Claude Code v2.1.83 Released — Auto Mode, Transcript Search, Credential Scrubbing & 40+ Fixes
Anthropic shipped Claude Code 2.1.83 yesterday and it's a meaty one. The headline feature is auto mode — a classifier-backed permissions layer that lets Claude power through long tasks without peppering you with approval prompts every thirty seconds, but without going full --dangerously-skip-permissions either. Before each tool call, an AI classifier checks for destructive actions, sensitive data exposure, and prompt injection; safe calls fire automatically, risky ones get blocked and Claude reroutes, and if it keeps hitting walls it'll finally surface a human prompt. It's the agentic loop developers have been asking for.
The rest of the changelog is equally dense. New managed-settings.d/ drop-in policy fragments let teams layer config without clobbering each other's settings. Two new hook events — CwdChanged and FileChanged — unlock proper direnv integration so your shell environment updates as Claude moves between project directories. The sandbox.failIfUnavailable flag is a welcome hardening option for CI pipelines that need to guarantee sandboxing or die trying. And CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 strips credentials from subprocess environments — a production security win for any team running Claude Code in CI with secrets in the environment.
On the quality-of-life side: transcript search is now available via / in Ctrl+O mode, Ctrl+X Ctrl+E drops you into your editor mid-session, and agents can now declare an initialPrompt in frontmatter. The 40+ bug fixes include the macOS hang-on-exit issue, scrollback jumping, and a caffeinate leak that was quietly preventing Macs from sleeping. Worth the upgrade on the caffeinate fix alone.