This Tool Exposes Your AI Coder

For two years, we've blindly handed codebases to AI agents, then reviewed the fallout via git diffs. This approach is a joke. Traditional logs only capture final outputs, never the actual thought process. They completely hide which files an agent explored, the dead ends it hit, or if it quietly strayed from its intended scope. We see the diff, but not the journey.

This invisible context cripples any true review of AI-generated code. You're presented with a result, not the underlying reasoning. This means you cannot properly assess the quality of its strategy—did it explore enough? did it understand the job? Without that insight, you're merely auditing the output, not the intelligence behind it.

Mindwalk tears down this opacity. Its central premise is making the hidden process visible, rendering your entire codebase as a 3D city. Every file becomes a building; taller structures signify more code. You then press play on your agent's session logs—Mindwalk reads existing Claude Code and Codex logs from your disk—and watch the lights. Files glow green when glanced, blue when read, and amber when edited. This transforms abstract log files into a tangible, observable journey, revealing the agent's actual path through your project. Finally, a way to visualize the invisible context.

Mindwalk renders your entire codebase as a dark 3D city. Each file becomes a building; its height scales directly with line count. This gives you an instant architectural overview, revealing project complexity and file prominence before the agent even touches a line.

Agent actions decode via a simple, powerful color language. Green means the agent merely glanced at a file, a quick pass. Blue indicates it read the contents, engaging with the code. Amber signifies an edit, a critical change to the file. This visual shorthand instantly reveals the agent's focus and intent.

The interactive timeline is the core of this observability. Scrub back and forth to literally trace the agent's path through the code. Clear markers delineate key events:

• Sub-agent spin-ups
• Context window compaction
• User interruptions

This lets you watch the agent's journey unfold, understanding its decision points and any external influences. No more guessing if it strayed or stayed on task. This is about seeing the actual thinking, not just the output. It exposes the AI's mind map, critical for debugging and optimizing agent behavior.

Mindwalk doesn't just visualize agent activity; it critiques it. Run the analyze command to deploy a local LLM, like Claude or Codex, as a judge. This isn't a linter; it's a meta-review system, auditing the AI agent's performance against strategic benchmarks, not just syntax.

This AI judge answers critical questions about agent execution. Did the agent explore enough before editing? Did it stay in scope or wander into unrelated files? Did it bother to verify its own work before committing? These are the real metrics of an agent's competence, exposing its planning and adherence.

Here's the kicker: every finding in the judge's report is a clickable link. Selecting a critique instantly navigates the 3D timeline to that precise moment. This transforms abstract feedback into an actionable review tool, letting you pinpoint exactly where an agent strayed or excelled.

No more guessing why an agent failed a task. This feature provides a direct conduit to the agent's decision points, invaluable for refining prompts, comparing different model behaviors, or simply understanding complex changes. It's the missing piece for truly debugging AI-driven development. For further technical deep dives and contributions, check out the project's GitHub - cosmtrek/mindwalk: A visualization tool that replays coding-agent sessions on a 3D map of your codebase..

Mindwalk's initial reception on Hacker News highlighted a valid critique: for a two-line fix, a plain git diff remains vastly more efficient. A 3D fly-through to inspect minor changes is overkill, slowing down workflows rather than accelerating them. Mindwalk isn't designed for every trivial modification.

Its true utility emerges in high-complexity scenarios, where traditional logs fail to convey the agent's internal state. Think:

• Comparing divergent approaches from Claude Code versus Codex on identical tasks
• Dissecting massive agent-led refactors
• Debugging elusive failures in complex agentic workflows

Here, Mindwalk provides unparalleled insight into exploration patterns and decision paths.

Mindwalk doesn't replace existing code review tools. Instead, it establishes a new category for understanding agentic systems, offering visibility into the AI's "mind map" that was previously invisible. It’s an early, yet crucial, swing at solving the persistent AI transparency problem in software development. This is a nascent but vital tool for serious agent builders.

What is Mindwalk?

Mindwalk is a developer tool that visualizes a codebase as an interactive 3D map and replays an AI coding agent's session, showing exactly how it explored and edited the code.

How does Mindwalk show what an AI agent is doing?

It uses a simple color language on file "buildings": green for a quick glance, blue for reading the file's contents, and amber for making an edit, all on a replayable timeline.

What AI agents does Mindwalk support?

As of its initial release (v0.3), Mindwalk is designed to read and replay session logs from AI coding tools like Claude Code and Codex.

Is Mindwalk practical or just a visual gimmick?

While visually impressive, its core purpose is practical: to understand an AI agent's reasoning, review its strategy, and identify inefficiencies that a standard text diff would miss, especially in large-scale changes.