Claude Code Digest — 2026-03-28 16:11:24

What the docs reveal

The privacy cost of automated feedback The /feedback command now transmits your entire conversation history—including all contextual code—directly to Anthropic. The engineering motive here is clear: Anthropic needs full state data to debug complex agent hallucination loops. Users routinely submit vague bug reports. By capturing the underlying context window automatically, Anthropic developers can reproduce failures exactly.

However, this feature prioritizes bug hunting over strict code privacy. A developer casually typing /feedback now functions as an inadvertent data leak for proprietary source code. If your organization operates under strict NDAs or compliance frameworks, you must permanently isolate your environment. Security administrators should immediately deploy DISABLE_FEEDBACK_COMMAND=1 across their corporate workstations.

Enforcing simplicity via environment variables Anthropic introduced the CLAUDE_CODE_SIMPLE=1 environment variable to enforce a strict, minimalist operating mode. This setting bypasses auto-discovered configuration hooks, disables extensible plugins, and ignores historical memory logic. It restricts Claude strictly to a basic system prompt, standard Bash execution, and raw file editing.

Anthropic recognizes a core product truth: agentic behavior often ruins trivial tasks. Complex system prompts burn unnecessary tokens, increase response latency, and trigger the model to overthink straightforward edits. While this update simply maps to the existing --bare CLI flag, environment variable support enables persistent configuration. You can now lock down specific shell environments, Makefile targets, or automated scripts globally. Use this setting in CI/CD deployments where you require deterministic outcomes over creative reasoning.

Fighting macOS sandboxing Developers missing system notifications must now manually grant "Script Editor" permissions within Apple's System Settings. This new troubleshooting step highlights an engineering compromise. Anthropic avoids maintaining a native Swift compilation target for macOS notifications. Instead, Claude Code relies on standard osascript sub-processes to trigger system alerts.

Apple's recent security updates aggressively restrict these legacy inter-process calls. To establish a reliable feedback loop for long-running refactoring tasks, you must fight the operating system. Run a dummy command, let the OS block the notification, and immediately authorize "Script Editor" in your security preferences.

Restoring terminal keyboard flow You can now exit the image attachment dialog using the Down or Escape keys. Previous versions of the multimodal UI trapped developers inside the selection modal, forcing an awkward shift from keyboard to mouse. This fix reflects vital user feedback: terminal-native developers expect seamless keyboard navigation. The update restores your continuous execution loop.

View commit