Claude Code Digest — 2026-04-22 00:13:12
Version updates
Release 2.1.117 fixes a major UX annoyance: /model selections now persist across restarts, even if the underlying project attempts to pin a different model. Anthropic also enabled frontmatter mcpServers for main-thread agents executed via --agent. This streamlines custom agent bootstrapping by dropping the requirement to define servers globally. Subagents can now execute in isolated forks via CLAUDE_CODE_FORK_SUBAGENT=1, giving developers strict process boundaries for dangerous tasks.
What the docs reveal
Enterprise fleets and strict sandboxing
Anthropic clearly wants Claude Code deployed across corporate IT environments. The documentation introduces macOS MDM support (com.anthropic.claudecode), allowing administrators to push central configurations like pre-defined, read-only SSH settings seamlessly. On Team and Enterprise plans, administrators now completely control MCP server management.
Security controls adopt a zero-trust posture. A new failIfUnavailable sandbox setting triggers a hard startup crash if the sandbox fails to initialize. This prevents accidental fallback to unprotected host execution. Furthermore, Auto mode now runs an internal classifier to intercept data exfiltration attempts, actively blocking unauthorized repository pushes or cloud writes. You can tune this classifier locally, but Anthropic forces it to ignore shared project settings. This prevents a malicious pull request from injecting relaxed security rules via .claude/settings.json.
Precision engineering for Agent workflows
The Agent SDK evolves from a wrapper into a robust orchestration framework. Developers gain an effort field to explicitly throttle or boost thinking depth per subagent. You no longer waste high-compute reasoning tokens on simple formatting scripts. Subagents also receive permissionMode controls to restrict tool access at the individual agent level.
Hooks finally achieve necessary state persistence. CwdChanged and FileChanged hooks write directly to CLAUDE_ENV_FILE, preserving environment variable transitions across long-running sessions. Custom slash commands powered by UserPromptExpansion accept additionalContext. This allows you to inject specific system memory into Claude exactly when the user invokes a command.
Skill definitions gain similar flexibility. Skills accept positional arguments mapped to variables like $issue or $branch. You can bypass automated tool selection entirely by setting disable-model-invocation, forcing the skill to trigger only via manual slash commands. You can even assign specific models to specific skills, automatically reverting to the default session model once the execution completes.
Context hygiene and compute offloading
Context windows fill rapidly during interactive sessions. Anthropic solves this with the /btw command. It opens a dismissible overlay for quick questions and explicitly drops the interaction from the main conversation history. Developers can query syntax or check documentation without polluting the primary agent's working state.
Anthropic also deployed /ultrareview, moving deep, multi-agent code reviews into a cloud sandbox environment. Local machines struggle to orchestrate parallel agent execution. By offloading review agents to Anthropic's infrastructure, developers reclaim local compute. Pro and Max users receive three free runs through May 5. This explicitly signals a future pay-per-compute monetization strategy for heavy agent workloads.
Deep observability and raw data extraction
Debugging complex LLM interactions requires ground-truth data. The new OTEL_LOG_RAW_API_BODIES OpenTelemetry flag captures complete Messages API requests and responses. Because massive JSON arrays crash terminals and exhaust local memory, developers can instruct the telemetry engine to dump these raw bodies directly to disk.
New hook tracking captures precise lifecycle events. OpenTelemetry spans now record when MCP servers drop back to prompt the user (Elicitation), track git worktree isolation stages, and measure background task durations. If your system relies on persistent multi-turn conversations, you can also deploy the new SessionStore configuration to mirror session transcripts to external serverless storage backends.