Claude Code Digest — 2026-03-26 04:50:56
Version updates
Version 2.1.84 ships the PowerShell tool for Windows as an opt-in preview. Anthropic historically built Claude Code around Unix-like environments, leaving Windows developers reliant on WSL or unreliable script translations. Native PowerShell execution reduces virtualization overhead and allows the agent to interact directly with the Windows host. The opt-in nature signals the agent likely still struggles with PowerShell's object-oriented pipeline outputs; text streams remain easier for LLMs to parse.
The release also introduces ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU}_MODEL_SUPPORTS environment variables. These allow developers to force-override specific features like effort levels. This change pairs directly with new third-party routing configurations detailed below.
What the docs reveal
Taking back the reasoning budget
The documentation introduces CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING. Anthropic recently shifted Opus 4.6 and Sonnet 4.6 to use adaptive reasoning by default. This shift silently ignores your standard MAX_THINKING_TOKENS budget. Adaptive thinking improves holistic code analysis but destroys predictability. If the model decides to overthink a simple variable rename, you pay the compute cost and suffer the latency. The new toggle hands control back to the developer. Disable adaptive reasoning when running boilerplate scripts or routine chores to enforce strict operational speed.
Unlocking enterprise endpoints
Corporate developers using Bedrock, Vertex AI, or Foundry can finally utilize native capabilities. Previously, routing Claude Code through enterprise endpoints meant using custom deployment IDs. Custom IDs broke Anthropic's internal model detection, silently disabling advanced features like interleaved reasoning or extended thinking. The new _SUPPORTED_CAPABILITIES environment variable lets you manually bypass this check. You can now declare feature support directly. Your enterprise pipeline no longer guarantees a degraded development experience.
Hardening git worktree isolation
WorktreeCreate hooks now enforce a strict contract: they must return an absolute path to the intended directory. Failure to do so aborts the creation. This update addresses directory drift. When Claude executes complex refactors across sub-directories, relative paths easily resolve to the wrong parent. Enforcing absolute paths guarantees the agent maintains its workspace context. Additionally, WorktreeCreate and WorktreeRemove no longer use the standard allow/block decision model. The CLI treats worktree manipulation as an explicit system state change, not a prompt intervention. WorktreeRemove failures now only log in debug mode to prevent disruptive execution halts.
Fixing startup and navigation ergonomics
The documentation clarifies SessionStart exclusively accepts command hooks. Do not attempt to embed interactive shells or blocking scripts during initialization; the agent demands an uninterrupted boot sequence. Lastly, Claude Code introduces proper keyboard navigation for the footer UI. You can traverse, select, and clear elements using the arrow keys, Enter, and Escape. Mouse reliance in a terminal application breaks developer momentum. These keybindings eliminate context switching and keep your hands on the home row.