Anthropic pushed Claude Code to v2.1.234 on August 17, while also shipping a new research-preview feature into the CLI and Desktop. This is not the usual "bug fix plus feature add" — it is a nearly symmetric dual-track update: one track welds shut the last gap from last week's NTLM credential leak, and the other officially brings Claude Design's artboard workflow into the coding environment. One side is security, the other boundary expansion; one tightens, the other lets go.
The Last Mile of the NTLM Patch
The headline fix in v2.1.234 is rejection handling for Windows NT-namespace (\??) paths. Last week's 2.1.233 patched the Windows credential-leak NTLM vector, but several pre-approved file-access paths — remote file reads, session resume, CLAUDE.md includes, workflow scripts, and file uploads — could still bypass approval via NT-namespace paths and touch the leaked credentials. v2.1.234 closes all five paths at once, so even an attacker holding leaked NTLM credentials cannot use path tricks to read files they shouldn't. This is the "last mile" of the NTLM vulnerability patch series, and the first time Claude Code approaches a complete credential isolation model on Windows.
Another subtle but engineering-heavy change in the same version: the built-in claude-api skill's context cost was cut from ~200k+ tokens to ~25k. The technique is "on-demand loading of reference docs" — at skill startup the full API documentation is no longer stuffed into context; instead reference docs are fetched externally as needed. The ~8x context reduction means that within a 200K context window, the claude-api skill is no longer a budget-hogging beast, and developers can mount multiple skills simultaneously without squeezing out readable length for core tasks.
/design Brings the Design Workflow into the Coding Environment
The /design skill (research preview), launched the same day, gives Claude Code CLI and Desktop Claude Design's artboard capabilities. Running /design, Claude generates editable artboards as visual UI artifacts — the developer picks one, fine-tunes it on the canvas, then has Claude implement it. This compresses the "design → implement" feedback loop from "switch to Design tool → screenshot → write text description → write code" down to "/design → pick one → tweak → implement." For front-end engineers, UI/UX designers, and full-stack developers, it means Claude Code no longer produces only text but directly outputs visual artifacts editable on an artboard.
Two More Hidden Engineering Changes
- Auto-resume on usage limit reset — when a user hits the claude.ai usage cap, the session pauses; once the limit lifts, Claude Code can now automatically resume the previous work. No restarting sessions or re-pasting context. A real experience improvement for developers running long coding tasks, especially agents doing multi-step reasoning that consumes large context budgets.
- GitLab merge request badges — when a repo has a GitLab remote configured and an authenticated glab CLI is present locally, the footer and statusline display MR status. Another step in Claude Code's support for collaboration platforms beyond GitHub.
What the Last Mile of the NTLM Patch Means
Claude Code 2.1.234 advances the Windows credential-leak vector from "main path blocked, sub-paths bypassable" to "sub-paths also blocked." For engineering teams deploying Claude Code on Windows workstations, this means the NTLM credential isolation model is complete for the first time — even with leaked credentials, an attacker cannot reach any file-access surface that exploits them. It's Anthropic paying down "prompt injection defense" and "credential isolation" as the same engineering debt, continuing the story that began when Auto mode was enabled by default on 8-14.
What It Changes
On this day Claude Code demonstrated two things at once: tightening defenses (all NTLM sub-paths closed) and expanding boundaries (/design bringing artboards into the CLI). These two tracks look contradictory but point to the same conclusion — the moat of AI coding tools is no longer "how many lines of code it can write," but "how much more it can do within an enterprise-grade security budget." With the NTLM patch and design artboard capability delivered in the same release, Claude Code has switched tracks from "coding assistant" to "coding platform."