Zat.Env

zat.env

Claude Code on an iPhone (ShellFish) with tmux, connected via Tailscale SSH to a Linux server in Germany

Minimal harness for verification-first autonomous coding with Claude Code. Verification quality, not prompt engineering, determines the ceiling on what agents can build. This repo implements the thinnest layer that matters: specs as the control plane, adversarial review as the verification loop, persistent artifacts as inter-session memory, and a pre-push gate that blocks unreviewed code from leaving the machine.

Clone this repo and run `zat.env-install.sh` to get spec-driven development, adversarial code review with builder/verifier separation, security auditing, architecture review, test strategy review, and a GitHub PR workflow as Claude Code skills, with a pre-push hook that gates `git push` on passing review. Optional multi-model reviewers (OpenAI, Google, local GPU) provide independent second opinions.

`zat.env-install.sh` is the only requirement: it wires skills, hooks, and conventions into any machine with git, jq, and Claude Code. `hw-bootstrap.sh` is optional and specific to [my always-on dev box](#current-hardware-hetzner-gex44) (a hosted dedicated bare-metal Linux server for GPU workloads with secure remote access). Skills are Markdown prompt files, hooks are bash scripts, conventions are plain text. Nothing is invasive: skills are opt-in (invoke them when you want them), and the pre-push gate has a bypass. Companion writing at [agent-hypervisor.ai](https://agent-hypervisor.ai).

If you're coming from [The Bitter Lesson of Agentic Coding](https://agent-hypervisor.ai/posts/bitter-lesson-of-agentic-coding/), this is the repo. The harness is deliberately minimal because the bitter lesson says it should be.

**Where this is headed.** Today a human is in the loop, reviewing outcomes and writing specs, and the harn