dx-aem-flow
Enterprise AI development platform for teams shipping on Azure DevOps and Atlassian. 75+ skills that run identically across Claude Code, GitHub Copilot CLI, and VS Code Chat — from ticket to PR, fully autonomous. Deep AEM specialization built in.
Why This Exists
Enterprise teams run complex sprints across multiple repos, multiple IDEs, and multiple team members. They need a system that handles the full lifecycle — requirements analysis, implementation planning, code generation, testing, verification, documentation, and PR creation — with governance at every step. Without re-explaining the project every session.
KAI is a structured development workflow built as a plugin system for enterprise teams. It encodes your entire sprint lifecycle — requirements, planning, execution, review, PR — into skills that orchestrate multi-agent pipelines across every major AI platform. A single command like /dx-req pulls the ticket from Azure DevOps or Jira, validates readiness against your DoR, distills developer requirements, researches the codebase with parallel subagents, and generates a team summary. Each skill chains specialized agents (Opus for deep review, Sonnet for execution, Haiku for lookups), gathers context from multiple sources (tickets, config, codebase, Figma designs, live AEM content), and writes structured output that the next skill picks up automatically.
What makes it different:
- Every AI platform — same 75+ skills work identically in Claude Code, GitHub Copilot CLI, and VS Code Chat. Same plugins, same config, same results — regardless of which IDE your team uses.
- Enterprise governance — DoR validation, 6-phase verification gate (compile, lint, test, secret scan, architecture review, AI code review), autonomous PR review, and branch protection. Ship with confidence.
- AEM full-flow — purpose-built AEM tooling covering Figma → component → dialog inspection → JCR content → editorial QA → browser automation → demo capture. Includes the AEM MCP server for live dialog and JCR access. The complete AEM development lifecycle, config-driven for your project.
- Config-driven, not prompt-driven — your build commands, branch names, and conventions live in one config file. Every skill reads it. No hardcoded values, no repeated instructions.
- Persistent memory between steps — each skill writes structured output to local files. The next skill picks it up automatically. Sessions can end and resume without losing context.
- Autonomous mode — the same skills that run locally also run unattended as ADO pipeline agents, triggered by Azure-native Service Hooks (no AWS required). Comment
@kai-bugfix on a ticket, get a verified bugfix with a PR.
Install
Prerequisite — your project should already have a CLAUDE.md.
These are plugins: they install into a project that is already set up. If yours has no
CLAUDE.md, run Claude Code's own /init first. The plugins never create or modify it.
This matters since Claude Code v2.1.277: in a project with no CLAUDE.md, Claude Code
falls back to reading AGENTS.md — and /dx-init generates AGENTS.md as an agent inventory
table, not project instructions. With CLAUDE.md present, the default mode
(claude-md-or-agents-md) ignores AGENTS.md and everything behaves as before.
Add the marketplace, then install the plugins you need:
# Add the marketplace (once)
/plugin marketplace add easingthemes/dx-aem-flow
# Install plugins
/plugin install dx-core@dx-aem-flow # Core workflow (all projects)
/plugin install dx-hub@dx-aem-flow # Multi-repo orchestration (optional)
/plugin install dx-aem@dx-aem-flow # AEM tools (AEM projects)
/plugin install dx-automation@dx-aem-flow # Autonomous agents (24/7 pipelines)
From a local clone:
/plugin marketplace add /path/to/dx-aem-flow
/plugin install dx-core@dx-aem-flow
Plugins
dx-core — Development Workflow
Full-stack development workflow for Azure DevOps and Jira projects: requirements gathering, implementation planning, step-by-step execution with testing and review, code review, bug fixes, and PR management.
Works with any tech stack.
dx-hub — Multi-Repo Orchestration
Hub directory management for coordinating work across multiple consumer repos — init, config, status.
dx-aem — AEM Full-Flow
The complete AEM development lifecycle: component dialog inspection, JCR content, page authoring, editorial QA with browser automation, snapshot/verify lifecycle, and demo capture. Includes the AEM MCP server for live JCR and dialog access. Purpose-built for AEM Cloud and on-premise projects.
Requires dx plugin.
Eleven autonomous AI agents (DoR checker, DoD checker, DoD fixer, PR reviewer, PR answerer, BugFix agent, QA agent, DevAgent, DOCAgent, Estimation, SimpleAgent) running 24/7 as ADO pipelines.
Triggers are Azure-native by default — no AWS account needed. An ADO Service Hook posts to an Incoming WebHook service connection the pipeline listens on. The AWS Lambda router is optional, for teams that want central dedupe, per-agent rate limits and token-budget gating in one place.
Requires dx plugin.
Client-Specific Plugins
Client-specific plugins can live alongside these plugins for project-specific enrichment (market data, brand config, etc.). See the authoring guide for how to create them.
Quick Start
Three ways to initialize — all produce the same project structure:
Claude Code
# 1. Install plugins
/plugin marketplace add easingthemes/dx-aem-flow
/plugin install dx-core@dx-aem-flow
# 2. Initialize (interactive — detects your project, asks a few questions)
/dx-init
# 3. Work on a story
/dx-req 2416553 # Fetch story, distill requirements, research codebase
/dx-plan # Generate implementation plan
/dx-step-all # Execute all steps autonomously
/dx-step-build # Build & deploy
/dx-step-verify # 6-phase verification
/dx-pr # Create pull request
GitHub Copilot CLI
# 1. Install plugins (same format, same skills)
/plugin install easingthemes/dx-aem-flow/dx-core
# 2. Initialize (same interactive flow)
/dx-init
# 3. Work on a story — same skills, same agents
@DxReqAll 2416553 # Full requirements workflow
@DxPlanExecutor # Generate and execute plan
@DxStepAll # Execute all steps
@DxCodeReview # Code review
@DxCommit pr # Commit and create PR
Standalone scaffold (any agent or no agent)
Bootstrap the project config without installing plugins first. Works with any AI coding agent — Claude Code, Copilot CLI, Codex, or others:
# From a clone of this repo:
node cli/bin/dx-scaffold.js /path/to/your-project --all
# Flags: --aem (AEM-specific files), --all (everything)
Auto-detects git remote, SCM provider, project type, and base branch. Generates config, rules, agents, and AGENTS.md — edit .ai/config.yaml to set your actual values. See cli/README.md for details.
Agent Compatibility
These plugins work across the AI coding agent ecosystem at three levels:
| Level | Agents | What Works |
|---|
| Full workflow | Claude Code, Copilot CLI | Plugin install, all 75+ skills, agent orchestration, hooks |
| Agents + rules | VS Code Chat, Copilot coding agent, Codex CLI | .github/agents/, AGENTS.md, rules, MCP servers |
| Rules + MCP | Cursor, Windsurf, Amazon Q, Cline, Continue | Coding conventions from .claude/rules/, MCP server configs |
The scaffold generates AGENTS.md (read by Codex CLI, Windsurf, Copilot coding agent) and .github/agents/ (read by VS Code Chat, Copilot CLI) so agent profiles work even without plugin installation.
Quality
Two layers, both in the open:
Structural validation runs in CI on every PR — skill and agent naming, frontmatter,
manifest/version consistency, no cross-plugin collisions (scripts/validate-*.sh).
Behavioral evals check what the skills actually do. Because a skill is instructions
for a model rather than code, the same prompt can give different answers on different
runs — so a skill is scored across repeated runs against fixtures whose correct answer
is known in advance, using the first-party claude plugin eval harness.
This is early: one suite so far, covering dx-plan-validate — a plan with one
planted defect, graded for both catching the real gap and not inventing extra ones.
It's a starting point, not coverage. Suites are added per skill as they're written, and
each lives beside the plugin it tests in plugins/<plugin>/evals/.
See the suite's README for how
a case and its graders are put together, including what the first run got wrong.
Documentation
Full documentation: KAI Website
License
MIT
