Claude Code v2.1.90 Release: New Interactive Lessons, Performance Improvements, and Bug Fixes

New Features and Environment Variables
Claude Code v2.1.90 adds /powerup — interactive lessons with animated demos that teach Claude Code features. The release also introduces the CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE environment variable, which keeps the existing marketplace cache when git pull fails, making it useful for offline environments. Additionally, .husky has been added to protected directories in acceptEdits mode.
Bug Fixes and UI Improvements
Several critical bugs have been addressed:
- Fixed an infinite loop where the rate-limit options dialog would repeatedly auto-open after hitting usage limits, eventually crashing sessions
- Fixed
--resumecausing a full prompt-cache miss on the first request for users with deferred tools, MCP servers, or custom agents (regression since v2.1.69) - Fixed Edit/Write failing with "File content has changed" when PostToolUse format-on-save hooks rewrite files between consecutive edits
- Fixed PreToolUse hooks that emit JSON to stdout and exit with code 2 not correctly blocking tool calls
- Fixed collapsed search/read summary badge appearing multiple times in fullscreen scrollback when CLAUDE.md files auto-load during tool calls
- Fixed auto mode not respecting explicit user boundaries ("don't push", "wait for X before Y") even when actions would otherwise be allowed
- Fixed click-to-expand hover text being nearly invisible on light terminal themes
- Fixed UI crash when malformed tool input reached the permission dialog
- Fixed headers disappearing when scrolling
/model,/config, and other selection screens
Security Hardening and Performance
PowerShell tool permission checks have been hardened with fixes for:
- Trailing
&background job bypass -ErrorAction Breakdebugger hang- Archive-extraction TOCTOU (Time-of-Check Time-of-Use) vulnerability
- Parse-fail fallback deny-rule degradation
Performance improvements include:
- Eliminated per-turn
JSON.stringifyof MCP tool schemas on cache-key lookup - SSE transport now handles large streamed frames in linear time (was quadratic)
- SDK sessions with long conversations no longer slow down quadratically on transcript writes
- Improved
/resumeall-projects view loads project sessions in parallel, improving load times for users with many projects
Changes and Removals
The --resume picker no longer shows sessions created by claude -p or SDK invocations. Get-DnsClientCache and ipconfig /displaydns have been removed from auto-allow lists for DNS cache privacy reasons.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Claude Opus 4.1 scores 17.75% on SWE-Bench Pro's private dataset, highlighting memorization vs. reasoning gap
Claude Opus 4.1 scored 80% on SWE-Bench Verified but dropped to 17.75% on SWE-Bench Pro's private dataset of 276 tasks from 18 proprietary startup codebases. Scale AI's analysis found models were navigating by memory rather than reasoning on familiar repositories.

Qwen3.5-122B on Blackwell SM120: fp8 KV Cache Corruption Issue and Performance Findings
Testing Qwen3.5-122B on 8x RTX PRO 6000 Blackwell hardware revealed that fp8_e4m3 KV cache silently produces corrupt output without errors, requiring bf16 KV cache instead. MTP optimization provided a 2.75x single-request speedup while DeltaNet constraints blocked other optimizations.

GitHub Copilot updates data usage policy for model training
GitHub will use interaction data from Copilot Free, Pro, and Pro+ users to train AI models starting April 24, 2026, unless users opt out. Copilot Business and Enterprise users are exempt from this change.

GitHub Copilot Code Review to Burn Actions Minutes Starting June 1, 2026
Starting June 1, 2026, GitHub Copilot code reviews on private repos will consume GitHub Actions minutes in addition to AI Credits. Public repos remain free.