Claude Code Digest — 2026-03-30 20:17:15
Version updates
2.1.87 (March 29, 2026): Anthropic patched Cowork Dispatch to fix silent message failures. Orchestrating multiple subagents heavily strains state management. This fix suggests Anthropic is hardening message queues to ensure complex, multi-agent workflows complete reliably rather than stalling mid-task.
What the docs reveal
The Arrival of Native macOS Computer Use
Anthropic brought native GUI control to Claude Code. It runs directly on your host—not in a sandbox.
This shifts Claude from a code editor into a genuine pair programmer. Claude can now click through the iOS Simulator, drive a browser to replicate a bug, or execute UI-heavy testing workflows. Because it runs unconstrained, developers must supervise the system. macOS per-app permissions serve as your primary defense.
Anthropic built clear guardrails into this architecture. Claude automatically excludes your active terminal from its spatial screenshots. This prevents the model from reading its own output, which typically triggers an infinite feedback loop. If an action goes rogue, pressing Esc globally aborts the execution. Native control requires a Pro or Max plan, an active macOS session, and manual activation via the /mcp menu.
Solving Approval Fatigue
Developers suffer from approval fatigue during extensive refactors. Anthropic addresses this through a revised permission architecture handling Auto and Bypass modes.
You can now toggle modes dynamically using Shift+Tab. Using the --allow-dangerously-skip-permissions flag unlocks these transitions without locking you into a persistent state.
Auto Mode: Anthropic delegates safety to a secondary LLM classifier. This server-side probe prevents prompt injections and sanitizes tool outputs before Claude processes them. Because this pipeline requires heavy computational lifting, Anthropic limits Auto mode to Sonnet 4.6 and Opus 4.6 models on supported plans. You cannot use it with Haiku or third-party providers.
Scoped Bypass Mode: Total bypass previously invited disaster. Anthropic refined the scope. Bypass mode now explicitly exempts critical configuration directories (.git, .vscode, .idea, and .claude). You can let Claude execute scripts freely in your src/ directory, but it cannot silently corrupt your repository history. Modifications to protected zones still demand specific manual authorization.
Model Constraints and Configuration Hierarchies
Enterprises constantly fight model drift and API costs. The model parameter previously functioned as a suggestion. Claude often swapped models dynamically during complex tasks.
Anthropic clarified that model only dictates the initial session model. If you require hard enforcement for compliance or budget limits, you must now compose a multi-layered configuration. Define explicit constraints using availableModels, model, and environment variables together. Update your repository's .claude.json to prevent junior developers from accidentally spinning up Opus 4.6 for trivial tasks.
Breaking External Integrations
Enterprise environments operate behind rigid walls. Anthropic overhauled network specifications to accommodate these boundaries. You must explicitly allowlist Anthropic API IP addresses to integrate with self-hosted GitHub Enterprise Server (GHES) environments for repository cloning and review comment posting. You can also configure internal plugin marketplaces using host-based regex pattern matching.
Voice dictation creates a schism in the user base. Audio processing requires cloud streaming directly through Anthropic's proprietary multi-modal pipeline. Because Anthropic refuses to proxy voice data through third-party APIs, Bring-Your-Own-Key users (Bedrock, Vertex AI, Microsoft Foundry) lose voice integration entirely. You must maintain a direct Claude.ai account to talk to your terminal.