grok-coding-observatory

grok-coding-observatory · Guide

How to review AI-generated code changes before you commit

AI agents write a lot of code quickly, and reviewing a big final diff is hard. Here is a simple workflow that makes it manageable, using git and the free grok-coding-observatory. 

# in the folder of your repo:
npx -y github:Eeliya/grok-coding-observatory   # open http://localhost:4477
Demo: an agent edits a small project; each edit is typed into the editor live, the session timeline on the right grows, and the agent status chip goes from working to done
A scripted agent at work: each edit replays as it lands, the Session timeline grows and the status chip goes from working to done.

A five-step review workflow

  1. Start from a clean tree. Commit or stash your own work first, or give the agent its own branch. Then everything that differs from HEAD is the agent's work, and undoing it is one git restore away.
  2. Watch it as it happens. Run npx -y github:Eeliya/grok-coding-observatory in the repo and open http://localhost:4477. Each saved edit replays live, so you notice a wrong direction after one file instead of twenty.
  3. Replay what surprised you. The Session timeline keeps every edit in order. Click one to replay exactly that change, including intermediate versions the final diff no longer shows.
  4. Check the whole change against HEAD. The sidebar lists every changed and untracked file; click one for a side-by-side diff. Unseen dots and the since last look view show only what is new since you last checked.
  5. Commit what you understand. Stage with git add -p, drop the rest with git restore, and commit. Commits the agent made itself can be replayed file by file with the « / » buttons.

What to look for

  • Files outside the task: config, CI workflows, unrelated modules.
  • Deleted or weakened tests, skipped assertions, new TODOs.
  • New dependencies and lock file churn.
  • Hard-coded secrets, URLs or credentials.
  • Error handling that swallows errors to make a test pass.
  • Large rewrites where a small fix was asked for.

Ask the agent to report its progress too. With a few lines in AGENTS.md or CLAUDE.md (see the agent status protocol), the observatory shows what it says it is doing, like "Running tests", next to what it actually changes, and markers in the timeline show when it said it was done.

FAQ

Can the observatory accept or reject changes?

No. It is deliberately read-only and never writes to your work tree. Use git for that: git add -p to stage what you keep, git restore to drop the rest.

Which AI coding tools does this work with?

Any tool that edits files in a git repo on your machine, for example Claude Code, Codex CLI, Cursor, Aider or GitHub Copilot's agent mode.

Does it review the code for me?

No, it doesn't judge the code. It makes changes easy to see and replay so that you can review them faster.

What does it need?

Node.js 22.6 or newer, git and a modern browser. On Windows, run it inside WSL.

More guides

See all features View on GitHub