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

Opus 4.6 Medium vs Low: Performance Differences and Pricing
Opus 4.6 medium costs approximately 50% more than the low version but addresses significant laziness issues found in the low-powered model. The medium version sits between low and high in performance benchmarks.

Berkeley Study: All AI Revision Prompts Drift Prose Toward Formality, Even "Preserve Voice"
New paper from Berkeley measures 300 personal narratives through Claude, ChatGPT, and Gemini under three prompt conditions. Every model and condition reduces contractions, first-person pronouns, and narrative closeness — the "preserve voice" prompt only reduces drift magnitude, not direction.

Coasty AI Agent Solves CAPTCHA Challenges Up to Level 6 Without Training
Coasty's Computer Using Agent (CUA) achieved 82% on the OSWorld benchmark, solving CAPTCHAs up to Level 6, browser popups, and cookie banners without specific training for 'I'm not a robot' challenges.

Meta's MCI Tool Captures Employee Interactions for AI Training
Meta is installing tracking software called Model Capability Initiative (MCI) on U.S. employee computers to capture mouse movements, keystrokes, clicks, and occasional screen snapshots for AI model training. The data aims to improve AI's ability to replicate human computer interactions like dropdown menu selection and keyboard shortcuts.