Claude Code: 75+ Bug Fixes & Windows PowerShell Support

Claude Code just dropped its largest stability update to date: 75+ bug fixes spanning Windows PowerShell support, memory leaks, schema failures, and cross-platform rendering issues. For developers building AI automation pipelines or pushing Claude Code into enterprise team environments, this release removes the most common blockers in one go.

One fix alone stands out for its scale: a bug causing approximately 50% failure rate when using multiple StructuredOutput schemas (the format Claude uses to return structured, machine-readable data — think predictable JSON responses instead of free-form text) was quietly breaking half of all complex data workflows. It's fixed now.

Windows Developers Get PowerShell Support in Claude Code

The headline addition is the new PowerShell tool for Windows, available as an opt-in preview in this update. Previously, Windows users running Claude Code were forced to work with bash-style commands, which don't natively map to Windows shell environments. Now, Claude Code can generate and execute PowerShell commands directly.

The tool is smart about versions: it distinguishes between PowerShell 5.1 (the version that ships with Windows by default) and PowerShell 7+ (the cross-platform, open-source successor). That means Claude Code generates version-appropriate syntax instead of giving you commands that fail silently on your system.

# Enable PowerShell tool (opt-in preview)
# Open Claude Code and run /env to configure settings
/env

# Verify your PowerShell version inside Claude Code
$PSVersionTable.PSVersion

# Resume a paused headless session (new in this update)
claude --resume

This puts Claude Code more directly in competition with GitHub Copilot and Cursor IDE, both of which have stronger Windows footprints. Enterprise teams standardizing on Windows workstations were the most underserved group in previous Claude Code releases. The inclusion of version-aware PowerShell guidance suggests this isn't a token gesture — Anthropic is building for Windows as a first-class platform.

Memory Leaks, Zombie Processes, and the 50MB Crash

The engineering fixes reveal how Claude Code has been scaling into real enterprise use — and hitting hard limits. Three critical stability issues were addressed:

• Memory leak in long sessions: Large JSON inputs (JavaScript Object Notation — a structured data format used everywhere in software) were being retained as LRU cache keys (LRU = Least Recently Used — a memory-saving mechanism that's supposed to keep only recent items, not accumulate everything). In long-running sessions, this caused memory to balloon without limit. Now cleaned up automatically.
• 50MB session file crash: Any saved session history exceeding 50 megabytes would crash when Claude Code tried to trim old messages. Fixed — sessions can now grow and be managed without crashing.
• LSP server zombie state: LSP (Language Server Protocol) is the system powering code intelligence features like autocomplete, go-to-definition, and inline error highlighting. When the LSP server crashed, it would enter a zombie state — technically still running but not responding — and block all code assistance until you restarted the entire Claude Code session. Now it self-heals on the next request.

These aren't minor inconveniences. They're exactly the class of bugs that force engineering teams to pull back a tool from their workflow after one bad day on a large codebase.

The Windows Edit Bug That Was Silently Corrupting Files

One of the more insidious fixes: Claude Code's Edit and Write tools were doubling CRLF line endings on Windows. CRLF (Carriage Return + Line Feed — the Windows standard for ending a text line, inherited from typewriter mechanics) was being inserted twice on every edit. The result: extra blank lines appearing in config files, JSON, YAML, and source code every time Claude Code touched them. Silent. Consistent. Easy to miss until a deployment breaks.

A second silent data loss: CJK characters and emoji were being dropped at 4KB boundaries in prompt history. CJK (Chinese, Japanese, Korean) characters are encoded differently from Latin text, and at the 4-kilobyte chunk boundary they were being silently discarded — meaning developers working in those languages were losing parts of their conversation history with zero error messages or warnings.

Developer Quality-of-Life: Claude Code Fixes for AI Coding Workflows

Beyond the crash-level fixes, this update tightens the everyday coding experience across several areas:

• Thinking summaries disabled by default — extended reasoning displays no longer clutter interactive sessions. Re-enable with showThinkingSummaries: true in settings if you need visibility into Claude's reasoning chain.
• Flicker-free rendering via CLAUDE_CODE_NO_FLICKER=1 — solves display jitter when running Claude Code inside tmux (a terminal multiplexer — software that splits one terminal into multiple panes) on iTerm2.
• 5,000 character deep links — URLs that pre-populate Claude Code with a task or query now support up to 5,000 characters, with a scroll-to-review warning for long inputs.
• 5-second MCP connection timeout — MCP (Model Context Protocol — the system Claude uses to connect to external tools like browsers, databases, and code execution environments) connections that fail now time out in 5 seconds instead of hanging indefinitely.
• 45-second Bash task notification — if a shell command hangs, you get notified after 45 seconds rather than staring at a frozen screen wondering if Claude Code froze or is just slow.
• 75-minute idle nudge — after 75+ minutes without activity, Claude Code reminds you that a session is still open and waiting.

Developers using vibe coding workflows — where Claude Code handles implementation details while you stay focused on product direction — will notice fewer disruptions across every session with these fixes in place.

The Hook System Gets Conditional Logic

Hooks (automated scripts that run before or after Claude Code takes an action, like editing a file or executing a command) received two significant upgrades that matter for power users and team setups.

Conditional if field: You can now define hooks that only trigger under specific conditions — using the same syntax as Claude Code's permission rules. This reduces unnecessary process spawning (launching extra background programs) on every single action, which is especially important when hooks involve expensive operations like linting checks or test runners.

Defer permission decisions for headless sessions: When Claude Code runs in automated CI/CD pipelines (CI/CD = Continuous Integration/Continuous Deployment — the systems that automatically test and deploy code changes), it can now pause at a sensitive tool call and wait for a human review rather than auto-approving or failing. Resume the session with claude --resume after making your decision.

# Environment variables introduced in this update:

# Flicker-free rendering in tmux/iTerm2 setups
export CLAUDE_CODE_NO_FLICKER=1

# Custom stream idle timeout in milliseconds (default: 90,000 = 90 seconds)
export CLAUDE_STREAM_IDLE_TIMEOUT_MS=90000

# Non-blocking connector mode for CI environments
export MCP_CONNECTION_NONBLOCKING=true

What This Release Says About Where Claude Code Is Heading

A 75-fix release is unusual for any developer tool — most ship in smaller, more targeted batches. The breadth here reflects two things simultaneously: rapid adoption by teams pushing Claude Code into edge cases it wasn't originally stress-tested for, and a deliberate push toward enterprise-grade reliability.

The combination of Windows PowerShell support, improved prompt cache performance for Bedrock, Vertex, and Foundry (cloud platforms from Amazon, Google, and Microsoft respectively — used by larger organizations to run Claude at scale without calling Anthropic's API directly), and headless session controls with human-in-the-loop approval tells a clear story: Anthropic is competing for team-level and organization-level contracts, not just winning over individual developers.

If you're already on Claude Code, update now — especially if you're on Windows or run long sessions with large files. You can check what version you're on with claude --version. If you haven't tried Claude Code yet, the AI coding tools comparison guide breaks down how it compares to alternatives, and the latest AI automation news tracks every major update as they ship.

Related Content — Set Up Claude Code | AI Coding Guides | More AI Automation News