Claude Code Digest — 2026-03-14

Version updates

Claude Code 2.1.76 introduces MCP elicitation. MCP servers can now pause mid-task and request structured user input via an interactive dialog or browser URL. Previously, tools requiring dynamic input or real-time authentication either failed outright or relied on brittle workarounds. This release standardizes the handoff. You provide the inputs, and the agent continues its loop uninterrupted.

Anthropic also enabled a 1-million token context window by default for Opus 4.6 on premium tiers in version 2.1.75. A window this massive changes how you scope projects. You no longer need to micro-manage file reads across large codebases. However, pushing context limits degrades response speed and inflates API costs. Anthropic knows this, which explains the granular context management tools hidden in the documentation update.

Finally, developers juggling multiple workspaces receive overdue session management tools. You set a session name with the -n CLI flag or /rename command. You assign terminal prompt colors mapped to different workspaces via /color. These simple additions prevent you from accidentally executing destructive bash commands in the wrong repository.

What the docs reveal

The end of interactive bottlenecks

The new Elicitation and ElicitationResult hooks reveal Anthropic's vision for headless agent operations. Interactive dialogs kill automation. If you run a long background task, a mid-task credential prompt leaves the agent permanently stalled.

Hooks solve this. You intercept the MCP server's request before the UI renders. You read the requested JSON schema, pull the required credentials from your local password manager, and return the hookSpecificOutput with an accept action. The agent receives the data and proceeds. You never see the prompt. You use the ElicitationResult hook to audit or block user responses when operating in strict enterprise environments.

Decoupling context from compaction

A 1M context window requires aggressive memory management. Auto-compaction historically triggered at 95% capacity. Waiting for a 1M window to hit 95% guarantees catastrophic latency and massive API bills.

The new environment variable CLAUDE_CODE_AUTO_COMPACT_WINDOW decouples the compaction threshold from the physical model limit. You instruct the engine to treat the context window as 200K tokens for compaction calculations, even while running Opus 4.6. You enforce this upper bound, keep loop latency low, and let the model retain the deeper 1M capacity for direct file reads.

Anthropic paired this with the new PostCompact hook. It fires synchronously after compaction completes and receives the compact_summary payload. You capture this summary and write it to external logs or persistent memory stores. You track exactly what the model forgets.

Adaptive reasoning overtakes fixed token budgets

Anthropic abandoned raw token math for Claude 4.6 models. The documentation deprecates MAX_THINKING_TOKENS for Opus 4.6 and Sonnet 4.6. The engine ignores the variable entirely.

Instead, Claude Code now uses adaptive reasoning. The model dynamically scales its thinking budget based on the session's conceptual difficulty. You control this threshold via the /effort command or the CLAUDE_CODE_EFFORT_LEVEL environment variable (low, medium, high, max).

This reflects a shift in prompt engineering. Hardcoding a 10,000-token limit forces the model to truncate complex thoughts or waste tokens on trivial tasks. Adaptive effort trusts the model to optimize its compute overhead. When you need extreme depth for a single query without permanently altering the session effort, you simply type the keyword ultrathink in your prompt. The agent temporarily spikes its reasoning budget, resolves the architecture problem, and drops back to its baseline constraint.