Anthropic's Claude Code, its terminal-first coding agent, pushed version 2.1.221 with a cluster of changes aimed at developers running it daily rather than casual users. The headline addition is a VS Code Focus view, a chat-panel toggle that collapses the usual scroll of tool calls into an expandable per-turn summary with a small live indicator of what is currently running, so a long agent session does not bury the actual conversation under pages of file reads and shell output. Alongside that, sandboxed commands on Linux and WSL gained a credential-masking mode: a sandboxed process reading a secrets file now sees a sentinel copy instead of the real values, with the sandbox proxy substituting the genuine credential only when traffic actually leaves the sandbox, which narrows the blast radius if a sandboxed command tries to exfiltrate a secret. The release also adds a prompt-audit subcommand to the claude-api skill, meant to flag prompts and tool descriptions that were written for older model generations and may no longer reflect how newer models behave, a housekeeping problem that accumulates in any codebase that has been iterating on agent prompts for a while. Several fixes close real correctness gaps too: a zsh permission-check bypass that let hidden commands execute inside double-bracket regex conditionals without a permission prompt, and MCP servers configured via a config flag not being connected before the first turn in print mode, which had been causing the model to emit tool calls as literal text instead of actually invoking them. None of these are flashy, but they are the kind of reliability and security hardening that matters once a coding agent moves from novelty to something running unattended against real credentials and real codebases.