cosmtrek/mindwalk

What it solves

Mindwalk transforms raw, linear session logs from coding agents into a visual, 3D map of a codebase. It solves the problem of understanding how an agent explored and modified a repository—identifying which files were relevant, where it wandered, and whether its footprint matched the intended scope—which is difficult to discern from reading JSONL logs.

How it works

Mindwalk creates a deterministic "citymap" of the repository and overlays a playback of the agent's session. It uses adapters to normalize logs from different agents (like Claude Code, Codex, and pi) into a stream of file-touch events.

Key visual elements include:

• Terrain and Tree Views: The codebase is rendered as a radial tree or treemap, where glow intensity indicates how often a file was touched.
• Touch States: Files are color-coded by their deepest interaction (e.g., moss green for seen, moonlight blue for read, warm amber for edited).
• Playback Deck: A timeline allows users to scrub through the session, distinguishing between observation (cool colors) and mutation (warm colors).
• Agent Lenses: If the session involved subagents, users can switch "lenses" to replay specific subagent traces on the same map.
• Session Evaluation: An optional feature that uses a local agent CLI to judge the session's trajectory based on process dimensions (exploration, scope, wandering, verification) and a task-specific scorecard.

Who it’s for

Developers and AI agent researchers who need to debug, analyze, and evaluate the performance of coding agents by visualizing their interaction with a codebase.

Highlights

• 3D Codebase Visualization: Visualizes agent activity as light moving through a 3D map of the repo.
• Local-First: The Go binary and UI run locally; viewing sessions sends no data externally.
• Multi-Agent Support: Supports logs from Claude Code, Codex, and pi.
• Lenses for Subagents: Ability to isolate and replay the subagent traces within a complex session.
• Evidence-Anchored Evaluation: LLM-based session reports that link findings directly to timeline events.
• Client-Side Video Export: Export playback sessions as .webm files entirely in the browser.