Pick
Claude Code Update: Memory Warnings, MCP Timeout Controls, and Stability Improvements
·2026.09.19 06:00
Key point
The latest Claude Code update introduces memory usage warnings, a new environment variable for controlling MCP server connection wait times, and various bug fixes and stability improvements.
Details
Key Feature Changes
The latest update to Claude Code (within versions 2.1.274–2.1.277) includes the following features:
- Memory Management: Added visible warnings when memory usage reaches a threshold, along with steps to release memory or safely restart.
- MCP Timeout Control: Added the
CLAUDE_CODE_MCP_STARTUP_WAIT_MSenvironment variable to limit the time the first non-interactive turn waits for MCP server connections (0means no wait). - Database Connection Timeout: Added
store.connect_timeout_secondsto the Claude App Gateway settings to adjust PostgreSQL connection timeouts, with a default value of 5 seconds.
Stability and Bug Fixes
- Sessions and Transcripts: Fixed an issue where sessions would retry indefinitely due to "unexpected tool_use_id" 400 errors, and corrupted transcripts are now self-healed when possible.
- MCP Connections: Resolved an issue where MCP servers configured as
httpfailed to connect when using legacy HTTP+SSE, responding with 422 or other 4xx errors on the first request. Also fixed a bug where Streamable HTTP MCP tool calls timed out after approximately 5 minutes regardless of per-servertimeoutsettings. - Permissions and Security: Fixed an issue where Bash commands that the permission checker could not fully analyze bypassed prompts under the
permissions.blockReadsOutsideWorkingDirectoriessetting. - Other: Improved 401/403 error messages for Bedrock, Vertex, and Foundry, and fixed an issue where previous thinking was deleted during
/login, causing prompt cache rewrites, among other stability improvements.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.