cavekit

Why we included this project

Working in Claude Code on a multi-step change, you've probably hit the moment when the context window resets and the agent forgets what it was doing. Cavekit addresses that with a spec-driven loop built around a single SPEC.md file: you sharpen a vague idea into a spec, build against it, and run a drift report that flags which invariants and tests are now violated. It skips the usual agent swarm and dashboard machinery, so the token cost stays low and everything lives in plain markdown. The backpropagation habit is the interesting part: every test failure gets written back into the spec as an invariant, so the model carries constraints across sessions. For teams that like test-driven discipline but don't want a heavy orchestration stack, this is a lean, opinionated fit.