DevLLMOps

Warning

This project is under active development. Resources provided (n8n workflows) should be used for experimentation only as no extensive testing has been performed yet.

DevLLMOps is a set of specifications, tools and introduces the AIgile methodology for teams to develop and deploy software at AI speed, keeping humans in the loop where it matters.

Ship fast or die slow. A methodology for AI-native software development in teams.

Why

The traditional SDLC (Requirements > Design > Code > Test > Review > Deploy > Monitor) assumed building was expensive. That constraint is gone. AI agents collapse these stages into a tight loop where intent, code, tests, and deployment converge simultaneously.

Based on OCPA specs for a testable project structure and rooted in DevOps principles for reliability and maintainability.

Demo Project

See devllmops-demo-quizapp for a reference implementation: a quiz app built end-to-end with DevLLMOps + OCPA specs, including CI workflows, AI review, intent-driven issues, and CLAUDE.md/TEAM.md/REVIEW.md configuration.

The "AIgile" Workflow

flowchart TD
    A["`**Human Intent**`"] --> B["`**AI Agent**`"]
    B --> C["`**Code + Tests**`"]
    C --> D["`**Automated CI**
    lint, test, security`"]
    D --> E["`**AI Agent Review**
    adversarial`"]
    E --> F{Pass?}
    F -- Yes --> G["`**Deploy**`"]
    G --> H["`**Observe**`"]
    H --> A
    F -- "Fail (routine)" --> I["`**Agent fixes**`"]
    I --> D
    F -- "Fail (novel)" --> J["`**Human Review**`"]
    J --> B

Loading

Stages don't get faster. They merge. The agent doesn't know what "phase" it's in. There's just intent, context, and iteration.

Team Roles

Role	Evolved From	Responsibility
Product Architect	CTO / Tech Lead	Defines intent, sets architecture guardrails, handles exceptions agents can't resolve, makes release decisions
Context Engineer	Developer	Crafts agent context, steers AI work, reviews complex/novel changes
Quality Sentinel	QA + DevOps	Owns security review of AI output, manages observability and CI/CD, monitors the feedback loop
AI Ops Lead (large teams)	New role	Manages agent costs, model selection, prompt optimization, orchestration pipelines

See Team Roles & Organization for full details, team structures, and transition from Agile.

Quick Start

1. Install Core Tools

Tool	Purpose	Install
Claude Code	AI agent (default)	npm install -g @anthropic-ai/claude-code
Docker	Containerization	See docs
Make	Command standardization	apt install make / brew install make
n8n	Workflow automation	Host org-wide or self-host with Docker
Gitleaks	Secrets scanning	Runs in CI (GitHub Actions)

See Tooling Setup for full installation and configuration.

2. Configure Your GitHub Repository

1. Enable branch protection on main and release (require status checks, no force-push)
2. Set up GitHub Projects board: Backlog > Ready > AI Ready > In Progress > Verification > Human Review > Done
3. Add secrets: ANTHROPIC_API_KEY, and optionally KUBE_CONFIG, SCW_ACCESS_KEY/SCW_SECRET_KEY
4. Copy the provided GitHub Actions workflows and issue template

See GitHub Setup for step-by-step configuration.

3. Project Structure (OCPA-Based)

.
├── .github/               # CI/CD pipelines (guardrails) and templates
├── app/                   # Service(s) with Dockerfile
├── docs/                  # All documentation (except README.md)
├── k8s/                   # Helm chart (if using K8s)
├── scripts/               # POSIX deployment scripts
├── compose.base.yml       # Shared service config
├── compose.dev.yml        # Dev environment
├── compose.prod.yml       # Production environment
├── compose.test.yml       # Test environment
├── Makefile               # Standardized commands
├── CLAUDE.md              # Agent context (architecture, conventions)
├── TEAM.md                # Team roster for agent review routing
├── REVIEW.md                # Team roster for agent review routing
├── VERSION                # Semantic version
└── .env.example           # Environment variables template

See OCPA Specs for full conventions (versioning, Dockerfiles, Makefile commands, env validation).

Templates available: CLAUDE.md, TEAM.md, REVIEW.md.

4. Create Your First Intent

Open a GitHub Issue using the intent template:

Intent: Add user authentication with JWT
Context: We need login/signup for the API. Using PostgreSQL for storage.
Acceptance: POST /auth/login returns a JWT. Protected routes return 401 without token.
Constraints: Tokens expire after 1h. Use bcrypt for password hashing.

Then let the agent work. Steer, iterate, ship.

Cost Warning

AI agents consume tokens at scale. A single Context Engineer running Claude can use $30-150+/day in API costs. Multiply by team size.

Monthly cost for a team of 4: $3,000-15,000+ in AI tokens alone, on top of infrastructure and salaries.

Mitigations:

• Set hard budget limits on your AI provider dashboard
• Use cheaper models (Haiku) for routine tasks, expensive models (Opus) for complex reasoning
• Monitor usage daily -- there is no "unlimited plan"
• See Cost Management for detailed strategies

Using a Different AI Agent

This spec defaults to Claude via Claude Code CLI. To use an alternative:

Tool	Type	Notes
aider	CLI agent	Works with any OpenAI-compatible API (OpenRouter, Ollama, local models)
Continue.dev	IDE extension	VS Code/JetBrains, any OpenAI-compatible endpoint
OpenHands	Agent platform	Open-source, self-hosted

Any model accessible via an OpenAI-compatible API (GPT-4, Llama, Mistral, DeepSeek) can replace Claude. Adapt the CLAUDE.md to your agent's context mechanism (e.g., .aider.conf.yml for aider).

Documentation

Document	Description
Methodology	The DevLLMOps workflow in detail
Team Roles & Organization	Roles, team structures, transition from Agile
Tooling Setup	Installation and configuration for all tools
n8n Setup	n8n deployment and automation workflows
GitHub Setup	Repository, Actions, Projects configuration
Security	Securing AI-generated code
Cost Management	Token cost control strategies

License

MIT -- See LICENSE