Claude Code Digest — 2026-03-27 04:51:20
Version updates
Anthropic released Version 2.1.85 to solve friction in enterprise configurations. Model Context Protocol (MCP) authentication previously forced developers to write duplicate helper scripts for each server. The new CLAUDE_CODE_MCP_SERVER_NAME and CLAUDE_CODE_MCP_SERVER_URL environment variables eliminate this redundancy. A single script now reads the server identity and routes the correct authentication headers. This change implies developers are scaling MCP aggressively across internal microservices.
Hook execution gains surgical precision. Developers can now attach an if field using permission rule syntax, limiting hook triggers to specific operations like Bash(git *). Broad hooks previously triggered constant permission interruptions, training developers to ignore security prompts. Filtering execution restores the value of these warnings. You should audit your hooks and apply strict filters to preserve your attention span.
What the docs reveal
Anthropic quietly deprecated the commands/ directory. Skills and commands now share the same /name invocation syntax, but documentation crowns skills/ as the standard. Commands encouraged brittle, single-file scripts. Skills introduce a robust architecture that bundles instructions with supporting reference files. Models write better code when you provide structural context alongside the prompt. Migrate your isolated scripts into skill bundles.
Context management shifts from universal to local. Developers can now restrict guidelines in .claude/rules/ to specific file paths using paths: frontmatter. Injecting universal project rules into every prompt exhausts the context window. It also degrades model reasoning. Scoping rules ensures Claude ignores your Python linting directives while editing React components. Update your rule library to target specific paths immediately.
Role definitions separate human shortcuts from model tools. New skill frontmatter allows developers to gate function access. Setting disable-model-invocation reserves functional macros for humans, preventing Claude from initiating unprompted workflows. Conversely, user-invocable: false hides backend validation macros from human autocomplete menus. Workflows fail when engineers try to build universal tools for both agents and humans. Build distinct toolchains for each audience. Parameters are now fully supported via $ARGUMENTS or positional flags, bringing shell-like execution power directly into these native workflows.
Telemetry captures raw agent capabilities. Documentation for OpenTelemetry (OTLP) now clarifies that enabling OTEL_LOG_TOOL_DETAILS logs exact bash commands and MCP parameters. Enterprise security teams demand verifiable audit trails for AI agents. They require proof of exactly what code Claude executed. You must treat these logs as high-risk assets, as credential leaks inevitably occur in raw bash inputs. Anthropic implemented length constraints to mitigate massive diff leaks, but you still need strict data masking on your OTLP sinks.
Telemetry polling gains custom timing. The newly documented CLAUDE_CODE_OTEL_HEADERS_HELPER_DEBOUNCE_MS variable stops OTLP header generation from overwhelming authentication servers. Dynamic telemetry triggers script executions. Without debouncing, rapid agent iterations crash internal gateways. Set this interval to match your specific token rotation lifecycle.