A Google Antigravity, Claude Code, and Codex plugin that teaches AI agents to build Moqui applications the way the Moqui ecosystem builds them: find what the framework already provides, use it, and prove the result — never reinvent, never trust an unverified green.
The Antigravity package is declared in .antigravity-plugin/plugin.json and .agents/plugin.json.
Keep the repository layout intact when using it. Check the package with:
python3 scripts/validate_antigravity_plugin.pySee Antigravity plugin layout and validation for details.
The Codex package is declared in .codex-plugin/plugin.json. Keep the
repository layout intact when installing it: skills use relative references to
the shared assets/, agents/, docs/, and scripts/ directories. The
package can be checked before publishing with:
python3 scripts/validate_codex_plugin.pySee Codex plugin layout and validation for the package contract and troubleshooting notes.
- 17 skills — task discipline for entities, services, screens, integrations, security, FreeMarker, the Maarg Data Manager, deployment, verification, running an agent team, and four role entry points for the Architect, Builder, Business Analyst, and QA Technician.
- 4 agent roles — a Business Analyst (writes requirements as business process stories), an Architect (reviews designs and PRs citation-by-citation against framework code), a Builder (implements one task, never certifies its own work), and a QA Technician (verifies claims adversarially; hunts tests that can never fail).
- 19 reference assets — the knowledge the skills and agents cite: framework pitfalls, authoring philosophy, integration patterns, the HEMP requirements method, agent orchestration, integration testing, and more.
Because coding was never the bottleneck — supervision was. One person feeding context to ten agents and checking every change is not ten engineers; it is one tired person with ten browser tabs. This plugin's agent roles work over a plain GitHub issue board, with mechanical gates between hand-offs — the code must compile, the tests must pass with exact expected counts, the design must use what the framework provides. The human states the goal and judges the result.
Read the story: The Supervision Bottleneck — how one campaign went from requirements to 78 passing tests with no human in the inner loop, and what went wrong on the way (including the green test suite that was lying).
Position statement: what this is and what it builds on · Verified sources: prior-art-agent-teams.md