Claude Code shipped version 2.1.191 on June 24 with a feature that looks minor in the changelog and is significant in practice: /rewind
can now recover a conversation you cleared with /clear
. If you’ve ever wiped a long session and immediately regretted it, this update is for you. Add a 37% CPU reduction during streaming and a fix for background agents that refused to die, and 2.1.191 is worth reading even if you’ve been skimming patch notes.
/rewind Finally Reaches Before /clear #
Previously, /rewind
only navigated within the current session. Clear your conversation and you lost everything — no path back. Version 2.1.191 changes that. Run /rewind
after a /clear
and you get a keyboard-navigable picker of your full conversation history, including the turns that existed before the wipe.
The mechanic itself is worth understanding. After each Claude response, Claude Code snapshots every file Claude has directly modified. When you invoke /rewind
, you choose a point in that history and pick one of three restoration modes:
Restore code and conversation— rolls back file edits and trims conversation history to that point** Restore conversation only**— preserves your current file state but rolls back the context Claude is working with** Restore code only**— reverts files to the selected point, removes subsequent messages
This release also adds a “Summarize up to here” option in the rewind menu. Rather than restoring the full verbose history, you can compress earlier context while keeping recent turns intact — useful when you want the gist of a long session without burning through tokens re-reading it.
One limitation worth knowing: /rewind
does not undo file changes made via bash commands. If Claude ran rm old.txt
or mv src dest
through the shell tool, those operations are outside the snapshot system. Only direct file writes from Claude’s file tool are tracked. The checkpointing documentation has the full reference on what is and isn’t captured.
37% Less CPU During Streaming #
Anthropic coalesced text updates into 100ms batches instead of rendering on every incoming token. The result: a 37% drop in CPU usage during streaming responses. On a MacBook Pro, that translates directly into battery life and fan noise during long sessions.
The performance work goes further. The /goal
status chip previously re-rendered the terminal at 5 Hz while idle — doing work even when nothing was happening. That’s fixed. UI re-renders during parallel subagent execution are also reduced. There’s also a memory suppression fix for the terminal output cache, which was leaking in sessions running longer than an hour or two.
For developers running Claude Code through an eight-hour workday, these compound. The tool was already capable; now it’s less likely to slow your machine down while being capable.
The Agent Resurrection Bug Is Finally Dead #
GitHub issue #58662 had over 200 upvotes. The complaint: stopping a background agent from the Tasks panel caused it to immediately resume. The fix is in 2.1.191 — stopping an agent from the Tasks panel is now permanent.
Related: background jobs were getting permanently stuck in a “working” state when an agent ended a turn without producing structured output. Also fixed. For teams running multi-agent workflows, both of these were serious reliability issues that made orchestration unpredictable.
Two agent policy fixes also shipped: Agent(type)
deny rules and Agent(x,y)
allowed-types restrictions were not being enforced for named subagent spawns. Enterprise teams setting allow/deny policies on agent types were getting no actual enforcement. That’s corrected now — policies apply to named subagent spawns as intended.
MCP, OAuth, and Credential Sandboxing #
If you run Claude Code with a custom API gateway or stateful MCP servers, two regressions were silently affecting you. Stateful MCP servers without the optional GET SSE stream were reconnect-looping on tools/list
— a bug introduced in v2.1.147. Custom API gateways were also receiving the user’s Anthropic OAuth credential instead of the gateway’s own token, breaking routing in any properly isolated setup. Both are fixed in 2.1.191.
On the security side, there’s a new sandbox.credentials
setting. It blocks credential reads from within sandboxes — meaning if you’re running Claude Code against an environment that has API keys or SSH credentials in scope, you can now prevent sandboxed code execution from reading them. A sensible addition for production setups where you’d rather not have generated code touching your credentials by default.
Update Now #
npm install -g @anthropic-ai/claude-code
The full changelog lists all 20 changes across versions 2.1.187 through 2.1.191. The five covered here are the ones that affect daily workflows. If you use GitHub releases to track updates, the full diff is there. For teams running Claude Code in enterprise setups, the agent policy and MCP fixes are the ones that matter most from a reliability and compliance standpoint.