Mindwalk replays Claude Code, Codex sessions on 3D code maps

A visualization tool for coding-agent sessions is an unglamorous thing to get excited about, and yet Mindwalk is the first one I have seen that answers a question every team using Claude Code or Codex has quietly been asking: what did the agent actually touch, and where. The project, posted on GitHub by Ricko Yu under an MIT license, ships as one Go binary that reads session logs off the local machine and replays them as light moving through a 3D map of the repo.

The design is deliberately simple. Two internal pieces do the work, a Trace layer that normalizes events from agent-specific adapters for Claude Code and Codex, and a Citymap layer that lays out the repository deterministically. A local Go server joins them and serves a React and Three.js frontend. Files are colored by the deepest state they reached during the session, moss green for seen, moon white for read, warm amber for edited, and dark for unvisited. You can flip between a radial tree and a treemap, scrub a timeline that marks context compactions, subagent launches, and user turns, and click any file to see its visit history.

The reason this matters is that agent audit tooling has been the missing piece of the let-the-agent-do-it story. If you cannot see where the agent went, you cannot review its work the way you would review a colleague's pull request. The session logs already contain the data; Mindwalk just makes them inspectable. And because the binary runs locally and, as the README puts it, no session data leaves your machine, it sidesteps the usual compliance objection to uploading traces to a hosted service.

The honest caveat is what a replay cannot tell you. Seeing that an agent edited three files is not the same as seeing whether the edits were correct, and a pretty visualization can create false confidence that busy paths equal good work. The README also does not address how the map holds up on very large monorepos, or whether two runs on the same task can be diffed against each other, which is the obvious next question for anyone trying to evaluate agent quality.

Still, if this style of local replay catches on, expect the vendors to ship their own versions inside Claude Code and Codex before long. The interesting bit is not the graphics, it is that agent runs are becoming a reviewable artifact, and that quietly changes what a code review is.