Claude Code Digest — 2026-05-07 00:19:57
Version updates
Anthropic released versions 2.1.129 through 2.1.132. The updates focus on terminal compatibility, session continuity, and plugin injection.
Terminal rendering commands. Version 2.1.132 introduces CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1. Developers hate when TUI applications hijack terminal scrollback. This flag forces Claude Code to run inline using standard output instead of spawning a pristine, full-screen canvas. Version 2.1.129 adds CLAUDE_CODE_FORCE_SYNC_OUTPUT=1 to fix visual tearing in obscure emulators like Emacs eat. You now dictate how the tool renders.
Session continuity. Bash tool subprocesses now inherit the CLAUDE_CODE_SESSION_ID environment variable. You can finally map external shell execution states directly to internal hook lifecycles.
Remote plugins. Version 2.1.129 introduces --plugin-url. You can load a zipped plugin payload into a live session without modifying your local filesystem. This accelerates prototyping and sharing ephemeral tools across teams.
What the docs reveal
The war on context bloat
Anthropic clearly hit a wall with context window limits. Previously, defining dozens of custom skills forced Claude to read every configuration manifest on startup. The latest documentation reveals an architectural fix: dynamic skill loading. Claude Code now loads full skill content only upon direct invocation.
You gain surgical control over this system via the new skillOverrides setting in .claude/settings.local.json. You can configure tools as name-only, user-invocable-only, or off. Apply disable-model-invocation: true to hoard third-party skills without paying an automatic token tax on every prompt. Anthropic recommends these budgets to stay under the internal SLASH_COMMAND_TOOL_CHAR_BUDGET. This marks a distinct product shift from automatic ingestion to manual context optimization.
Fine-grained tool streaming
The new CLAUDE_CODE_ENABLE_FINE_GRAINED_TOOL_STREAMING environment variable changes how the agent communicates state. Instead of hiding long file rewrites behind an agonizing, silent UI pause, Claude Code renders tool inputs as text streams in real time. This solves a major UX complaint for direct API users. Enable this if you demand immediate visual feedback during heavy compilation or rewrite tasks.
Gateway proxies go mainstream
The CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY flag proves developers operate Claude Code behind local inference gateways like LiteLLM. Instead of manually editing configuration files whenever an enterprise proxy rotates or adds a model, Claude Code now polls the standard /v1/models endpoint at startup. The interface dynamically tags these routing paths with a "From gateway" label. This eliminates configuration drift for corporate engineering teams.
Persistent environment ambitions
Plugins now support experimental.themes and experimental.monitors inside the plugin.json schema. You can define background jobs and inject custom color palettes into your CLI session. Though Anthropic warns the schemas will drift, this push toward visual customization proves Claude Code wants to be a persistent, personalized development environment. It no longer acts merely as a transient command-line utility.
Streamlined interactive workflows
Search boundaries default to narrower scopes. Pressing Ctrl+S toggles interactive searches between the current session, the current project, or all projects. Anthropic dropped the expensive "all projects" default, acknowledging that developers execute the vast majority of commands against localized file trees.
Package managers adopt silent operation. Setting CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE triggers automatic, background upgrades for WinGet and Homebrew users. A new autoUpdatesChannel setting lets you pin stability to the stable branch or constantly ride the latest release. Anthropic wants to push toolchain fixes immediately without waiting for developers to remember routine package operations.