What Changed — The v2.1.90 Release
Anthropic just shipped Claude Code v2.1.90 with 18 changes. The headline is /powerup — an interactive tutorial system built directly into the CLI. But the real meat is in performance improvements and a critical fix to Auto Mode that developers have been hitting.
What It Means For You — Concrete Impact
1. Learn Features Faster with /powerup
Instead of reading docs, run:
claude code /powerup
You'll get animated, interactive demos of features like --resume, MCP servers, and tool permissions. This is Anthropic's answer to the learning curve — especially useful if you're new to the Model Context Protocol architecture that Claude Code is built on.
2. Auto Mode Now Actually Listens
The fix for "auto mode not respecting explicit user boundaries" is huge. Previously, if you said "don't push to main" or "wait for tests before deploying," Auto Mode might ignore you if the action was otherwise allowed. Now it respects those boundaries. This follows recent community reports about agents executing destructive commands — a trend we've seen in our coverage of autonomous agent incidents.
3. Performance Wins You'll Feel
- MCP sessions are faster: Eliminated
JSON.stringifyof MCP tool schemas on every cache lookup - Large responses stream better: SSE transport now handles large frames in linear time (was quadratic)
- Long conversations don't slow down: SDK sessions with long transcripts no longer degrade quadratically
/resumeloads faster: Project sessions load in parallel if you have many projects
4. --resume Gets Smarter
The --resume picker no longer shows sessions created by claude -p or SDK invocations — cleaner interface. More importantly, they fixed a regression (since v2.1.69) where --resume caused full prompt-cache misses for users with deferred tools, MCP servers, or custom agents.
5. Offline/Enterprise Friendliness
The new CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE environment variable helps in offline environments. If git pull fails when updating the plugin marketplace, it keeps the existing cache instead of breaking.
Try It Now — Commands and Config
Update Immediately
# Update Claude Code
npm update -g @anthropic-ai/claude-code
# Try the tutorials
claude code /powerup
# Check your version
claude code --version
Configure for Offline Use
If you work in air-gapped environments:
export CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE=true
# Now marketplace updates won't break if git pull fails
Test the Auto Mode Fix
Try a boundary that previously failed:
claude code --auto
# Then in conversation: "Refactor this module but don't commit the changes yet"
# Auto Mode should now respect "don't commit"
Protected Directory Update
.husky is now in protected directories (acceptEdits mode). If you use husky hooks, Claude won't accidentally edit them.
Other Notable Fixes
- Rate limit crash fixed: No more infinite loop when you hit usage limits
- Edit/Write with formatters: Fixed "File content has changed" when PostToolUse hooks reformat between edits
- PreToolUse hooks work correctly: Hooks that exit with code 2 now properly block tool calls
- PowerShell security hardened: Fixed several edge cases in permission checks
This release shows Anthropic focusing on both onboarding (with /powerup) and hardening the core experience — especially around the Auto Mode behavior that's critical for agentic workflows.
