Decree

An AI orchestrator for spec-driven development. Write specs, run decree process, get working code.

Why Decree

AI coding assistants are powerful but ad hoc. You prompt, you review, you prompt again. Nothing is repeatable, nothing is tracked, and multi-step workflows need constant babysitting.

Decree treats AI work like database migrations. You write spec files describing what you want built. Routines define how AI processes each spec — implement, build, test, fix. Processing runs them in order, one at a time, each building on the last. Everything is logged.

The result: you focus on what to build. The routines handle how.

Quick Start

cargo install decree
decree init          # scaffold project, pick your AI tool

This creates .decree/ with routines, prompts, config, and a router.

Workflow: Spec-Driven Development

1. Write specs

Create migration files in .decree/migrations/, numbered for ordering:

.decree/migrations/
├── 01-auth-system.spec.md
├── 02-user-profiles.spec.md
└── 03-api-endpoints.spec.md

Each spec is markdown with optional YAML frontmatter:

---
routine: develop
---
# Auth System

Implement email/password authentication with session tokens.

## Requirements
- POST /auth/register creates a user
- POST /auth/login returns a session token
- Sessions expire after 24 hours

## Acceptance Criteria
- Registration with duplicate email returns 409
- Invalid credentials return 401
- Expired tokens are rejected

2. Process

decree process

Each spec is processed in order through the assigned routine. The default develop routine invokes your AI tool twice — once to implement, once to verify acceptance criteria. Failed specs retry with prior attempt logs as context.

3. Review

decree status        # see what's been processed
decree log 01        # see execution output for a spec

Blackbox Testing with Specs

Specs work well as blackbox test cases. Define inputs and expected outputs. The routine implements code to satisfy them. You never describe how — only what.

Write specs around observable behavior:

# Markdown Parser

Parse markdown to HTML.

## Acceptance Criteria
- `# Hello` produces `<h1>Hello</h1>`
- `**bold**` produces `<strong>bold</strong>`
- Empty input produces empty output
- Nested lists render correctly

The AI figures out the implementation. The acceptance criteria are the tests.

Routines

Routines are shell scripts in .decree/routines/ that define how work gets done. They receive the spec as a message file and call your AI tool directly.

The default develop routine:

1. Sends the spec to your AI tool for implementation
2. Sends it again for verification against acceptance criteria

The rust-develop routine adds build and test steps:

1. AI implements the spec
2. cargo build --release && cargo test
3. AI reads build/test output and fixes failures

Write your own routines for any workflow — linting passes, documentation generation, image creation, data pipelines. A routine is just a bash script that calls whatever tools you need.

decree routine       # list available routines
decree verify        # check all routine pre-checks pass

Chaining

Routines can write follow-up messages to .decree/outbox/. Decree processes them depth-first before moving to the next migration. This enables multi-step pipelines:

market-analysis → competitive-landscape → financial-model → executive-summary

One spec in, four documents out.

AI Tool Permissions

Your AI tool needs permission to read and write files in the repo. Configure this per-project so routines can operate non-interactively.

For Claude Code, create .claude/settings.local.json:

{
  "permissions": {
    "allow": [
      "Bash(cargo build:*)",
      "Bash(cargo test:*)",
      "Read",
      "Write",
      "Edit",
      "Glob",
      "Grep"
    ]
  }
}

Other tools have similar mechanisms — check your AI tool's docs for non-interactive / headless permissions.

Lifecycle Hooks

Configure hooks in .decree/config.yml for cross-cutting concerns:

hooks:
  beforeEach: git-baseline
  afterEach: git-stash-changes

The built-in git hooks stash a baseline before each spec and checkpoint changes after. Failed specs restore to baseline before retrying. Every attempt is preserved as a named stash.

Daemon & Cron

For recurring work, run the daemon:

decree daemon

It polls .decree/cron/ for scheduled messages and .decree/inbox/ for new work. Cron messages use standard cron syntax in frontmatter:

---
cron: "0 9 * * 1-5"
routine: daily-review
---
Run the morning code review.

Prompts

Interactive prompt templates live in .decree/prompts/. They inject project context — processed migrations, available routines, config — so your AI conversations start informed:

decree prompt migration    # plan next batch of specs
decree prompt routine      # get help writing a new routine

Project Structure

.decree/
├── config.yml          # AI tool config, retries, hooks
├── router.md           # instructions for automatic routine selection
├── processed.md        # tracks completed migrations
├── migrations/         # spec files (your input)
├── routines/           # shell scripts (your workflows)
├── prompts/            # interactive prompt templates
├── cron/               # scheduled messages
├── inbox/              # messages being processed
├── outbox/             # follow-up messages from routines
├── runs/               # execution logs (the audit trail)
└── dead/               # exhausted messages for review