Agent Constitution — AI skill for Claude Code

Agentic SDLC — a governance framework for production systems built by AI agents

A working skeleton for running **multiple AI coding agents against a critical production codebase** without losing control of what is true, what was decided, and what has actually been verified.

This is not a prompt collection and not an agent framework. It is the **governance layer** that sits above whichever agents you use — the documents, contracts, tests, and CI gates that make agent output reviewable, attributable, and safe to merge.

Install Stage 1 into an existing repo

git clone https://github.com/metafive-ai/agent-constitution.git
python agent-constitution/scripts/bootstrap.py --dest /path/to/your/repo

That copies the constitution, decision log, issue templates, and scripts. Existing files are left alone unless you pass `--force`. Then:

1. Fill every in CLAUDE.md. Keep it under ~150 lines.
2. Start docs/DECISIONS.md with your next real decision. Do not backfill history.
3. gh label create truth-gap --color B60205 --description "Two sources of truth disagree"
4. python scripts/session_brief.py

`--stage 2|3|4` adds the later layers. Read [`ADOPTION.md`](ADOPTION.md) before turning CI gates on. A solo project with one agent and reversible changes should skip this repo.

---

The problem it solves

A single agent on a toy repo needs no governance. The failure modes appear at the intersection of three things:

1. Production consequences — a wrong change costs money, data, or trust, and the worst ones fail silently.
2. Multiple agents — different models, different vendors, different sessions, no shared memory, each arriving with full confidence and zero context.
3. Continuous change — the system moves faster than any human can personally verify.

At that intersection, four things break in a predictable order:

Failure	What it looks like
Context rot	Hand-maintained docs drift from reality.