How should a permission-first AI workflow travel across Claude Code, Antigravity, and Codex?
AI coding agents are getting more capable, but my practical problem was keeping intent, approval, and verification consistent when switching between different hosts.
I built Full Stack HQ, an MIT-licensed open-source engineering workflow with:
One shared source of truth for agents, skills, workflows, and rules
Native adapters for Claude Code, Google Antigravity, and OpenAI Codex
Read-only environment diagnostics
Installer smoke tests
Windows and Ubuntu CI verification
The goal is not to create another AI coding host. It is to make the engineering workflow portable between the hosts people already use.
Repository:
https://github.com/sabahattink/antigravity-fullstack-hq
Release:
https://github.com/sabahattink/antigravity-fullstack-hq/releases/tag/v1.2.0
For people using multiple AI coding tools: what breaks most often in your workflow — shared rules, permissions, skills, or verification?
Good write-up. What would you do differently if you started again?
Good write-up. What would you do differently if you started again?
Good write-up. What would you do differently if you started again?
The portability thesis is clear, but which inconsistency is actually forcing users to change their workflow today: rules, permissions, skills, or verification?
Permissions are the forcing function; verification is the consequence that makes the difference visible.
Rules affect context, and skills affect packaging. Both create friction, but users can usually translate them once. Permissions change the actual operating procedure: when an agent must ask, what it may touch, whether shell, browser, or filesystem actions are trusted, and how approval state persists.
That is where Claude Code, Antigravity, and Codex diverge enough to make users change their workflow. Verification comes next: each host reports completion differently, so a portable workflow needs a shared evidence contract, not just portable commands.
So the priority should be:
Rules and skills should be delivery layers—not competing sources of truth.
That’s a much clearer picture of where the workflow actually breaks. What’s the best email to reach you on?