Claude Code Digest — 2026-03-27 00:38:51
What the docs reveal
Anthropic declared war on context bloat. The latest documentation changes reveal a clear product direction: developers must treat the LLM context window as a strict, finite budget. As engineers connected more Model Context Protocol (MCP) servers to their sessions, API costs spiked and instruction adherence degraded. Anthropic changed the defaults to protect your context window.
The End of MCP Bloat
Developers accumulated MCP servers, filling the initial context window with idle tool schemas. Claude Code now uses "tool search" by default. The system loads only tool names up front and defers full schemas until Claude explicitly invokes the tool.
You can override this strict lazy-loading by setting ENABLE_TOOL_SEARCH=auto. This hybrid approach loads tool schemas into the context window up to a 10% threshold, deferring the rest.
The docs now explicitly recommend native CLI tools (gh, aws, gcloud) over MCP servers. Native tools sidestep the context overhead of tool listings entirely. This signals a practical shift in how Anthropic views tool integration. Use MCP servers for bespoke internal data; use native CLI tools for standard infrastructure.
Real-Time Token Visibility
Flying blind wastes money. Anthropic added a custom status line to track context usage in real time. New commands like /context and /mcp break down exactly which files and servers drain your tokens.
The documentation also introduces an interactive visual guide for session token consumption. Anthropic wants developers to monitor token usage proactively rather than reacting to degraded performance. When the context window fills with noise, Claude loses the plot. Keep the context clean.
Subagents Break Out
Previously, spawning a subagent or parallel worktree often failed if the task required local secrets. The new .worktreeinclude file fixes this limitation. It automatically copies untracked, gitignored files like .env into new worktrees. Subagents now boot with necessary credentials intact.
Subagents operate in complete context isolation. They load their own instances of CLAUDE.md, skills, and MCP tools. You should offload memory-bound tasks, like deep repository analysis or sprawling refactors, to subagents. They execute the task without polluting the main session's token budget.
To maintain task continuity, subagents now automatically inherit up to 200 lines (25KB) of MEMORY.md. More importantly, Anthropic explicitly instructs subagents to curate and trim this memory file. Anthropic relies on bounded, self-pruning memory routines to prevent runaway token costs during long-running parallel tasks.
Strict Rules for Hooks and Instructions
Anthropic clarified the mechanics of PostToolUse hooks. Plain stdout feedback drops into the void. To inject hook feedback into the context, your scripts must format the output as JSON within the additionalContext field. The system demands structured data to ensure reliability.
Finally, the documentation reminds developers that CLAUDE.md is not a rigid configuration file. It consumes tokens and acts as a system prompt at the start of every session. Poorly written instructions degrade Claude's reliability. Write concise, specific prompts. Omit needless words.