Luna Maxxing with Antigravity
Purpose
Use Antigravity for bounded implementation, Luna Max for state coordination and routine review, and Sol only for bounded high-value decisions or takeover. Optimize model-weighted subscription consumption, not raw token count. Git, source, and independently run tests are authoritative; worker summaries are not.
Entry gate
Start only when the user has supplied or identified a human-approved specification and explicitly asked to run this workflow. Specification discovery, requirement approval, commit, integration, push, deployment, production access, and destructive changes remain separate authorization gates.
Read these references before acting:
- Always read references/workflow.md.
- Read references/contracts.md before creating artifacts or task prompts.
- Read references/codex-failover.md before creating Luna or Sol tasks.
- Read references/antigravity-rounds.md before any Antigravity run.
Model roles
| Role | Model | Boundary |
|---|---|---|
| Coordinator | Luna Max | State transitions, evidence integrity, routing, counters, and authority gates only; no substantive code or architecture review |
| Planner | Sol xhigh | One initial phase split and implementation contracts, unless already approved |
| Phase Reviewer | Luna Max | Reused across revision rounds for one phase; performs substantive routine review |
| Acceptance Auditor | Fresh Luna Max task | Independent acceptance check after an ACCEPT recommendation |
| Advisor | Sol xhigh | One explicit difficult decision with evidence; read-only |
| Finisher | Sol xhigh | Takes over only after the objective stopping policy |
| Implementer | Antigravity | Bounded edits and scoped verification only |
The invoking task must not silently substitute models. Create user-visible Codex tasks with explicit model and reasoning settings when a role requires another model. Record the returned task identity and runtime evidence. If the required model is unavailable, stop or disclose and obtain approval for a substitute.
Decisions
Every review returns exactly one decision:
ACCEPT, REVISE, EVIDENCE_REQUIRED, ESCALATE_SOL,
ESCALATE_HUMAN, CODEX_TAKEOVER, or STOP_SAFETY.
Only ACCEPT confirmed by a fresh Acceptance Auditor can move a phase to
approval. REVISE keeps the same phase scope and physical worktree. Evidence
collection does not authorize edits. No concurrent writers may target one
worktree.
Non-negotiable gates
- Never accept a walkthrough or claimed test result without Git and runner-owned test evidence.
- Never widen allowed files, dependencies, APIs, migrations, permissions, or acceptance criteria inside a revision round.
- Never bypass Antigravity permissions or use a paid API-key fallback.
- Every ownership transition must use the complete runner-owned workspace-state capture. Never reimplement the Git-pointer hash.
- Acquire a task-scoped writer lease only after path, HEAD, contract hash, state ID, task ID, model, and effort all match.
- Release the lease after parent-runner reconciliation and before creating the fresh Acceptance Auditor.
- Never commit, merge, push, deploy, publish, or clean worktrees until the user explicitly approves that action.
- Stop after the workflow limits in
workflow.md; do not continue an unproductive loop merely because the hard cap has not been reached.