Claude Code 2.1.70 fixes proxy gateway support and model compatibility issues
Key Fixes
Claude Code 2.1.70 addresses several high-impact bugs that affected proxy gateway compatibility and model inference:
- Proxy Gateway Support: Fixed API 400 errors when using
ANTHROPIC_BASE_URLwith third-party gateways. Tool search now correctly detects proxy endpoints and disablestool_referenceblocks to ensure compatibility. - Custom Model Support: Resolved "This model does not support the effort parameter" errors when using custom Bedrock inference profiles or non-standard Claude model identifiers.
- Response Quality: Fixed empty model responses after
ToolSearchby improving how tool schemas are rendered in the prompt to prevent early model termination. - Prompt Caching: Fixed prompt-cache invalidation when an MCP server with instructions connects after the first turn.
Windows Stability Improvements
Multiple Windows-specific issues have been resolved:
- Fixed clipboard corruption for non-ASCII text (CJK, emoji) by switching to PowerShell
Set-Clipboard - Fixed extra VS Code windows opening at startup when running from the integrated terminal
- Fixed voice mode failing with "native audio module could not be loaded" on Windows native binary
- Improved SSH connection handling to prevent Enter key inserting newlines instead of submitting commands
UI & UX Enhancements
- Added spark icon in VS Code activity bar listing all Claude Code sessions, with sessions opening as full editors
- Added full markdown document view for plans with comment support for feedback
- Added native MCP server management dialog accessible via
/mcpcommand to enable/disable servers and manage OAuth without terminal switching - Fixed
/colorcommand to support reset options (/color default,/color gray,/color reset,/color none)
Performance & Platform Improvements
- Reduced prompt input re-renders by approximately 74%
- Reduced startup memory by ~426KB for users without custom CA certificates
- Optimized Remote Control
/pollrate from 1–2s to once per 10 minutes (300× server load reduction) - Fixed skill listing re-injection on every
--resume(~600 tokens saved per resume) - Improved image preservation during compaction for faster prompt cache reuse