Claude Code v2.1.90: Rate-Limit Crash Fix, Auto Mode Boundaries, and Performance Gains
Claude CodeClaude Code v2.1.90 delivers a set of reliability and performance improvements. A critical crash triggered by rate-limit responses during long agentic sessions has been resolved, preventing unexpected session terminations. Auto Mode now enforces clearer decision boundaries so the model is less likely to attempt actions outside its configured permission scope. The release also includes startup time reductions and memory optimizations that make the tool noticeably faster, particularly on large codebases.
Sources & Mentions
5 external resources covering this update
Claude Code v2.1.90 — rate limit crash fix and Auto Mode improvements
Hacker News
Claude Code 2.1.90 patch notes — rate limit crash fix
Claude Code v2.1.90 ships with rate-limit crash fix and performance improvements
X (Twitter)
Claude Code v2.1.90 Release Notes
GitHub
What's New in Claude Code v2.1.90
Dev.to
Claude Code v2.1.90: Reliability, Auto Mode, and Performance
Version 2.1.90 of Claude Code focuses on stability and responsiveness, addressing a crash that affected long agentic sessions and refining the boundaries of Auto Mode behavior.
Rate-Limit Crash Fix
A bug introduced in recent versions caused Claude Code to crash ungracefully when it received a rate-limit (HTTP 429) response from Anthropic's API during an active agentic session. Instead of pausing and retrying or surfacing a clean error message, the CLI would terminate abruptly — losing session context and any pending work.
The fix ensures that rate-limit responses are caught and handled properly. Claude Code now pauses, displays an informative message about the rate limit, and resumes automatically once the limit window resets. Sessions are no longer lost due to transient API throttling.
Auto Mode Boundary Improvements
Auto Mode — which allows Claude Code to make decisions about tool use and command execution without prompting the user at each step — has been updated to respect permission boundaries more consistently. In previous versions, certain edge cases caused the model to attempt actions that were outside the configured scope, leading to confusing permission denials mid-session.
The v2.1.90 update tightens the decision logic so that Auto Mode evaluates its configured permissions before attempting any action, and falls back to prompting the user when an action is ambiguous. This makes Auto Mode sessions more predictable and reduces unexpected interruptions.
Performance Gains
Two performance improvements ship in this release:
- Faster startup: CLI initialization time has been reduced, most noticeably on machines with large numbers of files in the working directory.
- Lower memory footprint: Memory usage during long sessions has been optimized, reducing the likelihood of slowdowns when working in large monorepos or running many sequential tool calls.
Summary of Changes
| Area | Change |
|---|---|
| Reliability | Rate-limit (429) crash fixed; graceful pause-and-retry instead |
| Auto Mode | Stricter boundary checks before tool execution |
| Performance | Faster startup, reduced memory usage in long sessions |