Siegard Code

Autonomous agent infrastructure for Claude Code that ships features — from spec to delivery.

18 specialized agents. 3 coordinated teams. One command to orchestrate them all.

Most AI coding tools help you write code. Siegard Code manages the entire development lifecycle — it writes specifications, plans backlogs, implements features, runs QA, and delivers tested code. All autonomously, all traceable, all through Claude Code.

You describe what you want. Siegard's agents handle the rest: the Spec team writes and validates technical specifications, the Dev team plans and implements Stories with automated QA, and the Reverse Spec team can even document existing codebases. Every step produces artifacts. Every decision is logged. Every quality gate has teeth.

This is not a product. It is the agent infrastructure you install into your own projects.

Why Siegard Code?

Architecture

Three agent teams work in sequence, connected by formal handoff protocols:

graph TB
    subgraph INPUT["Input"]
        REQ[Requirement / Existing code]
    end

    subgraph RSPEC["Reverse Spec Team · 3 agents"]
        RSORC[Orchestrator]
        RSANA[Analyzer]
        RSWRT[Writer]
        RSORC --> RSANA --> RSWRT
    end

    subgraph SPEC["Spec Team · 6 agents"]
        SPORC[Orchestrator]
        SPWRT[Spec Writer]
        SPREV[Spec Reviewer]
        SPBCK[Back Spec Agent]
        SPFRT[Front Spec Agent]
        SPVAL[Spec Validator]
        SPORC --> SPWRT --> SPREV
        SPREV -->|APPROVED| SPBCK
        SPREV -->|REJECTED| SPWRT
        SPBCK --> SPVAL
        SPVAL -->|all back.md valid| SPFRT
        SPFRT --> SPVAL
        SPVAL -->|INVALID| SPBCK
        SPVAL -->|INVALID front| SPFRT
    end

    subgraph DEV["Dev Team · 9 agents"]
        DVORC[Orchestrator-Dev]
        PLAN[Planner]
        UIAG[UI Agent]
        DEVAG[Developer]
        QAAG[QA & Docs]
        DVORC --> PLAN --> UIAG --> DEVAG --> QAAG
        QAAG -->|REJECTED| DEVAG
    end

    REQ --> RSORC
    REQ --> SPORC
    RSWRT -->|specs · draft| SPORC
    SPVAL -->|VALID| DVORC
    QAAG -->|APPROVED| FIN[Delivery]

Quick Start

Prerequisites

• Claude Code installed and configured

Install into your project

Copy the dist/.claude/ directory into your project:

cp -r dev-team/dist/.claude/ /path/to/your-project/.claude/

This copies all agents, skills, and commands into your project's .claude/ directory.

Configure your project

Add these fields to your project's CLAUDE.md:

domain: backend           # or "frontend" — determines which pipeline runs
stack: Node.js, Express   # your tech stack (agents adapt, not hardcode)
specs_dir: docs/specs     # where specifications live

Run your first command

# Create a specification from requirements
/u-spec docs/specs

# Implement from approved specs
/u-dev docs/specs my-feature

That's it. The orchestrator detects the mode, activates the right agents, and guides you through every step.

Commands

Six slash commands cover the entire development lifecycle:

Workflows

Feature (full):       /u-spec  →  /u-dev
Quick improvement:    /u-improve  →  /u-dev
Bug fix:              /u-bug-report  →  /u-dev
Reverse + evolve:     /u-reverse-spec  →  /u-spec  →  /u-dev

Key Features

Spec-First Development

Every feature starts with a validated specification — openapi.yaml, use cases, business rules, state machines, screen specs, and navigation flows. No code is written without an approved spec.

Automatic Mode Detection

The orchestrator reads your project state and selects the right mode: Spec-first, Improve, Bug, Resume, or Reverse-eng review. No flags, no configuration — it just works.

Frontend & Backend Pipelines

One command, two pipelines. The domain: field in your CLAUDE.md routes to the correct agents. Frontend gets a UI Agent for visual specs; backend gets infrastructure dependency tracking.

Quality Gates with Escalation

• Spec Reviewer: Max 3 rejection cycles before human escalation
• Spec Validator: Cross-reference validation with coverage reports
• QA Agent: Test-gate + full mode with max 3 rework rounds
• Nothing passes silently — every gate produces artifacts

Token-Efficient Context Management

• Context mounting: Each agent receives only what it needs
• Short mode: Reactivations use ~2K tokens instead of ~15K
• Triage mode: Process 5–10 errors per session, not 20+ at once

Session Resilience

Interrupted? Run the same command again. The orchestrator reads its log file, reconstructs state from disk artifacts, and resumes from the exact stopping point.

Project Structure

your-project/
├── .claude/
│   ├── agents/
│   │   ├── dev/                  # Dev team (FE + BE agents + protocols)
│   │   ├── spec/                 # Spec team (6 agents + protocols)
│   │   └── reverse-spec/         # Reverse Spec team (3 agents + protocols)
│   ├── commands/                 # 6 slash commands
│   └── skills/                   # 20+ reusable skill libraries
├── CLAUDE.md                     # Your project configuration
├── {SPECS_DIR}/
│   ├── _global/                  # Shared: conventions, error codes, glossary
│   ├── domains/{domain}/         # Backend: openapi.yaml + spec.md + back.md
│   └── front/                    # Frontend: screens, flows, design system
└── {SESSIONS_DIR}/
    └── {SESSION}/                # Current session working directory
        ├── backlog.md            # Planned Stories
        ├── us-XX-delivery.md     # Delivery artifacts
        ├── us-XX-qa.md           # QA reports
        └── log-orchestrator-dev.md

Action Guide

Documentation

Estimates

License

This project is licensed under the Apache License 2.0. See NOTICE for attribution details.

Siegard Code — Because shipping features shouldn't mean losing control.