We need a dual-layer, rich metadata, version control system that gives agents reproducibility and gives humans legibility. Every commit has two faces: one the machine can replay, one the human can understand. It ships as an API, a CLI, and a native integration inside Claude Code.

Problems 1.The Lost Prompt 2.The Unreviewed Mountain (of commits) 3.The Regression Nobody Caught

This is how we can build this: ┌─────────────────────────────────────────────────────┐ │ HUMAN LAYER │ │ │ │ What developers see: │ │ • Narrative commit messages (auto-generated) │ │ • Intent diffs ("what was the agent trying to do") │ │ • Risk scores per commit │ │ • Architecture impact summaries │ │ • Standard git clone / push / pull │ │ │ ├─────────────────────────────────────────────────────┤ │ AGENT LAYER │ │ │ │ What the system records: │ │ • Full prompt + system prompt at time of execution │ │ • Model ID + version + hyperparameters │ │ • Tool calls (terminal, browser, file reads) │ │ • Reasoning traces (where available) │ │ • Token usage + cost │ │ • Session graph (which tasks spawned which) │ │ • Eval results (per-commit and per-session) │ │ │ └─────────────────────────────────────────────────────┘