Adding a “Apply First, Confirm Later” Agent Interaction to Roo Code Similar to Cursor

[yuanwer]

Hello everyone,

I would like to start a discussion on improving Roo Code’s interaction experience, specifically by adopting the Cursor Agent model where AI-generated code is first applied to files, allowing users to review the actual running effect before deciding whether to accept the changes.

Background
Currently, Roo Code’s AI interaction flow is that AI generates and directly modifies code files based on instructions. Users must first accept these changes before they can see the actual running effect of the code. If the generated code does not meet expectations, users have to manually find checkpoints and roll back the code. This process has the following pain points:

No instant preview mechanism; users cannot see AI’s suggested changes before code is written.

High rollback cost, especially in large or multi-file projects.

Interaction lacks flexibility; users lack proactive control over AI code changes.

Advantages of Cursor Agent
Cursor Agent presents code changes as suggestions or drafts first. Users can review or test them before confirming to apply, greatly reducing rollback cost and increasing trust.

Proposal
Add an “Agent interaction mode” to Roo Code with:

Code diff preview

Support for running changes in temporary environments or branches

User confirmation before syncing changes

Automatic snapshots and easy rollback

Thank you!

[InTEGr8or]

I usually work around this by allowing auto-approve edits to the files, and then having the agent add the fixed files and create a multiline git commit.

Often, an edit to one file won't work, and changes have to be made to multiple files.

git allows that.