Claude Code Digest — 2026-04-17 00:15:47

Version updates

What the docs reveal

The Shift to Effort Over Token Budgets

Anthropic retired the fixed token budget mechanism for newer models. Opus 4.7 relies entirely on dynamic "effort" levels ranging from low to max. This signals a fundamental shift in how Claude allocates compute. Instead of halting mid-thought when hitting a rigid token ceiling, the model scales its reasoning steps based on the abstract complexity tier you set via the /effort command.

If your application strictly requires the old fixed-budget behavior on 4.6 models, assert the CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING environment variable. Furthermore, the 1M-token context window is out of beta. Append [1m] to your model strings (e.g., claude-opus-4-7[1m]) to activate extended context.

Background Monitors Transform Claude into a Daemon

Plugins now support a monitors.json configuration. This allows background processes to stream system events directly into Claude's context, enabling proactive responses. Claude Code is evolving from a synchronous REPL into an autonomous daemon. You no longer have to paste raw log dumps to debug issues. A plugin can watch a test runner or a build directory, and Claude will react the moment it spots a failure. Keep in mind that background tasks explicitly do not persist when you pause and resume sessions.

Fullscreen TUI and Context Recovery

Long terminal sessions frequently suffer from scrolling artifacts and visual stutter. The new /tui fullscreen renderer eliminates display flickering and introduces native mouse interaction. Anthropic paired this visual overhaul with an automatic session recap feature. If you step away for three minutes, Claude prints a one-line summary of recent activity when you return. You can trigger this recap manually with /recap.

To manage complex project histories, the /resume picker now supports cross-worktree resumption and keyboard-driven filtering across git branches. This eliminates the friction of manually navigating directories to find isolated coding sessions.

Agent SDK Reaches Enterprise Maturity

The Agent SDK received aggressive reliability upgrades. The query() method now handles JSON schema validation automatically. If the model outputs malformed JSON, the SDK prompts the agent to fix the structure before returning the payload to your application. This eliminates a massive amount of scaffolding code.

Anthropic also introduced the startup() utility. This function pre-spawns the Claude CLI process in the background. Pulling initialization overhead out of the critical path drastically cuts latency for subsequent queries.

Configuration Scoping and Security

Configuration bleed across global, project, and local environments historically caused unpredictable SDK behavior. The new settingSources array fixes this by requiring you to explicitly define where settings originate. Pass settingSources: [] to guarantee configuration isolation—a mandatory practice for secure, multi-tenant deployments.

Use caution when initializing subagents. They automatically inherit their parent's permission structure. If you launch a session with bypassPermissions, every spawned subagent possesses full, autonomous system access.

Windows First-Class Support

Bash execution has consistently caused friction for Windows users. Anthropic addressed this by shipping a native PowerShell tool. Set CLAUDE_CODE_USE_POWERSHELL_TOOL=1 and Claude routes execution to pwsh.exe or powershell.exe. This grounds the agent in the host OS idiom rather than forcing WSL or emulated bash environments.

View commit