Run your first review

EV gives the same change a fresh review context. Let your coding agent run the common review-fix-repeat loop, or open the review in Desktop when you want to inspect and curate the result yourself.

Before you start

You need:

• EV installed on an Apple Silicon Mac
• a Git or jj repository containing the work you want reviewed
• a base branch or ref to compare it with, such as origin/main
• a terminal opened at that repository

Let your coding agent run the loop

EV includes an optional workflow skill for Codex and Claude Code. EV Desktop offers to install it when it detects a supported agent, or you can install it from the terminal:

# Codex
ev agent setup codex

# Claude Code
ev agent setup claude

The command shows its plan and asks before writing to your agent configuration.

Then invoke the workflow in your coding agent:

Agent	Invocation
Codex	$ev-review
Claude Code	/ev-review

The workflow asks EV to review the change in a separate context, reads the findings, fixes valid issues, records decisions, and reviews again when needed. You stay in the coding-agent conversation instead of moving findings between tools by hand.

Tip

You can ask for a narrower pass in the same prompt—for example, “Use $ev-review and focus on authentication and error handling.”

Review with Desktop

Use Desktop when you want to read the evidence and make the final decisions yourself:

git fetch origin --quiet
ev review -b origin/main --desktop

-b origin/main reviews your current work against its base on origin/main. EV opens the review in Desktop, runs the selected review skills, and places each finding beside the relevant code and evidence. Replace origin/main with the remote base branch your work actually uses.

Work through the result:

1. Read the summary and findings.
2. Address a valid issue in the code, then choose Resolve to record that outcome.
3. Use Dismiss for a false positive, accepted risk, or deliberate decision not to fix it. Add a reason so the decision remains useful later.
4. Add your own comments where the automated review missed context.
5. Choose Submit review.

The waiting terminal receives the review as Markdown. Desktop also copies that Markdown to the clipboard, ready to paste into a coding-agent conversation.

Note

Resolve records that a finding has been addressed. It does not ask an agent to make the change for you. Fix the issue—or decide how it should be handled—before recording the outcome.

Run an automated review directly

Agents and scripts can call EV without opening Desktop:

git fetch origin --quiet
ev review -b origin/main --stream --auto

• --stream emits newline-delimited JSON for an agent or script to consume.
• --auto selects relevant reviewer skills from the files in the change. AI review already runs by default.
• Using the remote ref avoids comparing against a stale local base branch.

Choose what to review

Comparing the current work with its remote base is a common starting point, but it is only one review scope. Run ev review without a scope option for an interactive choice, or select one directly:

Scope	Command
Current work against a base branch	ev review -b origin/main
Uncommitted changes against HEAD	ev review --working-copy
One commit	ev review --commit <sha>
Current work against an exact ref	ev review --from <ref>
Exact committed range	ev review --from <base> --to <tip>

Add --desktop to inspect and curate the review in Desktop, or --stream --auto for an agent-friendly automated run.

See Choose a review scope for the differences between base branches, exact refs, commits, ranges, working-copy reviews, and supplied diffs.

What EV keeps

The review is stored locally with its scope, evidence, findings, comments, and decisions. That durable record lets the next pass distinguish a new issue from something you already fixed or deliberately dismissed.