v2.1.211

Source: Anthropic Claude Code Date: 2026-07-15 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.211

Summary

Claude Code v2.1.211 is a fix-and-hardening release: it adds a --forward-subagent-text flag for surfacing subagent output in stream-json, closes a permission-preview spoofing gap (bidirectional-override, zero-width, and look-alike characters could visually alter approval text), hardens Claude-in-Chrome file-upload path validation, and fixes an auto-mode bug where a PreToolUse hook’s ask decision was being overridden for unsandboxed Bash. Also includes background-agent reliability fixes (respawn-after-kill, still-running status reporting) and cross-cloud auth/model-selection fixes for Vertex/Bedrock.

Implications

• Agent-runtime hardening thread: this is another entry in the multi-release arc (207→214→…) hardening the provenance and integrity of consent — the bidi/zero-width/look-alike character fix specifically prevents an approval prompt’s displayed text from being visually spoofed, the same “who said yes, and can you trust what you saw” concern as prior releases in this arc.
• Auto-mode/hook boundary: fixing auto mode overriding a hook’s explicit ask decision closes a gap where the harness’s own automation could bypass a safety control meant to stop it — consistent with the broader pattern of restraint-on-self-granted-authority seen elsewhere in the CC 2.1.2xx series.
• Routine maintenance otherwise; no capability expansion, no new CVE.