Fix findings with your AI agent

Fix with AI sends a finding’s instructions to a coding agent you choose, such as your editor’s chat, Claude Code, or Codex. The agent makes the change in your workspace, and you review it as you would any other edit. With the default settings, Fix with AI in VS Code opens VS Code chat in agent mode and submits the prompt right away, so the agent can start editing your files at once. Cursor and Windsurf paste the prompt into their chat without sending it, and other editors, such as Antigravity, copy it to the clipboard. To read the prompt before an agent runs, turn off coderabbit.autoStartTask, or choose another agent. This is different from CodeRabbit’s , which runs in CodeRabbit’s cloud, and from Fix with AI in CodeRabbit Security. See Names that collide.

​

Choose your agent

Set coderabbit.agentType to the agent you use; see Open the settings. The default is Native. Requirements by agent type:

• Terminal agents (Claude Code, Codex CLI, and opencode) run in the editor’s integrated terminal, which stays open for the next handoff. The command runs right away: claude and codex start an interactive session with the prompt, and opencode run runs the prompt once. If the agent from the previous handoff is still running in that terminal, the extension types the new command into it as input; exit the agent or close the terminal first to start a fresh run. The extension adds no permission flags, so the agent’s own approval prompts apply. The agent’s CLI must be installed and on your PATH on the machine where the extension runs, which is the remote machine in a remote window. The extension doesn’t check for it: if the CLI is missing, the terminal shows your shell’s error, and the finding is still marked “Fixed with AI”.
• Extension agents (Claude Code Extension, Cline, Roo, Kilo Code, and Augment Code) need that agent’s extension. If it’s missing, the extension offers View in Marketplace. Claude Code Extension needs Claude Code extension version 2.1.0 or later, below 3.0.0.

If a handoff fails for any reason, the prompt goes to the clipboard so you can paste it yourself. In a workspace you haven’t trusted, only your user values of coderabbit.agentType and coderabbit.autoStartTask apply, so a repository you clone can’t choose which agent runs or whether it starts on its own. See Restricted Mode.

​

What Native does in each editor

To have an agent start on its own in Cursor, Windsurf, or the other editors, choose a terminal agent, or Cline, Roo, or Kilo Code with coderabbit.autoStartTask on.

​

Start the agent automatically or review the prompt first

coderabbit.autoStartTask is on by default. It decides whether Native in VS Code, Cline, Roo, and Kilo Code start right away. Turn it off to read or edit the prompt before the agent runs.

​

Fix one finding

The finding is marked “Fixed with AI” as soon as the extension hands it off, including when the prompt only went to the clipboard. If the agent didn’t make the change, select Unresolve on the finding in Findings to reopen it.

​

Fix several findings at once

After a review completes with unresolved findings, Findings shows a Fix all issues button:

1. Select Fix all issues.
2. Choose Select files or Select comments.
3. Pick the files or findings to fix.

When the review also has unresolved nitpicks, More fix options appears next to Fix all issues. Choose Fix issues and nitpicks to include them, or Fix issues only to leave them out. In Nitpicks, Fix all nitpicks sends every unresolved nitpick, without asking you to pick. Everything you pick goes to your agent as one prompt, and every finding in it is marked “Fixed with AI”. Select Unresolve on any finding the agent didn’t fix.

​

Review, fix, and review again

1. Make changes, yourself or with an agent, and commit. With coderabbit.autoReviewMode set to auto, every new commit starts the automatic review unless you cancel it, including commits an agent makes in a terminal. With prompt, a commit starts a review only when you answer Yes to the notification.
2. Hand the findings to your agent with Fix with AI or Fix all issues.
3. Review again. An automatic review starts only after a commit and covers committed changes, so commit the agent’s edits, or start a review yourself with Uncommitted changes to check them before you commit. The next review covers only what changed since the last one. After an amend or a rebase, it covers the whole branch. See Review again after more changes.

The next review looks only at what changed since the last one and doesn’t re-check earlier findings, so a finding that no longer appears isn’t confirmed as fixed. See Review again after more changes. Each review, including each automatic one, counts toward your hourly limit. Fix with AI doesn’t start on its own when a review finishes; you start each handoff. If you’d rather have your agent start reviews and read the findings itself, use the CodeRabbit CLI with agent skills. To install both from the editor, run CodeRabbit: Set up CLI and skills; see Set up the CLI and agent skills.

​

What’s next