Claude Code v2.1.90 Released — /powerup Interactive Lessons, Critical Bug Fixes, Performance Wins
Claude Code 2.1.90 landed on April 1 (UTC), bringing one of the most developer-friendly additions in recent memory: /powerup, a new interactive lesson system with animated in-terminal demos that walks you through Claude Code's features as you use them. If you've ever wondered what you're missing in a tool this deep, this is Anthropic's answer — no docs tab required.
Under the hood, 2.1.90 closes several significant pain points. The --resume prompt-cache miss regression (introduced in v2.1.69) is finally fixed — every session start for users with deferred tools, MCP servers, or custom agents was burning unnecessary tokens since that release. The StructuredOutput schema cache bug that caused a roughly 50% failure rate when using multiple schemas in the same workflow is gone. A quadratic performance regression in SSE transport and transcript writes (long sessions slowing down measurably over time) has been corrected. And a recurring "File content has changed" error triggered by PostToolUse format-on-save hooks between consecutive edits is resolved.
Windows gets unusually thorough attention this release: PowerShell security hardening patches four separate edge cases including a trailing & background job bypass, a -ErrorAction Break debugger hang, an archive-extraction TOCTOU, and a parse-fail fallback deny-rule degradation. The new CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE env var preserves marketplace cache when git pull fails — a meaningful win for offline and air-gapped deployments. PreToolUse hooks that emit JSON and exit with code 2 now correctly block tool calls as documented. Update via npm install -g @anthropic-ai/claude-code@latest.