ai-sdlc-kit

A spec-driven, test-first code factory for AI coding agents.

ai-sdlc-kit gives your coding agent an engineering team's workflow. It plans with you, writes the tests before the code, builds to those tests, and has a different model try to break the result — for one ticket or a whole PRD. Works with Claude Code, Cursor, Windsurf, and Gemini CLI.

Install

pip install ai-sdlc-kit          # or: uv tool install ai-sdlc-kit
npm install -g ai-sdlc-kit       # or: npx ai-sdlc-kit install

ai-sdlc install --agent claude   # cursor | windsurf | gemini | all

install adds the skills, five stage subagents, the Bug Ledger MCP config, and — for Claude Code — two guard hooks. Existing config files are merged, never overwritten.

Quick start

/sdlc-init              once per project: stack, code rules, git workflow, models
/plan <ticket>          understand the problem and agree on a fix — no code yet
/factory <feature>      spec → review → tests → build → QA, unattended after one review

Commands

The factory

/factory <feature>          one feature or ticket
/factory f1 f2 f3           several features
/factory --roadmap          every open roadmap feature
/factory <PRD path>         a whole project: roadmap first, then every feature
   --auto                   no stops — a reviewer model replaces your review
   --ship                   open one PR per feature (never merged for you)

It plans every feature, has the reviewer critique each plan, and stops once for your approval. Then it builds unattended: independent features run in parallel git worktrees, each passing feature merges into an integration branch, and a stuck feature is parked while the rest continue. Progress and resume commands live in .sdlc/RUN.md.

How it works

• Separate roles. Every stage runs in a fresh agent. The tester writes tests before the code; the builder can't edit them.
• A second opinion. A reviewer on a different model attacks every plan and every diff. Claude Code defaults: Opus 5.5 plans, Sonnet 5.5 builds, Opus 4.8 reviews.
• Evidence, not confidence. Claims about the code carry [read path:line] or [ran cmd]; nothing ships on [assumed]. A QA finding blocks only with proof.
• Focused tests. About 15 per feature, each naming the bug it catches.
• Readable plans. Each feature keeps a human walkthrough.md, in simplified technical English, next to a machine spec.md.
• Bug memory. Fixed bugs go to Bug Ledger after your yes, and every stage checks it before working in that area.

CLI

ai-sdlc install [--agent NAME] [--target PATH] [--global] [--force] [--no-bugledger] [--no-hooks]
ai-sdlc list                      # bundled skills, agents, and hooks
ai-sdlc ste-check FILE...         # check a walkthrough against the writing rules
ai-sdlc --version

Requirements

• Python 3.9+ or Node 18+ for the CLI, and git
• uv for Bug Ledger
• python3 on your PATH for the guard hooks
• Optional: codebase-memory-mcp for faster caller search

Upgrading from 0.2.x

ai-sdlc install --agent claude --force   # replaces /automate with /factory

Then run /sdlc-init in each project — it upgrades .sdlc/ in place.

Links

• Changelog
• License (MIT)

The writing rules are an unofficial paraphrase in the style of ASD-STE100; this project is not affiliated with or endorsed by ASD. ASD-STE100 is a trademark of ASD.