Claude Code v2.1.169: Safe Mode, /cd Command, and Dozens of Bug Fixes

Claude Code v2.1.169 is out with several quality-of-life improvements and critical bug fixes for enterprise users. The headline additions are --safe-mode for troubleshooting and the /cd command for moving sessions without cache loss.
Safe Mode
The new --safe-mode flag (and CLAUDE_CODE_SAFE_MODE environment variable) starts Claude Code with all customizations disabled — CLAUDE.md, plugins, skills, hooks, and MCP servers. Useful for isolating whether a customization is causing issues.
/cd Command
Added /cd command to change the working directory of a session mid-session without breaking the prompt cache. No more starting a fresh session just to work in a different directory.
Enterprise & Admin Features
disableBundledSkillssetting andCLAUDE_CODE_DISABLE_BUNDLED_SKILLSenv var: hide bundled skills, workflows, and built-in slash commands from the model.- MCP policy enforcement fixed: Enterprise
allowedMcpServers/deniedMcpServerspolicies now apply consistently across reconnect, IDE-typed configs,--mcp-configservers, and first sessions after install. Also fixed slow cold starts for orgs without remote settings. - Remote-managed settings: Invalid entries now apply remaining valid policies and surface validation errors instead of silently dropping the whole payload.
- Background sessions now preserve flags like
--ide,--chrome,--bare,--remote-controlacross retire→wake, and respawn validation is hardened. - Background agents no longer ignore project-level
envvalues (e.g.,ANTHROPIC_MODEL) when dispatched onto a pre-warmed worker. - CLAUDE.md too long warning threshold now scales with the model's context window.
Performance Fixes
- Fixed a ~30-50ms UI stall at the start of each turn for macOS users logged in with claude.ai credentials.
- Fixed
claude -pbeing slow or hanging on Windows due to slash-command/skill scan (regression in v2.1.161). - Reduced CPU usage while responses stream and during spinner animations.
- Restored a default 5-minute idle timeout on Vertex/Foundry; set
API_FORCE_IDLE_TIMEOUT=0to opt out.
UI & UX Fixes
- Up/Down arrows now move through wrapped visual rows of a long input line before jumping to command history.
- Fixed Remote Control stuck on "reconnecting" when OAuth token refresh raced with session resume.
- Fixed Git Credential Manager popup on Windows at startup due to background git commands with no cached credentials.
- Fixed footer hints (e.g., "esc to interrupt") not showing with custom statusline.
- Fixed stale permission prompts reappearing after reattaching to a remote session whose worker died.
/workflowsnow opens immediately even while a turn is in progress.- Fixed
claude agents --jsonomitting blocked and just-dispatched sessions; added--allflag and newid/statefields. - Fixed agents view garbled frame on WSL in Windows Terminal after navigation.
Miscellaneous
- Improved
TaskCreatereliability: malformed inputs are repaired automatically, and validation errors include schema for unloaded tools. - Better error message when org disables API key authentication, with guidance based on active key source.
- Plugin
.in_usePID lock files now swept once per day to prevent accumulation. - Fixed untrusted project settings being able to set OTEL client-certificate paths without trust confirmation.
- Auto-updater on Windows stops retrying once
claude.exeis held by another process.
📖 Read the full source: GitHub Claude-Code
👀 See Also

ICML 2026 Desk-Rejects 2% of Papers for LLM Review Policy Violations
ICML 2026 rejected 497 papers (~2% of submissions) after detecting 795 reviews (~1% of all reviews) where reviewers violated explicit agreements not to use LLMs. The detection method involved watermarking PDFs with hidden LLM instructions.

Claude Consumer Terms Analysis: Data Retention, Liability Caps, and Service Termination
An analysis of Anthropic's Consumer Terms of Service reveals key details for $100/month Max plan subscribers: data training is on by default with 5-year retention for opted-in users, liability is capped at $600 maximum, and service can be terminated without refund for violations.

Claude AI Recovers 11-Year-Old Bitcoin Wallet Worth $400K by Finding Backup and Fixing Brute-Force Bug
A user recovered a 5 BTC wallet (worth ~$400K) after 11 years by feeding their entire college computer files into Claude. The AI found an older backup wallet and identified a bug in btcrecover's password combination logic, enabling successful decryption.

AI Detection Tools Push Students to Use AI Defensively, Study Finds
AI detection tools in education are causing students to intentionally write worse to avoid false positives, with some students turning to AI tools defensively to check if their own writing will be flagged.