Claude Code v2.1.221 brings two operational features aimed at frictionless daily use. The first is the Focus view in VSCode: a toggle (Ctrl+Alt+F or the “Claude Code: Toggle Focus view” command) that hides tool activity behind an expandable summary for each turn, with a live indicator for the running tool. Instead of scrolling through dozens of lines of command output, you see a single line telling you what the agent is doing, and you expand it only when needed.

The second is credential masking for the sandbox on Linux and WSL. When a sandboxed command reads a credential file, it receives a sentinel copy (the entire file or only portions captured by an extract regex). The sandbox proxy replaces the real value on exit, when the command actually needs to use it. On macOS, where masking isn’t yet available, the system falls back to deny.

For anyone using Claude Code daily, the Focus view removes the visual noise that makes it hard to follow what the agent is doing in long sessions. Credential masking solves a concrete problem: keeping keys out of the context that the model sees, even when the agent needs to launch commands that use them.

The release also closes a Bash permission bypass with zsh (commands hidden in [[ ]] regex), fixes for Windows paths with quotes, and a series of stability fixes for MCP, thinking toggle, and interrupted sessions.

As we covered on July 23 with v2.1.218, Claude Code is chasing daily use: each release removes a friction point or closes a gap, and the tool increasingly behaves like a work companion.

If you want to try it: in VSCode, Ctrl+Alt+F during a session with active tools.

In detail

Claude Code’s release series this summer tells the story of a product changing pace. Ten releases in three weeks, from mid-July to now, each one closing concrete bugs or adding operational controls. The pattern is clear: Anthropic is moving Claude Code from demo tool to daily work companion.

Focus view: less noise, same information. When Claude Code works, tool output (file reads, bash commands, MCP calls) can fill dozens of lines per turn. In a half-hour session, following the thread of what the agent did becomes exhausting. The Focus view compacts that output into an expandable summary for each turn, with an indicator showing which tool is running in real time. Toggle it with Ctrl+Alt+F or the “Claude Code: Toggle Focus view” command from the chat menu. It’s a VSCode feature, not the standalone CLI.

The comparison is to a build system’s verbosity level: when you compile, you don’t want to see every compiler command, you want to know where it is and if it’s done. If something breaks, you expand and look. Focus view brings the same idea to agent interaction.

Sandbox credential masking: keys don’t enter the context. The problem is concrete: if an agent needs to run a command using credentials (for example an API call with a token), those credentials end up in the session context, where the model can read and potentially repeat them. Credential masking solves this on Linux and WSL with a two-phase mechanism.

When a sandboxed command tries to read a credential file, the system gives it a sentinel copy: a placeholder. It can mask the entire file or only portions captured by a configured extract regex. When the command actually needs to use the real value (egress), the sandbox proxy replaces the placeholder with the real value on output. The model never sees the real credential.

On macOS the mechanism isn’t yet available and the system falls back to deny: if a sandboxed command tries to read credentials, it’s blocked. It’s a cruder but safe fallback.

Security fixes. The release closes a Bash tool permission bypass: zsh could execute commands hidden inside [[ ]] regex conditions without asking approval. Now those commands ask for permission. On Windows, PowerShell mishandled paths with quotes: those now ask for approval instead of passing silently.

A note for those deploying agents: the zsh bypass reminds us why agent security perimeters should be thought in layers, not as a single wall. The playbook on defending against prompt injection starts here.

What remains open. Credential masking is Linux/WSL only: macOS users only get deny. Focus view is VSCode only, not the CLI. And like every Claude Code release, stability fixes for MCP, interrupted sessions, and thinking toggle suggest the tool is still in active stabilization, not maintenance.