Claude Code v2.1.191: /rewind, CPU fixes, MCP reliability improvements
Anthropic released Claude Code v2.1.191 with a handful of fixes and improvements that directly affect daily agent usage. Notable additions include /rewind support, a ~37% CPU reduction during streaming, and several MCP reliability enhancements.
New command: /rewind
/rewind now lets you resume a conversation from before a /clear was run. If you accidentally wiped context, you can restore the prior session without restarting from scratch.
CPU usage and memory
Streaming response CPU usage is reduced by ~37% via coalescing text updates to a 100ms interval. Long-session memory growth from terminal output cache has also been reduced.
Agent fixes
- Stopping an agent from the tasks panel is now permanent — background agents no longer resurrect after being stopped.
- Builtin slash commands (like
/usage) sent to background sessions now show a hint instead of being passed as prompt text. - Job rows for pasted images now display
[Image #N]placeholder instead of full filesystem paths.
MCP improvements
- Capability discovery (
tools/list,prompts/list,resources/list) now retries on transient network errors with short backoff. - MCP OAuth discovery and token requests retry once after transient errors; headless environments skip the browser popup and go straight to the paste-the-URL prompt.
- HTTP 404 errors now show the URL and point to your MCP config.
Other fixes
- Scroll position no longer jumps to the bottom while reading earlier output during a streaming response.
/voicenow explains the organizational restriction instead of showing a generic message./loginURL no longer gets truncated in Windows Terminal when it wraps across lines.- Cmd+click on links in fullscreen mode works again for Ghostty over ssh/tmux.
- Hooks with comma-separated matchers (e.g.,
"Bash,PowerShell") now fire correctly. /permissionsRecently-denied tab: approving a denial now persists on close.- Managed settings:
forceRemoteSettingsRefreshnow works via MDM or file policy, and the fetch sendsCache-Control: no-cacheto prevent stale proxy responses. - Sandbox network permission dialog now remembers hosts allowed with "Yes" for the session.
- Vim mode prompt-history search (NORMAL
/) hints how to reach slash commands.
📖 Read the full source: GitHub Claude-Code
👀 See Also
The Mundane Risk: Why AI Safety's Biggest Threats Are Boring, Not Dramatic
An essay argues that mundane AI failures are already causing damage at scale, current alignment approaches depend too heavily on sandboxed environments, and capability convergence makes accidental open-world exposure increasingly plausible.
Analysis of 413K AI Agent Runs Reveals What Makes Them Succeed
An analysis of 413,278 AI software engineering agent runs from the CoderForge-Preview dataset shows that human software engineering best practices often harm agent performance. The data reveals specific patterns that separate successful from failing runs on the same problems.
Qwen KV Cache Quantization Deep Dive: PPL, KL Divergence, and Asymmetric K/V Results
Second round of benchmarks on Qwen 3.6-35B-A3B with KV cache quantization: perplexity, KL divergence, asymmetric K/V combos, and 64K context depth on Apple M5 Max.
AI Fails to Boost Productivity According to Recent CEO Study
Despite widespread AI adoption, a study of 6,000 executives reports negligible impacts on productivity and employment, echoing the productivity paradox identified in the IT era of the 1980s.