Skip to content

krukovden/sdlc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

11 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SDLC Multi-Agent System

An AI-powered Software Development Life Cycle system. Five specialized agents handle the work — you approve the decisions.

Works with Claude Code, GitHub Copilot, and OpenAI Codex.


The Idea

Traditional software development needs many roles: analyst, architect, developer, tester, reviewer. This system gives each role to a specialized AI agent:

  You say:     /sdlc feature "add user notifications"

  Lead agent    understands what you need     → clarify
  Lead agent    analyzes your codebase        → research
  Lead agent    designs the solution          → design
  Lead agent    breaks it into tasks          → plan
  Coder agent   writes the code              ─┐
  Tester agent  writes and runs tests         │ implement
  Reviewer agent checks code quality          │ (per task)
  Security agent scans for vulnerabilities   ─┘

You approve after each phase. Tasks run continuously without stopping between them.


Quick Start

npx github:krukovden/sdlc

# Setup asks which tool you use (Claude Code / Copilot / Codex / All)
# and generates platform-specific config files from .agents/

# Or specify the platform directly:
npx github:krukovden/sdlc claude

# Then start a workflow:
/sdlc feature "add user notifications"

How It Works

1. You pick a workflow type

Type When to use Ends at
feature Build something new implement
bugfix Fix broken behavior implement
refactor Restructure working code implement
spike Research a question design (no code)

2. The system walks through phases

  clarify     What are we building? Scope? Constraints?
     ⏸ you approve
  research    What exists in the codebase? What patterns?
     ⏸ you approve
  design      Architecture, API contracts, data model, test strategy
     ⏸ you approve
  plan        Task breakdown with dependencies and skill assignments
     ⏸ you approve
  implement   Agents execute tasks continuously (stops only on failure)
     ⏸ you approve final result

3. Each task goes through the agent pipeline

  Lead dispatches → Coder → Tester → Reviewer → Security → Lead checks compliance

If any agent fails, the system retries up to 3 times. If it still fails, it stops and asks you what to do.

4. Git worktree isolation

At workflow start, the system asks: worktree or current branch?

  • Worktree — creates an isolated copy of the repo. Your main branch stays clean while agents work in parallel. Recommended for features.
  • Current branch — works directly. Simpler, good for bugfixes.

5. Live dashboard

The system generates a live HTML dashboard (Kanban board) that shows agent progress in real time:

  ┌─ Queue ──────┐  ┌─ Active ─────┐  ┌─ Done ───────┐
  │ Task 3       │  │ Task 2       │  │ Task 1       │
  │ Task 4       │  │ Agent: Tester│  │ ✅ All passed │
  │              │  │ ✅⚙○○○       │  │              │
  └──────────────┘  └──────────────┘  └──────────────┘

It polls manifest.json every 2 seconds — no refresh needed. Launches automatically in your browser when you start a workflow.

6. Everything is saved to files

Each workflow creates a folder in docs/workflows/ with all artifacts. Next session — even from a different tool — the system finds it and offers to resume.


Five Agents

Agent What it does Can modify code?
Lead Orchestrates phases, dispatches agents, checks design compliance No
Coder Writes implementation code Yes
Tester Writes and runs tests Yes
Reviewer Reviews code quality, outputs PASS or FAIL No
Security Scans for vulnerabilities, outputs PASS or FAIL No

Cross-Platform

The system works from any AI coding tool. .agents/ is the single source of truth — platform adapters are generated.

                         .agents/
            ┌──────── source of truth ────────┐
            │    skills, agents, guidelines,   │
            │    workflows                     │
            └───────┬──────────┬──────────┬───┘
                    │          │          │
                    ▼          ▼          ▼
              .claude/    .github/    .codex/
             Claude Code   Copilot     Codex

Run npx github:krukovden/sdlc once per project — it generates the right config for your tool.


Project Structure

.agents/                  ← committed, shared across all tools
├── skills/               ← what each agent knows (17 skills)
├── agents/               ← how each agent behaves (5 agents)
├── guidelines/           ← coding standards everyone follows
└── workflows/            ← phase definitions per workflow type

.claude/                  ← generated by setup.js
.github/                  ← generated by setup.js
.codex/                   ← generated by setup.js

docs/workflows/           ← gitignored, your active work

Skills

Skills are domain knowledge that agents load on demand:

Skill What it knows
architect API design, system boundaries, ADRs
frontend-angular Angular 18+, signals, RxJS
backend-node Node/TS, controllers → services → repos
backend-csharp C# Azure Functions, EF Core
devops-azure Azure Pipelines, templates
devops-github GitHub Actions, reusable workflows
enhanced-reviewer Code review checklist

Commands

Setup

Command What it does
npx github:krukovden/sdlc Interactive setup — asks which tool, generates config
npx github:krukovden/sdlc claude Generate Claude Code config only (also: copilot, codex, all)

Full Workflow

/sdlc feature "add user notifications"    # full pipeline: clarify → implement
/sdlc bugfix "login returns 500"           # same pipeline, bugfix-focused
/sdlc refactor "extract payment service"   # same pipeline, refactor-focused
/sdlc spike "evaluate auth libraries"      # stops at design, no code

Flags

Flag Description Example
--auto-approve Skip approval prompts between phases /sdlc feature --auto-approve "add notifications"
--stop-at <phase> Run only up to the specified phase, then stop /sdlc feature --stop-at design "add notifications"

Flags can be combined:

# Run clarify + research only, no approval prompts
/sdlc feature --auto-approve --stop-at research "add notifications"

# Resume and stop after plan
/sdlc:resume --auto-approve --stop-at plan

Single Phase

Run one phase at a time when you want more control:

/sdlc:clarify "add user notifications"     # just clarify scope
/sdlc:research                             # just analyze codebase
/sdlc:design                               # just produce design artifacts
/sdlc:plan                                 # just break into tasks
/sdlc:implement                            # just execute the plan

Resume

Pick up where you (or another tool) left off. The system scans docs/workflows/ for incomplete workflows and continues from the first phase that isn't approved yet.

/sdlc:resume                               # resume from where you left off
/sdlc:resume design                        # resume from a specific phase
/sdlc:resume --auto-approve                # resume without approval prompts
/sdlc:resume --auto-approve --stop-at plan # resume, do plan only, stop

What if there are multiple workflows?

  • One incomplete workflow → resumes it automatically
  • Multiple incomplete workflows → asks which one to resume
  • With --auto-approve → picks the most recent automatically
  • All workflows complete → tells you and suggests /sdlc to start a new one

Command Reference Across Platforms

Action Claude Code Copilot (VS Code) Codex
Setup npx github:krukovden/sdlc npx github:krukovden/sdlc npx github:krukovden/sdlc
Full pipeline /sdlc feature "..." /sdlc feature "..." $sdlc feature "..."
Clarify only /sdlc:clarify "..." /sdlc:clarify "..." $sdlc-clarify "..."
Research only /sdlc:research /sdlc:research $sdlc-research
Design only /sdlc:design /sdlc:design $sdlc-design
Plan only /sdlc:plan /sdlc:plan $sdlc-plan
Implement only /sdlc:implement /sdlc:implement $sdlc-implement
Resume workflow /sdlc:resume /sdlc:resume $sdlc resume
Resume from phase /sdlc:resume design /sdlc:resume design $sdlc resume design

Headless / CI Usage

Each tool can run the workflow non-interactively via its CLI:

# Claude Code
claude -p '/sdlc feature --auto-approve "add notifications"' --dangerously-skip-permissions

# GitHub Copilot
copilot -p '/sdlc feature --auto-approve "add notifications"' --allow-all --no-auto-update --silent

# OpenAI Codex
codex exec --skip-git-repo-check --full-auto '/sdlc feature --auto-approve "add notifications"'

Context switching — one tool starts, another resumes from the same docs/workflows/:

# Tool A: start and stop after research
claude -p '/sdlc feature --auto-approve --stop-at research "add notifications"' --dangerously-skip-permissions

# Tool B: resume and stop after design
copilot -p '/sdlc:resume --auto-approve --stop-at design' --allow-all --no-auto-update --silent

# Tool C: resume and finish
codex exec --skip-git-repo-check --full-auto '/sdlc:resume --auto-approve'

Typical Session

# First time setup
npx github:krukovden/sdlc

# Start a feature
/sdlc feature "add email notifications for order updates"
  → clarify: answers scope questions, approve
  → research: reviews codebase, approve
  → design: produces API contracts + data model, approve
  → plan: 4 tasks with dependencies, approve
  → implement: agents execute all 4 tasks, approve final result
  → done: merge, create PR, or keep branch

# Next day, different tool — resume where you left off
/sdlc:resume

Architecture Diagram

Open docs/architecture.html in a browser for an interactive visual reference covering cross-platform architecture, workflow phases, agent pipeline, context isolation, and platform capabilities.

About

AI-powered SDLC system with 5 specialized agents (Lead, Coder, Tester, Reviewer, Security). Walks through clarify → research → design → plan → implement with approval gates. Works with Claude Code, GitHub Copilot, and OpenAI Codex from a single source of truth.

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors