renderer

Setting CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1 keeps the conversation in your terminal's native scrollback instead of the alternate-screen fullscreen renderer. Helpful if you prefer to keep transcripts in scrollback or your terminal handles the alternate screen poorly.

Ctrl+R History Search Goes Global by Default Again

The Ctrl+R prompt history picker once again searches across all projects by default, restoring pre-2.1.124 behavior. Press Ctrl+S inside the picker to narrow to the current project or session.

Skill Overrides Setting Now Takes Effect

The skillOverrides setting now actually applies: off hides a skill from both the model and slash menu, user-invocable-only hides it from the model while keeping it in /, and name-only collapses the description so the model sees only the name. A practical knob if you want certain skills available on demand but out of the model's default toolbox.

MCP Server List Shows Tool Counts and Failure States

The /mcp view now shows tool counts for each connected server and flags servers that connected but exposed zero tools or failed tools/list. If you run several MCP servers, it's now obvious at a glance which ones are actually delivering tools versus quietly failing.

Under the Hood

Multiple OAuth and MCP Token Refresh Races Fixed

Several concurrency bugs around credential refresh have been fixed: MCP OAuth refresh tokens are no longer lost when multiple servers refresh at once (so you shouldn't need daily re-auth), parallel sessions no longer all 401 after a refresh-token race, and a wake-from-sleep refresh race that could log out every running session is resolved.

Memory and Cache Fixes for MCP and Sub-Agents

A stdio MCP server writing non-protocol data to stdout no longer balloons RSS past 10GB, warm-spare background workers now release under memory pressure, and sub-agent progress summaries now hit the prompt cache for roughly a 3× reduction in cache_creation tokens.

From Anthropic

In the News