Tool Approval Flow
Your AI agent never touches a file or runs a destructive command without your say-so.
What you'll see
- Launch
ryshand double-press Escape into prompt mode. - Ask the agent to edit a file (e.g. "add a comment to the top of README.md").
- The footer displays the pending action and a diff preview before anything is written.
- Press
yto approve -- the edit applies and the agent continues;nwould cancel it. - Note which tools are gated: file_edit, file_write, multi_edit, apply_patch, git_commit, and dangerous bash; reads and searches run freely.
Commands shown
rysh
add a hello world comment to the top of README.md
y
Keys used
EnterEscape
Transcript
0:00Your AI agent never touches a file or runs a destructive command without your say-so.
0:03Start rysh. We are in shell mode -- your familiar terminal.
0:08Double-press Escape for prompt mode, then ask the agent to change a file. Tools like file_edit always need approval.
0:16Before anything is written, the footer shows the action and a diff preview. Nothing changes until you decide.
0:30Press y to approve. The edit applies and the agent continues. Press n to reject and the action is cancelled.
0:40Reads and searches run freely. But writes, patches, commits, and dangerous bash are all gated behind you.
0:47You are always the final approver.
Key takeaway
Destructive or file-mutating tools always pause for your approval, with a diff preview, before they run.