Watching Agents Wander Your Code at Night

What it does

Mindwalk replays coding-agent sessions from Claude Code, Codex, or pi as a glowing 3D map of your repository. Files light up in moss green, moonlight blue, or warm amber depending on whether the agent merely saw, read, or edited them, while untouched files stay dark. A built-in playback deck lets you scrub through the session, jump to edits, errors, or subagent launches, and even export the replay to a .webm entirely in the browser.

The interesting bit

The tool separates concerns into three rigid layers—a normalized trace, a deterministic citymap layout, and an optional LLM judge report—so the same repository always renders identically and verdicts are rolled up mechanically rather than left to the model’s discretion. The optional evaluation is also unusually careful: it only calls your local claude or codex CLI when you explicitly run it, sends only that session’s summary, and seals the judge subprocess from tools, MCP servers, and persistence.

Key highlights

• Supports Claude Code, Codex, and pi logs out of the box; adapters normalize each into a comparable event stream.
• Deterministic 3D layout means you can compare multiple sessions on the same repo map.
• Subagent traces get their own replay lens on the same map, so you can inspect nested agent runs.
• Evaluation drafts task-specific scorecards from your own request wording, with every finding anchored to a clickable timeline event.
• Runs fully offline; the only network traffic is the opt-in evaluation step using your own local agent CLI.

Verdict

Grab it if you need to audit where your agent actually went before it edited—or if you suspect it never looked at the right files. Ignore it if you don’t use the supported agents or if you prefer your logs raw and your repo maps imaginary.