Review Claude Code changes with Play-by-Play

Start with a finished task

A useful example is a refactor that changes a component, its styles and its tests. Wait for the turn to finish, then review what actually changed before asking Claude to continue.

Run these inside Claude Code. Add the marketplace once, then install the mod. Skip the marketplace command if you already added it.

Use /plugin to confirm it is enabled. If you installed from a shell while the session was open, run /reload-plugins in that session.

Review one file at a time

Run /replay after a turn that changed files. Start with the scoreboard: which files changed, how much changed, and which file contains the largest edit?

• Does the change match the task you requested?
• Did it affect behavior or files outside that task?
• Is there a test or another concrete check for the behavior that changed?

While the pane has keyboard focus, n and p move between steps. Keys 1-9 jump to the first change of a numbered file. Close with q or Escape. Use /replay 2 to revisit the previous recorded changed turn.

For the final commit review, also inspect your working tree with your usual Git tools. Replay is not a complete or permanent record of the repository.

When changes are missing or unexpected

Shell tracking needs a Git repository. Ignored files are omitted, and slow or large snapshots can pause shell tracking while file-tool tracking continues.

An editor save made while a command is running can appear as a shell change. Background commands are compared when they enter the background, so later writes may not be represented. These limits matter when attributing an edit.

Access and compatibility

The reviewed version keeps up to ten changed turns in session memory. It reads edit-event contents and invokes read-only Git commands; it makes no network requests or model calls of its own. Recorded changes disappear when the session ends.

Interface and keyboard support depend on the Claude Code surface. Check the Play-by-Play detail page for the reviewed compatibility and complete limits.