Claude Code 2.1.120 adds `claude ultrareview` and fixes `DISABLE_TELEMETRY` opt-out

• ClaudeCodeLog's release post says Claude Code 2.1.120 shipped with 22 CLI changes, led by a new claude ultrareview subcommand for running /ultrareview non-interactively in CI and scripts.
• According to ClaudeCodeLog's changelog thread, the release also restores DISABLE_TELEMETRY and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC for API and enterprise usage metrics, closing a broken opt-out path.
• ClaudeCodeLog's changelog thread lists fixes for an MCP Esc disconnect regression, broken keyboard input after claude --resume, scrollback duplication, fullscreen clipping, and a Bash-tool file descriptor leak that could crash hosts on large directory trees.
• ClaudeCodeLog's additional updates shows a much smaller prompt bundle in 2.1.120, with 148 prompt files removed and prompt tokens down 91,956, while adding an auto-mode rule reviewer and conversation summarization helpers.

You can read the tagged v2.1.120 release, inspect the CLI surface diff, and dig into the prompt stats. The weirdly useful bit is the auto-mode rule reviewer prompt, which suggests Anthropic is now checking classifier rules before they hit production.

The headline feature is claude ultrareview [target]. As ClaudeCodeLog's changelog thread describes it, the command runs /ultrareview headlessly, prints findings to stdout, supports --json, and exits 0 on completion or 1 on failure.

That turns a research-preview review flow into something CI can actually consume. WesRoth's overview and trq212's repost of ClaudeDevs both frame /ultrareview as a cloud review system that spins up parallel bug-hunting agents rather than doing a single local pass.

The most consequential fix is boring in exactly the right way. ClaudeCodeLog's release post says DISABLE_TELEMETRY and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC had stopped suppressing usage metrics for API and enterprise users, and 2.1.120 restores that behavior.

The same changelog groups several recent paper cuts into one patch set:

• Esc during a stdio MCP tool call no longer closes the whole server connection, per ClaudeCodeLog's changelog thread.
• /rewind and other overlays work again after launching with claude --resume, according to ClaudeCodeLog's changelog thread.
• False-positive "Dangerous rm operation" prompts in auto mode were fixed for multi-line bash commands with both a pipe and a redirect, per ClaudeCodeLog's changelog thread.
• The Bash tool's find no longer exhausts open file descriptors on large directory trees and crash the host, according to ClaudeCodeLog's changelog thread.
• Fullscreen menu clipping, collapsed Write output, and jumpy slash-command matching also got patched, per ClaudeCodeLog's changelog thread.

The release adds one visible command and several small plumbing hooks. According to ClaudeCodeLog's additional updates and the linked CLI surface diff, 2.1.120 adds ultrareview, a --timeout option, and new env vars including AI_AGENT, CLAUDE_CODE_DAEMON_COLD_START, CLAUDE_CODE_LEAN_PROMPT, CLAUDE_COWORK_MEMORY_GUIDELINES, and COMSPEC, while removing the claude-code-screenshots model slug.

The prompt bundle changed more than the binary did. ClaudeCodeLog's additional updates reports bundle size up only 136.4 kB, while prompt files fell by 148 and prompt tokens dropped 57.1 percent. The linked metadata and prompt stats back that up.

Four prompt-level additions stand out in ClaudeCodeLog's additional updates:

1. An auto-mode rule reviewer prompt that checks classifier rules for clarity and completeness.
2. A conversation summarization prompt focused on user requests and prior assistant actions.
3. A session title and branch-name generator.
4. A scheduling guardrail prompt that checks prerequisites before offering scheduled tasks.

The tail of the changelog is a small platform cleanup pass. ClaudeCodeLog's changelog thread says Windows no longer requires Git for Windows, because Claude Code now falls back to PowerShell when Git Bash is absent.

A few other platform-specific changes land in the same release:

• Skills can reference the current effort level with ${CLAUDE_EFFORT}, per ClaudeCodeLog's changelog thread.
• Subprocesses now get AI_AGENT, which ClaudeCodeLog's changelog thread says lets gh attribute traffic to Claude Code.
• In VS Code, /usage now opens the native Account & Usage dialog instead of returning plain text, according to ClaudeCodeLog's changelog thread.
• In VS Code, voice dictation now respects the language setting in ~/.claude/settings.json, per ClaudeCodeLog's changelog thread.
• Auto-compact in auto mode now shows auto instead of a misleading token count, according to ClaudeCodeLog's changelog thread.