Claude Code Digest — 2026-03-20 16:18:22
What the docs reveal
Anthropic reengineered context loading to solve the notorious token bloat problem in large codebases. Claude Code now parses CLAUDE.md files incrementally. Instructions spanning from your repository root down to the current working directory load explicitly at launch. Claude ignores subdirectory context until you navigate into those paths.
This dynamic mapping demands deliberate architecture. Developers working in corporate monorepos can now use the claudeMdExcludes parameter to blacklist documentation from adjacent teams, preserving the token window for relevant code. However, the documentation currently contradicts itself regarding instruction conflicts. One update claims Claude prioritizes the most specific instructions available; another file warns that conflict resolution remains arbitrary. You must prune contradictory directives manually until Anthropic stabilizes this parsing logic.
Architectural shifts continue in the agent framework. Anthropic is migrating Claude Code away from monolithic prompt executions toward isolated, sandboxed micro-agents. You can instruct local skills to execute within a separate subagent using context: 'fork' and specify the exact agent runtime via the agent field.
To secure these autonomous processes, the CLI implements strict tool filtering. You restrict a subagent using the tools allowlist and the disallowedTools denylist. Denylists always execute first. Any tool appearing in both definitions drops from the final permissions matrix. You must audit your subagent configurations to enforce the principle of least privilege. Furthermore, a new --channels research preview flag lets Claude listen for Model Context Protocol (MCP) server notifications asynchronously. Anthropic intends to turn Claude Code into an event-driven daemon rather than a purely reactive terminal prompt.
Corporate security mandates drove the new server-managed settings pipeline. Organizations can now enforce telemetry and security rules globally via JSON. Administrators gain access to restricted overrides like disableBypassPermissionsMode. Furthermore, custom PostToolUse lifecycle hooks guarantee rigid intervention. If an administrator configures this hook for shell executions, Claude halts and forces an explicit user approval dialog before running the bash command. Anthropic willingly sacrifices developer velocity to secure enterprise compliance.
Plugin distribution shed unnecessary friction. Teams previously required a hosted repository to distribute internal tools. You can now define localized plugin bundles directly inside settings.json using the source: 'settings' directive. For broad distribution, Anthropic formalized the default syntax for authenticated marketplace installations (plugin:<name>@claude-plugins-official).
Finally, Anthropic eliminated rate limit opacity. Exhausting your context window mid-refactor breaks developer flow. The CLI status line now queries Pro and Max subscriber data directly, extracting used_percentage and resets_at metrics after your initial API response. Claude formats both 5-hour and 7-day usage thresholds. You no longer have to guess your remaining message allowance.