A glowing 3D replay of AI coding sessions leaves HN asking what it's for
Mindwalk renders a Claude Code or Codex session as a night map of the repository: files glow amber where the agent edited, white where it merely read, and a scrubber replays the whole run frame by frame. The Show HN pulled "Xerox PARC days" praise from one reply and a blunter question from others: who has time to watch a light show when the diff already says what changed?
Mindwalk replays a coding agent's session (Claude Code or Codex) as a 3D city built from your own repo. Files sit dark until touched, then light up moss green for a glance, white for a read, amber for an edit, while a timeline scrubs through the whole run.
One reply called it 'Xerox PARC days' for agent interfaces. A rival builder said he skips the visuals entirely, tracking the same file changes on a binlog in 3-5 milliseconds. A third commenter asked why anyone would rewatch a session instead of just reading the diff.
Two artifacts do the work: a trace of the agent, a map of the repo
Mindwalk is a Go and TypeScript tool, posted on GitHub by developer cosmtrek, that turns a coding-agent's session log into an explorable 3D scene of the codebase it worked in. Before it renders anything it builds two artifacts: a trace, the session's reads, edits and tool calls normalized into one format by adapters written for Claude Code and Codex logs, and a citymap, a deterministic layout of the repository so the same codebase produces the same shape across different sessions. A local Go server merges the two and serves a React/Three.js frontend; the README is explicit that session data never leaves the machine.
Amber for edits, white for reads, moss green for a glance
The result reads like a night map of the repo. Files stay dark until the agent touches them, then light up by what happened to them: moss green for a file the agent only glanced at, white for one it read closely, amber for one it actually edited. A scrubber replays the session step by step, showing observation passes in cool tones and mutations in warm ones, with markers on the timeline for context compactions and subagent launches. The pitch, per the Show HN thread, is that a session's shape becomes legible at a glance: whether the agent's footprint matches the task it was given, or wandered off it.
"Xerox PARC days," then a rebuttal three replies down
The first reaction staked out the strongest possible claim: that this isn't a demo toy but a preview of how people will eventually work with agents at all.
"I'm becoming convinced the optimal UI to engage with agents, long term is going to be something spatial... I really feel what you've made might be Xerox PARC days in terms of metaphor maturity" — an HN member since 2012 with close to 4,000 karma
A few replies later, a sitcom reference turned into an actual objection.
"Yes, its cool to see and watch but it seems to be adding more friction than reducing it. Like who is going to spend that much time watching what their agents did when there are far quicker and efficient ways of scanning through changes" — an HN member since 2025 with a bit over 600 karma
Between those two, several commenters, including one with a bare one-point account, asked the same plain question without an answer in the thread: what's the use case, beyond looking good?
Two other builders had already shipped something adjacent
The sharpest engagement came from people who'd solved a version of the same problem their own way. One reply had already tracked agent file activity without a renderer at all:
"I started tracking file changes across sessions with file hunk method on a binlog... not for visual, but operational for the harness I use... all file change lookups are 3-5ms" — the builder of a similar session-tracking tool, haicue-brew
Another builder showed up with a near-identical project already running, glyph3d.dev, and offered to compare notes on rendering:
"I've built an adjacent 3D visualizer that might complement the rendering portion of your system by providing actual glyph-level rendering of each file instead of block-based representations... I have a similar 'trace' feature that shows files pre/post edit from CC" — the creator of a similar 3D code visualizer, glyph3d.dev
The rough edges showed up in smaller reports rather than arguments: the 30-second demo link 404'd for one visitor, another found the tree view empty once the local project it had recorded no longer existed on disk, and a third wanted history merged across agent sessions instead of siloed per run.
The unbuilt feature that would settle the argument
No one in the thread filled the gap the skeptics pointed at. The closest anyone came was a proposal to compare two models, or a hundred runs of the same model, on the identical task, spatially and side by side, to see how much an agent's path actually varies. That's a different tool than the one that shipped: Mindwalk today replays one run at a time, well, and leaves the multi-run comparison as the feature that would turn the pretty picture into a diagnostic.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
An engineer designed a RISC-V game console's CPU, GPU, and PCB from scratch
Luke Wren spent his spare time writing his own RISC-V processor core, graphics pipeline, and PCB layout, then fit the whole console onto a low-end FPGA with room left over. By day he designs chips at Raspberry Pi; RISCBoy is what he builds after hours.