Codex, Cursor, Amp

Reviewing what Codex, Cursor or Amp changed

Whichever coding agent you use, the review problem is the same: fast changes, long summaries, decisions you never made. How to see them on the lines instead.

Every coding agent has the same shape of problem at the end of a task. The work is done, it is mostly right, and the explanation is a block of prose with file names in it.

The agent is not the problem

Codex, Cursor and Amp are all good at writing code. The gap is on the other side: you need to understand what was decided before you ship it, and prose is a poor way to carry that.

One format for any agent

deck works with any agent that can run a shell command. The agent writes a deck with four commands: start it, open it, add a group per claim, seal it. Each group says one thing and points at the lines that show it. deck checks that every file and line it points at exists before it accepts the group.

On your side, a bar appears. You open the deck when you are ready, click through the claims, see the code lit beside each one, and comment where you disagree. Your review goes back to the same agent, which carries on from there.

What changes in practice

• Reviews get shorter. You look at the decisions, not the formatting.
• Disagreements get earlier. You push back on the plan before the code exists.
• Context survives. The agent that did the work is the one answering your questions.

Setup

Install deck and run deck setup. It finds the agents on your machine and installs a short skill into each, so they know when a deck is the better answer. See which agents deck works with.