Claude Code v2.1.133: worktree.baseRef revert, sandbox paths, proxy fix for MCP OAuth

Anthropic released Claude Code v2.1.133 with several configuration additions and critical fixes. The notable update is the worktree.baseRef setting, which controls whether --worktree, EnterWorktree, and agent-isolation worktrees branch from origin/<default> or local HEAD. The default is fresh, reverting EnterWorktree to branch from origin/<default> — this changes the behavior from v2.1.128 which used local HEAD. To keep unpushed commits in new worktrees, set worktree.baseRef: "head".
New Settings & Features
- worktree.baseRef (
fresh|head): Choose worktree base branch. Defaultfreshusesorigin/<default>; set toheadto inherit localHEAD. - sandbox.bwrapPath and sandbox.socatPath (Linux/WSL): Specify custom bubblewrap and socat binary paths.
- parentSettingsBehavior (admin-tier key,
first-wins|merge): Allows admins to opt SDKmanagedSettingsinto policy merge. - Hooks now receive effort level: Via
effort.levelJSON input and$CLAUDE_EFFORTenvironment variable. Bash tool commands can read$CLAUDE_EFFORT.
Improvements & Fixes
- Focus mode: Behavior improved.
- Memory usage: Warm-spare background workers released under memory pressure.
- Parallel sessions 401: Fixed dead-end after refresh-token race wiping credentials.
- Edit/Write allow rules: Fixed scope matching on drive root (
C:\) or POSIX/incorrectly always prompting. - ECOMPROMISED lock error: Fixed unhandled rejection from clock skew or slow disk corrupting session-log file locks.
- Spurious compaction error: Pressing Esc during compaction no longer shows "Error compacting conversation".
- HTTP(S)_PROXY / NO_PROXY / mTLS: Now respected for the full MCP OAuth flow (discovery, dynamic client registration, token exchange, token refresh).
- Mapped network drives: Read/Write/Edit allowed when passed via
--add-diror SDKadditionalDirectories. - Remote Control stop: From claude.ai now fully cancels CLI session like local Esc, preventing stuck queued messages.
- Effort cross-session leak:
/effortin one session no longer changes others; IDE effort changes no longer silently dropped. - Subagent skill discovery: Fixed subagents not finding project, user, or plugin skills via the Skill tool.
CLI & VSCode
claude --helpnow lists--remote-controlalongside--remote-control-session-name-prefix.- [VSCode] Fixed
claudeCode.claudeProcessWrapperfailing with "Unsupported platform" when extension build lacks bundled Claude binary.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Claude's policy filter blocks bioinformatics work with pathogen names
A computational virology researcher reports Claude's usage policy filter flags legitimate bioinformatics scripts when pathogens are named, requiring workarounds like describing tasks without organism names or downgrading to Sonnet 4. The issue affects Claude Code, claude.ai, and both Opus 4.6 and Sonnet 4.6 models.

Constraint Decay: Why LLM Agents Fail at Structured Backend Code
New research introduces 'constraint decay': as structural requirements accumulate, LLM agent performance drops drastically — capable agents lose 30 points in assertion pass rates, weaker ones approach zero. Actionable insights for anyone using AI coding agents.

OpenClaw loses cost-effective access to GPT and Claude models
OpenClaw users can no longer use Anthropic models without paying high API fees, and OpenAI has severely reduced Business and Teams account quotas to near free-tier levels, forcing users toward Chinese or local model alternatives.

Anthropic enforces policy: third-party Claude harnesses no longer covered by subscription limits
Anthropic is enforcing a policy change effective April 4 where third-party harnesses like OpenClaw no longer draw from Claude subscription usage limits, requiring users to turn on extra usage or cancel by April 9 for a refund.