v2.1.206
protocolsagentsmodels
read at source ↗ github.com
v2.1.206
Source: Anthropic Claude Code Date: 2026-07-10 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.206
Summary
Claude Code v2.1.206 adds a confirmation gate before EnterWorktree enters a git worktree outside .claude/worktrees/, makes background agents self-upgrade right after a CC update instead of stalling on next attach, and fixes --bg/background-worker env-var propagation (CLAUDE_CODE_EXTRA_BODY) and MCP request_timeout_ms being ignored. A new /doctor check proposes trimming checked-in CLAUDE.md content that Claude can already derive from the codebase, and /code-review findings quality improves on claude-opus-4-8.
Implications
- Autonomy-sprint hardening. The worktree confirmation gate is the same pattern as recent releases’ session-transcript and rm-rf guards: as agents get more filesystem and process autonomy, each release adds one more place where an irreversible or scope-expanding action needs an explicit checkpoint. The self-upgrade fix for background agents is the inverse move — removing friction where staleness (not autonomy) was the risk.
- Agent-layer convergence. The
/doctorcheck that flags derivableCLAUDE.mdcontent is notable as tooling starting to police the operator’s own instructions-to-the-agent, not just the agent’s actions — a small step toward the harness maintaining its own context budget. - Routine reliability fixes (MCP timeout, OAuth re-auth,
/modelpicker pricing rows) — bookkeeping, not thread-worthy on their own.