AI Orchestration — AI skill for Claude Code

AI Orchestration

A small, host-neutral protocol for using Codex, Google Antigravity/Gemini, optional OmniRoute reviewers, and ESRA without turning ordinary engineering work into an agent ceremony.

The repository has one source of truth:

• `protocol/AI_ARCHITECTURE_BOARD.md` — maintenance source for roles, capability modes, review triggers, implementation loop, and the bounded ESRA cycle. The root AGENTS.md retains a compatibility mirror so existing Codex consumers still receive the full policy.

Everything else is an adapter, not a competing protocol:

Surface	What belongs there	Entry point
Codex	Thin always-on instruction and bootstrap prompt	`adapters/codex/AGENTS.md`
Antigravity / Gemini	Narrow independent-review rule	`adapters/antigravity/.agents/rules/orchestration.md`
OmniRoute	Reviewer-routing boundary	`adapters/omniroute/README.md`
ESRA	Trigger boundary for one task-scoped cycle	`adapters/esra/INTEGRATION.md`

The bundled [`esra-antigravity-plugin.zip`](esra-antigravity-plugin.zip) is a separate Antigravity ESRA skill package. It is not the Architecture Board protocol and should not become its duplicate source.

Design boundary

Task
  └─ Primary Agent
       ├─ routine work ───────────────► implement and verify
       ├─ consequential uncertainty ──► targeted Board review
       │                                  └─ advisory synthesis by Primary Agent
       └─ ESRA trigger ───────────────► one bounded ESRA cycle

• A model router chooses a feasible execution or reviewer route.
• The Architecture Board protocol decides when an independent view is useful.
• ESRA supplies a selective reasoning/control loop.
• Codex and Antigravity/Gemini are execution environments.

Use

Keep this reposi