Claude Code Digest — 2026-03-27 16:22:45
What the docs reveal
Anthropic fundamentally changed how Claude handles Model Context Protocol (MCP) tools to solve context window bloat. Integrating dozens of external MCP tools previously forced Claude to load every schema into the prompt upfront. This saturated the context window and wasted tokens. The new ENABLE_TOOL_SEARCH environment variable introduces deferred loading. Developers can set this to true to force deferral, false to load upfront, or auto:N to dynamically defer based on context limits. Anthropic built this because sprawling toolchains made daily usage prohibitively expensive. Set this to auto:20 to keep agent latency low while reserving context for your actual codebase.
The new teammateMode setting isolates a shift from single-agent workflows to multi-agent orchestration. The documentation quietly reveals rendering options for teammate panes via tmux, iTerm2, or in-process rendering. Anthropic envisions developers spinning up concurrent, specialized agents rather than relying on one monolith. A frontend agent translates React components while a backend agent writes the database migrations. This configuring prepares your terminal for parallel execution. Review your multiplexer configs to accommodate these new panes.
Claude now handles Git worktree state automatically using the .worktreeinclude file. Previous versions stranded agents without required database credentials when spawning background worktrees. Developers can now create a .worktreeinclude file at the repository root to whitelist ignored files, ensuring Claude copies .env files automatically. Anthropic designed a strict override mechanism for custom scripts. If you configure a custom WorktreeCreate hook—which now officially supports non-Git systems like Subversion and Perforce—Claude disables internal .worktreeinclude parsing entirely. If you write the hook, you own the pipeline.
Several configuration updates refine the prompt-to-execution pipeline. The showClearContextOnPlanAccept toggle enforces strict session hygiene. Enabling this clears the conversation history the moment you approve a generated coding plan. This forces the agent to execute instructions using a clean slate, eliminating hallucinations caused by meandering planning conversations. The fastModePerSessionOptIn setting overrides global preferences, letting developers default to the smartest model and drop down to fast mode only for specific, repetitive sessions.
Voice interactions establish Claude Code as an active pair programmer. The new voiceEnabled feature introduces push-to-talk dictation explicitly tied to your Anthropic account. Combined with UI tweaks like spinnerTipsOverride and prefersReducedMotion, the terminal experience moves away from a rigid text prompt. Anthropic expects developers to dictate complex debugging instructions while stepping through code, eliminating context switching. Enable voice dictation to accelerate your feedback loops.