explainx.ainewsletter3.5k
TrendingNewsPathwaysSkills
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

corporate training

[email protected]

get started

Find your pathTake Free Evaluation

learn

pathways — start freeworkshopsbootcampscoursescertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcptoolsagentsllmsdesignsagi trackerranks

company

aboutvisionmissionteaminstructorscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportprivacytermsdata rightssubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

On this page

  • Highlights at a glance
  • /rewind: recover from /clear
  • Background agents: stops stick
  • Performance: 37% less CPU while streaming
  • Hooks: comma-separated matchers fixed
  • MCP improvements
  • Permissions, sandbox, and enterprise
  • Other fixes (selected)
  • New env vars and models (CLI surface)
  • How to update
  • Related reading
← Back to blog

explainx / blog

Claude Code 2.1.191: /rewind After /clear, Background Agent Fixes, MCP

Claude Code CLI 2.1.191 adds /rewind to resume before /clear, permanent background agent stops, 37% lower streaming CPU, MCP OAuth retries, hook matcher fixes, and 20 CLI changes — full changelog.

Jun 25, 2026·4 min read·Yash Thakker
Claude CodeChangelogDeveloper ToolsMCPAnthropic
go deep
Claude Code 2.1.191: /rewind After /clear, Background Agent Fixes, MCP

TL;DR: Claude Code 2.1.191 dropped June 25, 2026 (@ClaudeCodeLog / @marc_krenn). Headline features: /rewind resumes context from before /clear, background agent stops are permanent, ~37% lower CPU during streaming (100ms coalescing), comma-separated hook matchers fixed, and a pile of MCP reliability improvements. Twenty CLI changes total.


Highlights at a glance

ChangeWhy it matters
/rewind after /clearUndo an accidental clear without losing the thread
Background agents stay stoppedKills from tasks panel no longer resurrect
Streaming coalescing (100ms)~37% CPU reduction; smoother long sessions
Hook matcher fix"Bash,PowerShell" matchers actually fire now
MCP retries + OAuthTransient network errors backoff; headless OAuth skips browser popup
Sandbox network memory"Yes" on host allow persists for the session
Permissions UI fixRecently-denied approvals persist on close
Weekly digest3.5k readers

Catch up on AI

Curated AI updates on agents, skills, and MCP — delivered to your inbox. Unsubscribe anytime.


/rewind: recover from /clear

The biggest workflow fix in 2.1.191:

Added /rewind support for resuming a conversation from before /clear was run.

Before: /clear (aliases /reset, /new) wiped context. If you cleared too aggressively, you relied on /resume for a saved session — not the in-flight thread.

After: /rewind rolls back to a checkpoint before the clear, restoring prior context so the session continues.

This sits alongside existing rollback tools:

MechanismTriggerScope
ESC × 2KeyboardCode and/or conversation checkpoints
/rewindSlash commandCheckpoints + pre-/clear restore (2.1.191+)
/resumeSlash commandSaved named sessions
/branchSlash commandFork conversation at a point

Deep dive: Claude Code commands reference · Checkpoint restore guide.


Background agents: stops stick

Fix: Background agents stopped from the tasks panel no longer restart — stopping is permanent.

If you run parallel work via /fork, /agents, or /tasks, a stop that silently undid itself was dangerous (duplicate edits, double bash). 2.1.191 closes that hole.

Related: claude agents no longer sends builtin slash commands like /usage to background sessions as raw prompt text — you get a hint instead.


Performance: 37% less CPU while streaming

Two related improvements:

  1. Coalesce text updates to 100ms → ~37% CPU reduction during streaming responses
  2. Reduced long-session memory growth from terminal output cache
  3. Scroll fix — reading earlier output no longer jumps you to the bottom mid-stream

For developers running sound and traffic-light hooks on long agent sessions, lower CPU and stable scroll make unattended monitoring less painful.


Hooks: comma-separated matchers fixed

Bug: Hooks with matchers like "Bash,PowerShell" silently never fired.

Fix in 2.1.191: Multi-tool matchers parse correctly.

If you copied hook recipes from forums and wondered why nothing ran, update first, then audit ~/.claude/settings.json:

json
{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash,PowerShell",
        "hooks": [{ "type": "command", "command": "echo triggered" }]
      }
    ]
  }
}

MCP improvements

AreaChange
Capability discoverytools/list, prompts/list, resources/list retry transient errors with backoff
OAuthDiscovery/token requests retry once; headless skips browser popup → paste URL prompt
ErrorsHTTP 404 shows URL + pointer to MCP config
ReliabilityFewer flaky server dropouts on slow networks

Connect tools: Claude Code MCP guide.


Permissions, sandbox, and enterprise

  • /permissions Recently-denied tab — approving a denial now persists on close (was silently discarded)
  • Sandbox network dialog — hosts allowed with "Yes" are remembered for the session
  • /voice — when disabled by org policy, message explains the restriction (not generic "not available")
  • Managed settings — forceRemoteSettingsRefresh works via MDM/file policy; fetch sends Cache-Control: no-cache

Other fixes (selected)

  • /login URL — no longer truncated when wrapped in Windows Terminal
  • Cmd+click links — fixed in Ghostty over ssh/tmux fullscreen
  • Agent panel — no longer jumps one row when scrolling roster past overflow cap
  • Image placeholders — job rows show [Image #N] not full filesystem paths
  • Welcome splash — no longer overflows default 80×24 macOS Terminal
  • Vim / search — hints how to reach slash commands from NORMAL mode

New env vars and models (CLI surface)

Per @ClaudeCodeLog extended thread, 2.1.191 also adds:

Environment variables: BUF_BIGINT_DISABLE, CLAUDE_CODE_DD_ERROR_TRACKING_FLUSH_INTERVAL_MS, NODE_USE_ENV_PROXY, MINIMATCH_TESTING_PLATFORM

Model aliases (internal/flags): claude-bg-, claude-code-error-tracking, o5-qn, o5-u:, opus-on-pro-near-limit

Treat model slug additions as feature-flag plumbing unless documented on code.claude.com.


How to update

bash
# If installed via npm/homebrew — follow your usual upgrade path
claude --version   # confirm 2.1.191+

# Or use Claude Code's built-in update mechanism per official docs

Changelog source: @ClaudeCodeLog (unofficial, tolerated bot) · full history in linked GitHub repo.


Related reading

  • screenpipe — desktop work memory via MCP (Jul 2026)
  • Claude Code commands — full reference
  • /rewind on slash commands registry
  • Restore checkpoints (ESC × 2)
  • Hooks automation guide
  • Sound + traffic-light approval alerts
  • /goal long-running agents

Changelog items sourced from @ClaudeCodeLog posts on June 25, 2026. Verify behavior on your installed version — patch details may vary by platform.

Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

Related posts

Jul 16, 2026

Claude Code Artifacts + MCP: Live Dashboards With Viewer-Scoped Auth

Anthropic's ClaudeDevs account announced artifact MCP connectors: build a dashboard once, and each viewer pulls live data through their own connectors. explainx.ai explains viewer-scoped auth, plan limits, admin toggles, and how this extends Thariq's thick-artifacts framework.

Aug 8, 2026

Auto Mode Becomes the Default in Claude Code for Pro, Max, and Team

Anthropic is flipping Claude Code's default permission mode to "auto" for Pro, Max, and Team plans starting August 14, 2026 — replacing manual approval prompts with a classifier that screens every tool call. The controlled study behind the switch found humans catch a planted dangerous command 13.6% of the time; auto mode catches it 89% of the time.

Aug 8, 2026

Claude Code Cross-Session Messaging: Sessions Can Now Talk to Each Other

Anthropic shipped cross-session messaging for Claude Code: one session can now send a summary, a finding, or a status update straight to another session you're running, and get a reply back. This is what changes for anyone juggling multiple terminals or worktrees.