- [new] Added
--safe-modeflag (and correspondingCLAUDE_CODE_SAFE_MODEenvironment variable) that saves settings edits without loading them until safe mode is disabled. - [new] Added
mcp__github__getMCP tool. - [new] Added
meta+storykeybinding. - [changed] Refined multiple built-in prompts covering agent spawning rules, focus mode behavior, memory updates, refusal handling, and context management instructions.
Models & prompts
Prompt updates emphasize writing teammate-oriented summaries, avoiding unnecessary agent delegation, respecting explicit user action names for consent, and handling focus mode output visibility. Additional guidance was added around safe mode interactions and design system syncing workflows.
Under the hood
New environment variables were added for background session auth (CLAUDE_BG_*), session persistence forcing, MCP framing, bundled skills disabling, user dialog timeouts, and web fetch proxy behavior. Two new API beta headers (fallback-credit-2026-06-09, server-side-fallback-2026-06-01) were introduced alongside a new security advisories route.
Upgrade notes
Users relying on settings edits or custom permission rules should verify behavior when running with --safe-mode enabled.