Claude Code Digest — 2026-04-19 00:13:25
Version updates
2.1.114 Anthropic fixed a permission dialog crash triggered when an agent's teammate requested tool access. This crash reveals Anthropic's focus on multi-agent orchestration. As autonomous teams spawn sub-agents and share local tools, the interface struggled to render nested permission prompts. This fix removes a critical roadblock for running complex agent swarms locally.
What the docs reveal
Defeating alert fatigue
Anthropic faces a persistent terminal challenge: alert fatigue degrades security. Developers blindly approve repetitive permission prompts. The new /fewer-permission-prompts command solves this behavior. It analyzes routine read-only Bash and MCP tool calls and automatically generates exact allowlists in your .claude/settings.json.
To counter this leniency, Anthropic tightened execution sandboxing. The documentation isolates command wrappers (watch, setsid) and destructive find flags (-exec, -delete) from broad prefix auto-approvals. You must write exact-match rules to permit them. A careless wildcard approval on find previously granted an agent unrestricted capability to modify system files. Anthropic closed that vector.
Enterprise networking and secure egress
Corporate environments dictate the new network parameters. The TypeScript Agent SDK introduces httpProxyPort and socksProxyPort configurations, allowing Claude Code agents to operate reliably behind strict corporate firewalls.
Parallel to proxies, Anthropic introduced deniedDomains. This setting establishes a rigorous blocklist that supersedes any allowedDomains rules. You can approve broad wildcard domains for an agent's web-scraping tool while explicitly blackholing your company's internal infrastructure endpoints.
Treating the local daemon as a server
Remote control architecture underwent a massive upgrade. Web and mobile remote sessions now map directly to your local project context. When you trigger an @ command from your phone, the application streams local file trees and MCP server states from your laptop instantly.
This parity indicates the local CLI now runs a robust, asynchronous API server in the background. However, the architecture handles interactive UI poorly. Anthropic restricts commands requiring native terminal rendering—like /mcp, /plugin, or /resume—to the local CLI interface. You cannot manage plugins remotely.
Upgrading terminal ergonomics
Claude Code steadily absorbs standard IDE text-editing mechanics. Anthropic shipped full selection expansion mapping (Shift plus arrow keys, Home, and End) alongside vertical viewport auto-scrolling. You span massive logs faster using the new scroll:fullPageUp and scroll:fullPageDown actions. Anthropic builds these native keyboard ergonomics to capture developer attention. They want you operating entirely inside the Claude Code interface, minimizing round-trips to traditional editors.
Finally, task management keystrokes require greater precision. Pressing Esc only halts tasks waiting for their next loop iteration. You must manually delete explicitly scheduled standalone tasks. This specific isolation prevents developers from accidentally killing background deployments when quickly clearing an active prompt buffer.