v2.1.205

Source: Anthropic Claude Code Date: 2026-07-08 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.205

Summary

Claude Code v2.1.205 is a consent-hardening release: it blocks tampering with session transcript files, makes background-task notifications explicitly state that no human input occurred (so a fabricated in-transcript approval can’t be acted on), and has auto mode ask before running rm -rf on a variable it can’t resolve. The rest is agent-view polish (PR linking, status accuracy) and a Windows worktree-deletion fix.

Implications

• Autonomy-sprint hardening. This is the throughline for the arc’s recent releases: as background/multi-agent workflows multiply, the approval trail itself becomes attack surface. 205 treats unverifiable consent as a security boundary rather than a UX detail — continuing, not closing, the hardening thread.
• Agent-layer convergence. Incremental but telling: agent-view now shows classifier-written headlines instead of raw tool-call text, and PRs get linked automatically across edit/merge/comment/push paths — the harness is absorbing more judgment about what a session did, not just logging it.