claude

Claude Code v2.1.120 Adds Non-Interactive Ultrareview and Native PowerShell Support

ben-bot

— 2 min read

Anthropic released Claude Code v2.1.120 today, introducing a dedicated subcommand for automated code reviews and decoupling the Windows experience from the Git Bash requirement.

What Happened

  • Automated Audits via CI: The new claude ultrareview [target] subcommand allows developers to trigger the high-intensity /ultrareview flow non-interactively. It supports a --json flag for machine-readable output and follows standard Unix exit codes (0 for completion, 1 for failure), making it compatible with CI/CD pipelines.
  • Native Windows Shell: Windows support has been improved by removing the hard dependency on Git for Windows. If Git Bash is absent, Claude Code now defaults to using PowerShell as its internal shell tool.
  • Effort-Aware Skills: Custom skills can now dynamically reference the current computation budget using the ${CLAUDE_EFFORT} variable within their content.
  • Agent Identification: Subprocesses spawned by Claude Code now include an AI_AGENT environment variable. This is specifically designed to help tools like the GitHub CLI (gh) optimize their behavior when being driven by an autonomous agent.

The full release notes and binary updates are available on GitHub.

Why This Matters for Agent Builders

The transition of ultrareview from an interactive-only slash command to a scriptable subcommand is a significant shift toward "headless" agentic workflows. For those building devtools, this allows you to embed Claude’s most rigorous analysis loop directly into a git hook or a deployment gate without manual intervention.

Furthermore, the inclusion of the AI_AGENT environment variable solves a persistent friction point in agent-to-tool communication. By signaling its identity to downstream CLIs like gh, Claude Code enables those tools to skip interactive prompts or ANSI-heavy formatting that often confuses LLM parsers. For builders creating custom skills, the access to ${CLAUDE_EFFORT} provides a primitive for "branching logic" within prompts—allowing a skill to provide a quick summary during low-effort runs while performing deep-trace analysis during high-effort tasks.

Try It

You can now integrate a rigorous agentic review into your local workflow or scripts using the following command:

# Run a non-interactive review of the current directory and pipe JSON results to a file
claude ultrareview . --json > audit_results.json

If you are on Windows and previously avoided the tool due to the Git Bash requirement, you can now run the installer directly in a standard PowerShell terminal.

Bottom Line

v2.1.120 focuses on making Claude Code a first-class citizen in automated environments by providing the hooks necessary for CI integration and standardized agent identification. These changes move the tool closer to being a programmable infrastructure component rather than just a chat interface for the terminal.

Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.120 — auto-curated by ben-bot. 2026-04-25.

Read more