Claude Code Digest — 2026-03-26 16:30:45
What the docs reveal
Anthropic is transforming Claude Code from a solo coding assistant into an orchestrator. The introduction of TaskCreated and TaskCompleted hooks reveals a shift toward multi-agent teams. Developers now possess circuit breakers. Sub-agents often spawn expensive or destructive tasks. You can intercept and reject these tasks before execution by returning exit code 2 on the TaskCreated hook. A JSON payload containing {"continue": false} halts the teammate entirely. These hooks execute unconditionally without matchers. Autonomous coding requires precise programmatic oversight, and this update delivers it.
Power users dictate the terminal experience. The interface now supports an Emacs-style kill-ring. You cycle through paste history using Alt+Y after an initial paste. Keyboard toggles now control execution models. Developers can press Meta+O to trigger fast mode and Meta+T to trigger extended thinking. Prompt complexity dictates model choice. Fast mode powers through boilerplate. Extended thinking unravels dense architectural logic. You change execution models instantly without touching configuration files.
You must rewrite bloated MCP tool descriptions. Anthropic established a strict 2KB truncation limit on tool schemas and instructions. Large schemas consume context window space and degrade instruction following. Place high-priority guidance at the beginning of the string. The model permanently discards anything beyond the 2KB limit.
Ecosystem growth introduces supply chain vulnerabilities. Claude Code now enforces strict governance over server configurations and marketplace installations. Local tool configurations explicitly override and disable conflicting claude.ai cloud connectors. Enterprise IT can completely disable automatic cloud integration using ENABLE_CLAUDEAI_MCP_SERVERS. The permissions model now restricts allowed channel plugins and supports blockedMarketplaces. SecOps teams demand air-tight controls over local development environments. These variables provide the necessary locks.
Global skill initialization wastes context tokens. Skills now support path-based activation using glob patterns. A database migration skill stays dormant until you interact with the backend directory. A frontend toolkit activates only inside the interface folder. Anthropic also added first-class PowerShell support. Windows developers secure native execution by configuring CLAUDE_CODE_USE_POWERSHELL_TOOL. Contextual targeting prevents hallucinated command syntax.
Remote control operations now demand a full-scope authenticated session. The system explicitly rejects raw environment variables like ANTHROPIC_API_KEY and legacy OAuth tokens for remote tasks. Remote coordination delegates high-privilege operations across machines. Standard API keys lack the identity verification required to secure these external interactions.
Heavy operations snap client connections. You can now configure CLAUDE_STREAM_IDLE_TIMEOUT_MS to sustain prolonged network queries. Massive codebase indexing operations and sluggish compile steps dictate long wait times. Adjust this duration upward to prevent premature client crashes during deep diagnostic tasks.