Claude Code v2.1.147: Pinned Sessions, /code-review, and Dozens of Fixes

Claude Code v2.1.147 (released May 21) is a substantial maintenance release with one new feature, one renamed command, and a long tail of fixes covering Windows, PowerShell, MCP servers, and more.
Pinned Background Sessions
You can now pin background sessions (Ctrl+T in agent mode). Pinned sessions stay alive when idle, are restarted in-place to apply Claude Code updates, and are only killed under memory pressure after unpinned sessions. This is useful for long-running agent workspaces you don't want accidentally evicted.
/simplify Becomes /code-review
The old /simplify slash command (which tried to cleanup and fix code) has been removed. It's replaced by /code-review, which reports correctness bugs at a chosen effort level. Run /code-review high for deep analysis, or /code-review low for quick checks. Pass --comment to post findings as inline GitHub PR comments.
Auto-Updater Improvements
The auto-updater now retries transient network failures, reports specific error categories and OS error codes on failure, and shows the current version when an update fails. This should reduce silent update failures and give you actionable error messages.
Performance & Behavior
- Improved diff rendering performance for large file edits.
- Prompt history no longer records consecutive duplicate entries — pressing arrow-up and re-submitting the same prompt won't add another copy.
- Pasted text is now delivered correctly instead of the placeholder
[Pasted text #N]. - The CLI's auto-mode no longer suppresses
AskUserQuestionwhen the user or a skill relies on it — answers are now treated as intent signal.
Key Fixes
- Enterprise login:
forceLoginOrgUUIDandforceLoginMethodsettings now work against third-party-provider and API-key sessions. - Shell snapshot: User functions starting with a single underscore are no longer dropped, fixing broken aliases that reference them.
- PowerShell: The tool no longer drops output for commands relying on the default formatter. The 'Yes, and don't ask again' dialog now writes matching rules on subsequent runs. Fixes exit code 1 when
pwshis installed via winget or the Microsoft Store. - MCP servers: Paginating MCP servers no longer drops resources, templates, and prompts past page 1.
- Windows: Removes background-job worktrees without following NTFS junctions into the main repo. Fixes full-screen strobing in attached background sessions on Windows Terminal while streaming. Fixes stale/doubled rows in agent view list with wide (CJK) characters.
- Plugin agents: Multiple
Agent(...)types intools:frontmatter are now all registered, not just the last one. - Hooks: Fixes
ifconditions likePowerShell(git push*)matching correctly; previously onlyPowerShell(*)worked. - Background sessions:
/backgroundno longer rejects sessions whose only input was a skill or custom slash command. - Streaming: Fixes an uncaught exception at the end of streaming sessions when using the Agent SDK.
- GNOME Terminal: Right-click and middle-click paste now insert text.
- Subagent model:
CLAUDE_CODE_SUBAGENT_MODELnow applies to teammate processes spawned by agent teams. - Slash commands: Unknown slash commands in headless/SDK mode now show an error instead of silently doing nothing. Commands followed by tab or newline are no longer treated as unknown.
- Misc:
/helprenders correctly on small terminals./effortstarts at the correct level./themedialog responds toEsc. Stripped images no longer cause the model to repeatedly re-read missing media.&in!command output displays properly for copy-pasting URLs. Plugin component counts are no longer doubled. Backgrounded sessions don't re-prompt for already-granted tool permissions.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Diagnosing Operational Drift and Task Amnesia in OpenClaw with Gemini 2.5 Flash on Proxmox
OpenClaw users report issues with persistent workflows on a Proxmox VM, citing operational drift and task amnesia. Despite stable performance in one-off tasks, the Gemini 2.5 Flash model struggles with automation and memory in this setup.

Auditing API Logs Reveals AI Agents Waste Tokens on Context Window Bloat
A Reddit audit finds Claude agents burn 30k+ tokens on file exploration and verbose logs before writing code, causing architectural decay as context fills with noise.

GPT 5.5 vs Claude: A Developer's Refactoring Battle Report
A developer used GPT 5.5 to plan and Claude to code a massive 36k-line C refactoring. GPT 5.5 impressed with clear plans but burned through 85% of usage in 2 hours on the $30 plan.

Anthropic to stream live briefing on Enterprise Agents today
Anthropic will stream a live virtual briefing today, February 24, 2026, focused on Enterprise Agents. The event is accessible via their website.