Alfred SuperAgent Core Rules

Foundation for Agent-Based Development

Architecture: Commands → Agents → Skills
Workflow: ADAP (Analyze, Design, Assure, Produce)
Principles: Agent-First, Delegation, Quality Gates

Overview

Defines essential rules for Alfred's decision-making and execution.

Key Responsibilities

1. 3-Layer Architecture: Commands orchestrate, Agents execute, Skills provide knowledge.
2. 4-Step Workflow: Analyze (research) → Design → Assure (quality) → Produce.
3. Agent-First: Delegate all execution to specialized agents.
4. TRUST 4: Enforce Test, Readable, Unified, Secured principles.

Rule 1: 3-Layer Architecture

┌──────────────────────┐
|  generate Commands (Orchestration) │  ← User entry points
│  /alfred:1-plan      │    No direct execution
└───────┬──────────────┘
        │ Task(subagent_type=...)
        ↓
┌──────────────────────┐
│  Agents (Execution)  │  ← Spec-builder, tdd-implementer
│  Deep reasoning      │    Domain expertise
└───────┬──────────────┘
        │ Skill("skill-name")
        ↓
┌──────────────────────┐
│  Skills (Knowledge)  │  ← Reusable patterns
│  Stateless playbooks │    <1000 lines
└──────────────────────┘

Command Layer (Orchestration Only)

Do: Delegate tasks, ask user questions.
Don't: Execute bash, write files, run tests directly.

# ❌ Wrong: Direct execution
bash("git commit -m 'message'")

# ✅ Correct: Delegate to agent
Task(subagent_type="git-manager", description="Commit changes")

Agent Layer (Execution)

Do: Complex analysis, planning, skill invocation.
Don't: Call other agents directly (use Commands for multi-agent orchestration).

Success Skill Layer (Knowledge Capsules)

Do: Provide patterns, best practices.
Don't: Invoke other skills, execute tasks, maintain state.

Rule 2: 4-Step Workflow (ADAP)

1. Analyze: Research latest docs via WebSearch/WebFetch.
2. Design: Architecture based on November 2025 info.
3. Assure: Validate quality (TRUST 4).
4. Produce: Generate files, commit.

Rule 3: Agent-First Paradigm

Forbidden: Commands MUST NOT directly:

• Execute bash commands
• Read/write files
• Run git operations
• Execute tests

Mandatory Delegation:

Rule 4: TRUST 4 Quality Gates

Before commit:

• Test: ≥85% coverage
• Readable: Pylint ≥8.0, clear naming
• Unified: Consistent patterns, no duplication
• Secured: OWASP scan passed (bandit)

Rule 5: AskUserQuestion Usage

When to use:

• Ambiguous requirements ("Create a module" → "What kind?")
• Technology choice ("Which framework?")
• Architecture decisions ("SQL vs NoSQL?")

Example:

AskUserQuestion({
  "question": "Which database?",
  "options": [
    {"label": "PostgreSQL", "description": "Relational"},
    {"label": "MongoDB", "description": "Document"}
  ]
})

Rule 6: Commit Message Standards

TDD Cycle Commits:

• RED: test(@TAG-001): Add tests (tests fail)
• GREEN: feat(@TAG-001): Implement feature (tests pass)
• REFACTOR: refactor(@TAG-001): Optimize code (tests still pass)

Validation Checklist

• Command: Only orchestration, no direct execution?
• Agent: Proper delegation to subagents?
• Skill: Invoked correctly via Skill("name")?
• TRUST 4: All quality gates passed?
• Commit: Follows conventional commits?

Related Skills

• moai-foundation-trust: TRUST 4 principles
• moai-foundation-tags: TAG traceability
• moai-foundation-git: Git workflows

Last Updated: 2025-11-20