Mindwalk Maps Your Agent's Session in 3D — And Shows Where Tokens Get Wasted

Seeing What Your Agent Actually Did

Mindwalk is a visualization tool that replays the session logs of coding agents like Claude Code and Codex on a 3D map of your repository. It renders the codebase as a night-time cityscape: files the agent searched, read, and edited glow, while untouched regions stay dark. At a glance you can see how broadly — or how wildly — the agent ranged across your code to complete a task.

It ships as a single Go binary, processes all session data locally (nothing leaves your machine), and offers treemap and terrain views. Files are color-coded by touch state — unvisited, viewed, read, edited — and the playback UI surfaces friction signals like error rates and file-modification volume.

Why Visualization Is a Cost Tool

Token bills are abstract. "You spent $4.20 on this task" tells you nothing about why. Mindwalk-style replay makes the waste visible, and in agent workflows the waste is almost always one of a few recognizable patterns:

• Over-reading — the agent loads dozens of files it never edits, pouring irrelevant code into the context window at input-token cost
• Re-reading — the same file gets pulled into context repeatedly across turns because context was reset or truncated
• Wandering — the agent explores unrelated corners of the repo before finding the right file, a sign the task or context was under-specified
• Thrashing — high error rates and repeated edits to the same file, each retry burning more output tokens

A bright-but-untouched cluster of files on the map is a literal picture of tokens you paid for and did not use.

The Token Math of Over-Reading

Suppose an agent on a $5/M input model reads 40 files averaging 400 lines (~5K tokens each) to make a change that actually only needed 3 of them. That is 200K input tokens ingested, ~185K of them irrelevant. At $5/M, you paid roughly $0.93 in wasted input on a single task — and because that bloated context is resent on every subsequent turn, the waste multiplies across the conversation.

Multiply by hundreds of daily agent runs and over-reading becomes one of the largest silent line items in an AI coding budget. You cannot fix what you cannot see; a session map turns the invisible into an obvious hotspot.

Turning Insight Into Savings

Once replay reveals the pattern, the fixes are concrete:

• Scope context better — if the agent wandered, your prompt or file-attachment strategy was too vague; point it at the right directory
• Tune retrieval — if it over-read, tighten the file-search or RAG step so it pulls fewer, more relevant files
• Manage context resets — if it re-read, adjust when you reset versus continue a session to avoid re-ingesting the same code
• Catch thrashing early — high error-rate signals tell you to intervene or switch models before retries pile up

The Broader Point: Observability Is Cost Control

Mindwalk is one of a growing class of agent-observability tools, and the trend matters for anyone managing spend. As agents get more autonomous, their token usage becomes harder to reason about from a bill alone. The teams that keep costs under control are the ones instrumenting why tokens are spent — file access patterns, retry rates, context growth — not just how many. Local, privacy-preserving tooling like a single Go binary lowers the barrier to adopting that discipline.

The Takeaway

A 3D session replay is more than eye candy — it is a debugging tool for your token budget. Over-reading, re-reading, and thrashing are expensive and invisible until you visualize them. Pair session-level observability with task-level cost modeling to find and eliminate waste. Estimate what a given workflow should cost with our AI coding cost calculator, then use replay to see where reality diverges.