Claude Code hardens MCP, sessions, worktrees

This is the kind of release that matters most once teams start running Claude Code in anger: less glamour, more guardrails. Anthropic is clearly smoothing the rough edges around parallel sessions, tool-heavy workflows, and worktree-based agent isolation.

• –`worktree.baseRef` is the notable workflow knob here; it restores predictable branching behavior for `--worktree` and subagents, while preserving the option to keep unpushed commits with `head`
• –MCP got a real stability pass: OAuth now respects `HTTP(S)_PROXY`, `NO_PROXY`, and mTLS across discovery, registration, token exchange, and refresh
• –The session fixes are practical, not cosmetic: parallel sessions no longer dead-end after a refresh-token race, and compromised session locks now fail more cleanly
• –Subagents can now discover project, user, and plugin skills via the Skill tool, which should reduce weird “why didn’t the agent see that skill?” failures
• –Exposing `$CLAUDE_EFFORT` to hooks and Bash makes downstream automation more aware of the current operating mode, which is useful for policy and observability glue