Keep a human decision in the coding-agent loop

Read the action before the button

A familiar-looking test command can belong to an unfamiliar project. Before allowing a tool, identify the session, read the actual command or change, and compare it with the task you intended.

Perch shows the supported request on a card so the decision can happen without first locating the terminal. The goal is to make a deliberate review easier to reach. A short route to Allow is not a reason to skip reading.

Separate different kinds of consent

Approving a tool, choosing a product direction, accepting a plan and replying to a finished turn are different decisions. For Claude Code, Perch gives these their own supported interactions: tool requests, numbered questions, plan review and replies.

Codex and Cursor have narrower action sets. The compatibility guide matters because a single “works with” label cannot explain what you can actually decide from the shelf.

Use undo for a mistaken input

Every answer has a two-second Sending window. Esc or Undo can stop an accidental answer during that period. This is a correction for a mistaken click or keypress, not a rollback system.

Once the agent receives the answer, undo cannot reverse the command it runs or the files it edits. Reversing a change belongs in the project’s normal development workflow.

Keep a route back to the agent

Sometimes the card is enough. Sometimes you need the rest of the conversation, the repository or a command’s full output. Returning to the original app is part of a good review loop.

Perch steps aside when you are in the session’s app and hands held requests back after your chosen wait. Answering stays off until you enable it. Those boundaries leave the decision with the person and the work with the agent.

A small checklist before responding

• Is this the intended project and session?
• Does the action fit the task’s scope?
• Do I have enough context to decide here?
• Am I authorizing the action I actually mean to authorize?