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.
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.
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"| 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) |
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
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.
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.
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.
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.
| 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 |
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.
.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 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 |
| 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) |
/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| 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 planRun 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 planPick 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, stopWhat 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
/sdlcto start a new one
| 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 |
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'# 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:resumeOpen docs/architecture.html in a browser for an interactive visual reference covering cross-platform architecture, workflow phases, agent pipeline, context isolation, and platform capabilities.