The governance floor for AI coding agents: install, guard, prove. Guards, git hooks, an executable contract and a receipt per run — no hosted control plane, no provider lock-in.
This repository also configures its own agents. See what ai-engineering tells them →
Agent Claude Code
Hostile, isolated code review of a feature checkpoint's changes. Assumes the code is wrong until proven otherwise, checks it against this repo's conventions, architecture rules and decisions, and returns a PASS/FAIL verdict with findings. Read-only. Use as the final gate of an orchestrated checkpoint, or when the user…
Agent Claude Code
Breaks a large feature into small, ordered, gated checkpoints and writes them as JSON to .ai-engineering/workflow/checkpoints/ .json. Checkpoint 1 is the smallest thing that can pass; each later checkpoint is a bit bigger and builds on the one before. Use when the user wants a feature planned, split into steps, or…
Agent Claude Code
The adversarial-reviewer's counterpart in an /ai-adversarial-loop debate. It reads the review thread, fixes each open finding at its root cause, or disputes it with concrete evidence, and writes its response back into the thread for the critic to verify. Use only inside /ai-adversarial-loop.
Agent Claude Code
Compares prototype vs live UI behavior - state mismatches from the capture fingerprints, copy, labels, component states, interactions and data states - from states.json, the prototype HTML and the live source. Read-only. Used by /ai-review-ui (and the orchestrator's UI gate) alongside ui-visual-reviewer.
Agent Claude Code
Compares state-matched screenshot pairs (prototype vs live) for layout, spacing, typography, color and responsive fidelity, judging sizes proportionally. Read-only. Used by /ai-review-ui (and the orchestrator's UI gate) alongside ui-behavior-reviewer.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: