Claude Code Digest — 2026-05-19 00:39:03
What the docs reveal
The End of Monolithic Task Tracking
Anthropic replaced the legacy TodoWrite tool with a granular CRUD toolset (TaskCreate, TaskGet, TaskUpdate, TaskList) as of Claude Code 2.1.142 and SDK 0.3.142. Previously, Claude maintained state by overwriting the entire task list. This approach wasted tokens and risked data loss during multi-agent workflows.
The new system enforces precise record-keeping. You must update your monitoring logic to track task state using maps keyed by task ID. If your observability scripts rely on parsing a single todo list payload, they will break. Set CLAUDE_CODE_ENABLE_TASKS=0 if you require an emergency fallback to the old behavior.
Background Initialization for MCP Servers
Startup wait times are gone. Claude Code now connects to external Model Context Protocol (MCP) servers in the background by default. Bypassing the mandatory five-second blocking period speeds up interactive sessions significantly. If your initial command requires immediate access to an external tool, set alwaysLoad: true for that specific MCP server to force the client to wait.
Anthropic paired this change with the new ToolSearch discovery category, which loads tools on-demand. Resolving the tension between robust tool availability and strict context limits is clearly a current engineering priority. Preloading comprehensive toolsets degrades model reasoning and inflates prompt costs; dynamic discovery solves both.
Guarding the Token Budget
Token awareness dictates the latest configuration updates. The workspace details pane now displays estimated token costs for plugins before you install them. Furthermore, subagents like Plan and Explore now explicitly ignore your global CLAUDE.md and repository git status. Anthropic hardcoded this limitation. Stripping global context protects your inputs when delegating narrow, isolated requests to specialized agents.
In interactive pricing updates, Fast Mode now defaults to Opus 4.7. This pipeline drains usage credits at $30/$150 per million tokens (input/output). Exceed your rate limits, and the client automatically bumps you to standard models to maintain uptime. Note that accessing the 1-million token context window for Sonnet now consumes usage credits across all tiers. Max, Team, and Enterprise users only receive unlimited access to the Opus 1M context window.
Closing Repository Security Loopholes
You can no longer enable auto permission mode from project-level configuration files. Anthropic restricted defaultMode: "auto" exclusively to global user settings (~/.claude/settings.json). This closes a glaring supply-chain vulnerability: a malicious repository could previously declare elevated permissions in its .claude directory and compromise your machine upon instantiation.
IDE integrations for VS Code and JetBrains also saw security refinements. The client now automatically pulls your active editor selection directly into the prompt context. You must explicitly configure Read deny rules for files containing secrets. The documentation confirms that standard deny rules strictly block both terminal reads and active IDE selection sharing.
API Alignment and Hook Caps
The experimental V2 session API for TypeScript (unstable_v2_createSession) is dead. TypeScript developers must adopt the standard query() function with the continue: true parameter to persist sessions. Removing the experimental TypeScript interface eliminates a fractured developer experience and aligns the ecosystem around the mature Python SDK patterns.
Finally, Anthropic introduced a hard safeguard against runaway background logic. Claude Code now intervenes and overrides custom Stop hooks if they iterate eight consecutive times without yielding actionable progress. Infinite loops will no longer silently drain your prepaid usage credits. To dictate custom limits for intentionally long-running build processes, scripts must check the new stop_hook_active boolean or define the CLAUDE_CODE_STOP_HOOK_BLOCK_CAP environment variable.