Agent Replay: a calmer way to review code

01A review you can focus on0:00

Narrator: I'm Khalil, and this is Maya. Open a sizeable pull request: files everywhere, a wall of red and green, and a limited amount of afternoon.

Maya: You're reviewing the size of the review before you've read any code.

Narrator: The interface makes it hard to focus. Changes are condensed into fragments, spread across files. You have to reconstruct the original function, the relationships, and why the agent changed any of it.

Maya: While keeping a mental list of everything you haven't reviewed yet. That list is remarkably talkative.

Narrator: That's why we built Agent Replay. Lines arrive and disappear smoothly, with the agent's recorded explanation beside the change. Your attention stays on what's happening now.

Maya: Pause there. Read that function. Then move on when you're ready.

02One change, with its context0:45

Maya: Start with the launch command in the article. You'll need Node twenty or newer, and Git. It opens a local player for existing Claude Code, Codex, and Gemini sessions.

Narrator: Pick a project and a session. If the project isn't listed, use Open folder in the project picker and point it at the repository.

Maya: That finds the folder. You'll still need a supported session log or a saved replay for the session itself.

Narrator: Space plays or pauses. The right arrow finishes a change or advances one step; the left arrow goes back. N jumps between review notes.

Maya: And you can keep the file tree open. You always have somewhere to look for which file you're in. Switch to File to read the surrounding function, or step back to see its earlier state.

Narrator: Our little payments demo had an invoice total one cent off. At the fix, a note points out the rounding assumptions. That's the moment to stop and check whether those assumptions fit.

Maya: The code has already been written. It can survive you reading it.

03What did the checks actually show?1:47

Narrator: Then open Evidence. It lists recorded checks and their results, and flags things worth looking at: deleted tests, a failure hidden by a shell pipe, or changes made after a passing check.

Maya: Our demo flags two deleted test files. Should we be worried?

Narrator: They were temporary probes. One ran the new tests against the old implementation to make sure they caught the bug. Later, the final test suite passed.

Maya: So the sequence matters. A deleted test deserves a look; here, we can see what it was doing.

Narrator: And the panel reads the recorded commands and output. It doesn't rerun the tests or decide whether the code is correct. That's still our job.

04Two ways to learn the repository2:28

Maya: Sometimes you can follow the change perfectly and still have no idea how the rest of the repository fits together.

Narrator: There are two ways in. Replay the Git history, commit by commit. You can start with the first commit, or choose a later revision and have the existing files introduced before following the changes.

Maya: The default history is limited to the latest three hundred commits. For the full history, choose the first commit as your starting revision.

Narrator: Or ask your agent for concise tutorials covering the repository as it is now. The article shows how to install the teaching skills for Claude Code.

Maya: Give it a useful brief: start with the data types, follow one request, explain the important decisions, and keep going until the repository is covered. Keep the boilerplate brief. We have suffered enough package files.

Narrator: Open the course under Learn. Real source arrives in small pieces with explanations beside the relevant lines. Pause, inspect the function, and carry on at your own pace.

Maya: The source belongs to the course's pinned revision. The explanations are written by an agent, so read those critically too.

05Share a calmer review3:38

Narrator: You can export the replay as one HTML file that opens offline, and share it alongside the pull request.

Maya: Check it before sharing: it includes code and session context. Then the next reviewer can follow the same steps.

Narrator: Agent Replay is open source. The GitHub link and the commands are in the article. Try it on one small change you'd otherwise skim.

Maya: One change, and permission to take your time.

Narrator: The agent can keep the speed. We get to keep a little understanding.

06Until next time4:08

Narrator: Thanks for listening. Follow along for more articles, experiments, and things that worked on my machine.

Maya: Some of them might even work on yours.

Narrator: No promises. Have a lovely day. Go make something.

Maya: Preferably a cup of tea first. Bye!