Claude Code Digest — 2026-04-03 04:50:02
Version updates
Version 2.1.91 shipped on April 2, 2026. It introduces an MCP tool result persistence override via _meta["anthropic/maxResultSizeChars"], expanding limits up to 500,000 characters. Anthropic also added the disableSkillShellExecution setting to block inline shell execution within custom skills and slash commands.
What the docs reveal
Fixing the context truncation trap
Developers use MCP servers to pull large resources like database schemas or monorepo file trees. Until now, Claude Code aggressively truncated these mass outputs or redirected them to disk files. Writing to disk breaks the agent's continuous reasoning loop, forcing the model to issue secondary commands just to read what it asked for.
Version 2.1.91 introduces the 500K character override. Anthropic trades token economy for execution speed. By keeping massive tool outputs in active memory, the context window absorbs the sheer weight of your infrastructure. The agent sees the full schema immediately and writes the correct query on the first attempt.
Local context overrides team chaos
Team repositories previously forced all developers to share identical CLAUDE.md instructions. This caused friction between repository standards and individual developer workflows.
The new CLAUDE.local.md solves the clash. Claude Code appends this unversioned file after the central configuration during load. Your personal preferences always win the conflict resolution. Anthropic also added the /memory command to display the active prompt hierarchy. You no longer guess which rules dominate your session. Add CLAUDE.local.md to your global .gitignore today.
Autonomy demands sandboxing
Anthropic successfully solved prompt fatigue for basic operations. When autoAllowBashIfSandboxed is active, Claude runs Bash commands without asking your permission. Anthropic trusts the sandbox boundaries enough to let the agent experiment, fail, and retry silently.
However, they explicitly clarify an execution boundary: Computer Use features bypass the sandbox entirely. Because controlling the mouse and keyboard requires host-level access (restricted to macOS Pro/Max models via the CLI), you still pay the manual approval tax for every action.
Simultaneously, the new disableSkillShellExecution flag restricts custom skills from running arbitrary shell code. Anthropic isolates execution risk deliberately. The core agent acts autonomously inside the trusted sandbox, while unpredictable user-defined skills face stricter execution boundaries.
Plugins become system extensions
Plugin architecture matures significantly in this documentation update. Executables dropped into a plugin's bin/ directory inject directly into the agent's Bash PATH. Anthropic transformed plugins from simple prompt wrappers into native system extensions. When deployed using the --scope project flag, these plugin configurations write to .claude/settings.json, guaranteeing repository-wide availability for your entire team.
Deep linking prepares the ecosystem
Claude Code now registers the claude-cli:// protocol handler. You can launch sessions directly via claude-cli://open?q=... links, including multi-line prompts via URL encoding.
Anthropic built this to bridge the web browser and the terminal. Watch for issue trackers, continuous integration dashboards, and internal developer portals adding "Fix with Claude" buttons next to stack traces. Enterprise administrators managing locked-down environments get an escape hatch: the disableDeepLinkRegistration setting satisfies strict IT policies prohibiting custom URI handlers.