v2.1.157

Source: Anthropic Claude Code Date: 2026-05-29 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.157

Summary

Claude Code v2.1.157 ships two structurally significant changes alongside a large bug-fix sweep: plugins in .claude/skills/ directories now auto-load without marketplace registration, and the agent field in settings.json is now honored for dispatched sessions with a --agent <name> override flag. EnterWorktree gains mid-session worktree switching, and tool_decision telemetry can now log tool parameters (bash commands, MCP/skill names) under OTEL_LOG_TOOL_DETAILS=1. The release also resolves a cluster of multi-agent lifecycle bugs — orphaned worktrees, leaked background shells, sleep/wake date desync, and --resume not surfacing subagents running at prior exit.

Implications

• Feeds agentic engineering patterns: auto-loading local skills without a marketplace step lowers the friction for agent-local capability extension significantly — skills become first-class artifacts that ship with the worktree, not afterthoughts registered centrally.
• Feeds fleet-ops hardening: the subagent lifecycle fixes (orphaned worktrees, leaked shells, retirement on idle, resume continuity) represent incremental hardening of multi-agent session management — each fixed edge case is a reliability invariant that fleet operators can now depend on.
• Feeds inter-agent trust: honoring the agent field in dispatched sessions means the dispatching agent can now explicitly name the sub-agent model/persona, tightening the identity contract between orchestrator and worker — relevant to any trust model built on session-level agent identity.