Agents propose.
A human merges.
CI decides.

A GitHub template that runs your software delivery lifecycle with autonomous agents — every change reviewed by two model families, gated by 23 checks, and merged only by you.

Use the template The seven ideas (10 min) See it live 
issue opened ──▶ steward triages ──▶ PR opened ──▶ two reviews
                                                       │
      ┌────────────────────────────────────────────────┘
      ▼
23-gate gauntlet ──▶ YOU merge ──▶ filing agent verifies the fix landed

Four ways in

New repo Use this template, then tools/adopt.sh — one resumable command, nothing happens without your yes. First merged agent loop in ~30 minutes.
Existing repo tools/upgrade.sh --install grafts the harness onto your repo — never overwrites your files, and future template releases become a computable three-way merge.
Your agent does it Tell Claude Code or Codex: "read ONBOARDING.md and adopt this." Ready-made answer profiles included — a platform team ships one, every team adopts in one command.
Just watching Open in a Codespace: a three-minute, zero-credential tour — ~650 tests, the adoption map, a dry-run agent command. Or set mode: observe: the whole fleet runs report-only, and pushing is mechanically impossible.

What you get

A steward on your issues Triages every new issue, answers mentions, writes fixes as pull requests. It never merges — that button is yours, by design.
Reviews that argue Two reviews from different model families, plus a referee that settles their disagreements against the code. Verdicts are advice; you overrule.
A gauntlet with your numbers 23 gates — tests, coverage, mutation, architecture, secrets, a11y — with floors calibrated to your codebase, then ratcheted: they only move up.
A fleet that remembers Eleven scheduled agents (all shipped off, enabled one at a time) and a second brain: agents propose distilled lessons, a human merges them, every future session reads them first.
Guards on the machine itself ~650 tests cover the agents' own plumbing; 129 incident-earned lessons are pinned so they can never be lost quietly.
No silent failures, no vendor lock Everything optional degrades by announcing itself. Models are addressed by role, never vendor; works on a flat CLI subscription, behind corporate proxies, on GitHub Enterprise.

The seven ideas underneath

  1. Branch protection as epistemics — write permission, not trust, separates instruction from history.
  2. Three-tier agent memory — history, distilled knowledge, and steering, each with its own write permission.
  3. Degrade visibly, never silently — a dead agent and a healthy one must never look the same.
  4. The ratchet — floors only move up, and suppression counts as lowering.
  5. Test the machine that builds the software — not just the software.
  6. The adversary never decides — different model families review; a referee rules; a human merges.
  7. Your baseline, never someone else's — floors ship unarmed and loud until measured against your code.

Each one earned by an incident and enforced by a test or a permission — the ten-minute read, including how to steal any one of them without adopting the rest.

Honestly, when not to use it

Solo prototyping doesn't need a gauntlet. If you want unattended merges, this is deliberately the wrong tool. If you only want dependency bumps or spec discipline, lighter tools exist. This is for teams who want agents doing real work and a mechanical reason to trust every change that lands.