mindwalk

Mindwalk ingests the JSONL session logs produced by coding agents and renders them as a 3D spatial replay — the repository becomes a map, and the agent’s activity plays back as movement through it. The core workflow is local: install the Go binary via the provided script or build from source, point it at a log file, and the tool produces an interactive visualization showing which files were read, which were written, and in what sequence the agent explored before committing to action.

The differentiating design choice is the spatial metaphor. Rather than a flat timeline or a raw log diff, Mindwalk draws the repository as a night-city map and animates the session as light traveling through it. The vendor’s own framing — ‘where it explored before it acted, whether its footprint matched the scope you had in mind’ — captures what this unlocks: the gap between what an agent did and what you assumed it did becomes a geometric shape you can point at, not an inference you have to argue from line numbers.

Mindwalk fits a specific moment in the debugging workflow: after a session completes, before you decide whether to ship the output or re-prompt. It is a human review aid, not an automated gate. There is no API, no webhook, and no integration surface for CI pipelines. Teams that want anomaly detection or automated scope-boundary enforcement need a separate layer — Mindwalk does not provide it and the project does not describe plans to add it.

The tool is MIT-licensed and written in Go, with build instructions and a release script in the repository. Log support targets Claude Code and Codex session formats as described in the README. The schema directory in the repository suggests log format expectations are explicit, which matters if your agent produces a non-standard JSONL structure — the docs describe the expected format and deviations will require preprocessing.