v2.1.191

Source: Anthropic Claude Code Date: 2026-06-24 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.191

Summary

Claude Code v2.1.191, released 2026-06-24 (~6 hours after v2.1.190). Substantive patch: adds /rewind command for resuming conversation from before a /clear was run; fixes stopped background agents resurrecting in the task panel; fixes hooks with comma-separated matchers silently never firing; fixes /permissions Recently-denied tab discarding approvals on close; improves MCP server reliability with retry logic for transient errors on capability discovery; reduces CPU usage during streaming by ~37% via 100ms coalescing; reduces long-session memory growth from terminal output cache.

Implications

• Agent-host correctness. The background-agent resurrection fix and the hooks comma-matcher silent-failure fix are both correctness bugs in the agent-management layer — exactly the class of bug that degrades trust in autonomous operation. Both were presumably surfaced at scale quickly given same-day patching.
• MCP reliability hardening. Retry logic on tools/list / prompts/list / resources/list with backoff is a maturity signal: the MCP integration is now resilient enough to be worth protecting against transient network errors rather than fast-failing.
• Performance baseline. 37% CPU reduction during streaming is non-trivial for interactive use. As sessions grow longer and agentic loops emit more output, streaming efficiency becomes a real UX factor.
• /rewind is architecturally interesting. Resuming from before /clear implies the engine is retaining conversation state that was previously treated as discarded. Worth watching for whether this enables longer-horizon state management in future iterations.