Mindwalk
GitHub
New tool called Mindwalk turns your coding agent's session logs into a glowing 3D map you can watch play out. It shows exactly what the AI actually looked at, read, and edited — not just what the transcript says.
Based on reporting by GitHub — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anyone who has stared at a raw JSONL log from Claude Code or Codex knows the frustration: you get a wall of text describing actions, but almost no sense of the agent's actual mental model of your repo. Did it wander into three unrelated modules before touching the right file? Did its edits actually match the scope you asked for? Mindwalk, a new open-source Go tool from developer Ricko Yu, tries to answer that by turning session logs into something you can literally watch.
The core trick is visual: your repository becomes a night map, rendered as either a radial tree or a treemap, and the agent's session replays as light moving across it. Files the agent merely glanced at glow moss green, ones it read turn moonlight blue, and anything it actually edited lights up warm amber. Untouched code stays dark. Files the agent touched but that have since vanished from the repo linger as ghostly wireframes, which is a small but clever touch. A playback deck lets you scrub through the session like a video timeline, with observation events colored cool and mutation events colored warm, so the moments where the agent actually started changing code jump out instantly.
Mindwalk ships as a single Go binary, installable via a shell script or buildable from source with make. Run it with no arguments and it scans your ~/.claude/projects and ~/.codex/sessions directories, spins up a local server on a random port, and opens a browser automatically. There's a repo-only mode too, mindwalk map, that renders a codebase's structure with height mapped to lines of code, no session log required.
The privacy angle is worth dwelling on because it's handled thoughtfully. Viewing sessions is entirely local — nothing leaves your machine. The one exception is an optional evaluation feature, where you can ask your own claude or codex CLI to judge a session's trajectory for things like wandering, scope creep, or missed verification steps. Even then, only a summary — task wording, file paths, one-line event digests — gets sent to the model behind your own account, and the judging subprocess runs sealed off from tools, MCP servers, and your other settings.
Under the hood, Yu split the system into three deliberately separate pieces: a trace (the normalized event stream), a citymap (a deterministic repo layout so replays stay comparable across sessions), and a report (the judge's findings, always rolled up mechanically rather than left to LLM whim). That separation, enforced through Go package boundaries, is the kind of architectural discipline that suggests this project is built to be extended rather than treated as a one-off demo.
My take — AI-written commentary, not fact-checked reporting
This is the sort of tool that should have existed the moment agentic coding CLIs became normal — we've been debugging AI behavior with grep and vibes for way too long. I like that Yu kept the LLM judging strictly opt-in and sandboxed; it's a rare example of an AI tool treating 'send my code to a model' as a decision the user makes on purpose, not a default. More agent tooling needs this kind of visibility-first, privacy-second-guessing design instead of just trusting the black box.
Read more about this at: GitHub
