Claude Code: 75+ Bug Fixes, PowerShell & Enterprise Features

If you rely on Claude Code for AI automation workflows — writing, reviewing, or shipping code — the latest update just cleared a long list of frustrations. Anthropic pushed 75+ distinct fixes in a single changelog burst — covering memory leaks, startup freezes, broken Windows behavior, and a brand-new PowerShell tool that Windows developers have been waiting for.

The breadth of this release signals something beyond routine maintenance: Claude Code is now handling serious production workloads — from vibe coding sessions to full enterprise deployments — and Anthropic is fixing the rough edges that only appear when real teams use a tool every single day.

The Scale of This Claude Code Update

This isn't a minor patch. The changelog covers over 75 numbered improvements across performance, stability, terminal compatibility, and enterprise policy controls. Some headline numbers:

• 16× faster stats screenshots (Ctrl+S in /stats) — now instant across all builds
• 30ms improvement in interactive startup from running setup() in parallel
• 1–8 second UI freeze on startup (caused by voice input initialization) — eliminated
• 5-second graceful timeout for diffing very large files — no more infinite hangs
• 90-second streaming idle watchdog — now configurable via CLAUDE_STREAM_IDLE_TIMEOUT_MS
• 2KB cap on MCP tool descriptions — prevents context bloat that was slowing everyone down

The fixes span every major operating system. But the most significant shift: Windows is no longer a second-class platform for Claude Code users.

The Fixes That Actually Matter Day-to-Day

Memory and Performance in Long Sessions

One of the highest-impact fixes: memory growth in long sessions. When you're working on a complex feature for hours, Claude Code's markdown and highlight render caches (the systems that color-code your code and format text on-screen) were holding onto full session content indefinitely. This caused gradual slowdowns and, on long enough sessions, crashes.

That's now fixed. Related: the /feedback command was triggering out-of-memory crashes on very long sessions with large transcripts — also resolved. The scroll engine for long conversations was replaced entirely, switching from WASM (WebAssembly, a compiled binary format that runs fast but is harder to debug) to TypeScript (the standard web programming language) for better reliability. Wheel scroll and click-to-select now behave correctly at the bottom of long conversations.

Windows: PowerShell Tool + Corruption Fixes

Windows users had two serious problems. First, unnecessary config disk writes were firing on every skill invocation — causing performance slowdowns and, on slower disks or antivirus-active machines, occasional file corruption. Second, there was no native Windows shell: users were stuck with Bash emulation (a compatibility layer that translates Linux-style commands, which breaks on purely Windows file paths).

Both are now addressed. The disk write issue is fixed. And Anthropic has shipped a PowerShell tool for Windows as an opt-in preview — you can run PowerShell commands directly inside Claude Code without workarounds. Enable it through the settings UI:

# Check your Claude Code version
claude --version

# View the full changelog
# Visit: https://code.claude.com/docs/en/changelog

# Enable PowerShell tool (Windows only - opt-in preview)
# Go to: Settings → Tools → Enable PowerShell Preview

Authentication and Security Fixes

Two subtle but important security issues were patched in the same release:

• OAuth login URL copying was truncating to roughly 20 characters instead of the full URL — you'd paste an invalid link with no obvious reason why login failed
• Masked input for OAuth token paste (the hidden field where you enter your API key) was leaking the start of your token when text wrapped on narrow terminals — a potential credential exposure fixed in this update

Enterprise AI Automation Features in the Changelog

The most strategically important additions aren't bug fixes — they're new infrastructure aimed directly at team and enterprise deployments:

• allowedChannelPlugins managed setting — IT admins can now enforce an allowlist of approved channel plugins across an entire organization, replacing individual user discretion with top-down policy
• X-Claude-Code-Session-Id header — added to API requests so corporate proxy servers (the network intermediaries companies use to monitor and log traffic) can aggregate session-level data without parsing request bodies
• TaskCreated hook — fires automatically when a task is created via TaskCreate, enabling workflow automation and audit logging for compliance teams
• WorktreeCreate hook with HTTP support — teams using custom repository workflows (like monorepos with specialized branch management) now have native hook support with worktreePath in the response JSON
• ANTHROPIC_DEFAULT_*_MODEL_SUPPORTS env variables — third-party platforms integrating Claude Code can now declare their own model capability profiles, removing the need for hardcoded Anthropic assumptions

The managed settings framework — a policy enforcement system that lets administrators control tool behavior across their entire organization — is the clearest signal yet that Anthropic is competing for enterprise software budgets, not just developer side projects. Check the Claude Code setup guide if you're configuring Claude Code for a team.

MCP, AI Automation & Plugin Improvements

MCP (Model Context Protocol — the open standard that lets Claude Code connect to external tools, databases, and services) received several important fixes and new guardrails:

• The 2KB cap on MCP tool descriptions and server instructions prevents context bloat (when documentation takes up so much space it crowds out the actual work you're trying to do)
• A roughly 3-second MCP config fetch wait on startup was eliminated — Claude Code now loads noticeably faster for users with multiple MCP servers configured
• Official marketplace plugin script failures on macOS and Linux (broken since v2.1.83) are now fixed
• The --bare mode (a minimal UI designed for use in scripts and automation pipelines) was silently dropping MCP tools and discarding queued messages — now corrected
• A 250-character skill description cap in /skills listing now reduces context usage automatically

For developers building MCP integrations: trim your tool descriptions now. Anything over 2KB gets cut off regardless, and verbose documentation was degrading performance for all users sharing the same server. Browse our AI automation and coding guides for best practices on MCP server setup.

What to Know Before You Update

A few honest limitations worth flagging before you upgrade:

• The PowerShell tool is opt-in preview only — not default, not stable. Don't build critical workflows around it until it graduates from preview status.
• Large file diffing now has a 5-second timeout as a graceful fallback — the comparison algorithm doesn't scale past certain file sizes, so it now gives up and falls back instead of hanging indefinitely.
• Voice push-to-talk character leaking is fixed, but audio hardware dependencies remain — voice mode still doesn't work reliably in headless environments (servers or Docker containers without audio hardware attached).
• The 75-minute idle threshold that surfaces a /clear reminder is a cost-management workaround for token re-caching overhead (the processing cost of re-reading your full conversation history on every new message) — the scaling problem is managed, not eliminated.

Update Claude Code now through your standard channel — VS Code marketplace, npm, or direct download. Windows users: head to Settings → Tools to enable the PowerShell preview. Enterprise admins: the allowedChannelPlugins managed setting gives you organizational control that wasn't previously available — worth exploring immediately if you're managing a team deployment.

Related Content — Get Started | Guides | More News