multiagent-protocol

What it does

• L1 pre-merge gate — 5 conditions a PR must satisfy before merge.
• L2 post-merge re-validation — detect real regressions on main and file an incident with the revert command; ignore infra-only failures (Actions outages).
• L3 race-guard — re-check the base just before merging.
• L4 identity gate — validate Agent-* commit trailers against your registry.
• L5 break-glass auditor — surface `[break-glass-*]` commits + demand ADR within 24h.
• Decision Inbox — Quadrant D (irreversible + critical) PRs reach you as GitHub Issues with a structured ballot.

Why "multiagent" not "Claude" / "Codex" / etc.

The protocol treats every agent equally. Adding a new vendor takes one line in config/agent_registry.yml. Identity is enforced through commit trailers, not API endpoints.

What it is NOT

• Not a CI/CD system — bring your own tests.
• Not a code reviewer — it routes to you or auto-approves, never opines on diff quality.
• Not a replacement for GitHub Pro's branch protection if you can pay.
• Not a multi-tenant SaaS — each operator runs their own copy.

Documentation

• Quick start (15 min)
• Architecture — the 4-module bot
• Four-quadrant autonomy classifier
• Five-tier file taxonomy
• Decision Inbox
• Break-glass
• Skills plugin interface
• Mirror cascade
• General preferences (built-in defaults)
• Writing your own skill