Initializer Protocol

The Initializer Agent is responsible for analysis and planning. It runs once and hands off to Coding Agents for execution.

Anthropic Best Practice: Two-Agent Pattern

Separating initialization from execution provides:

•Cleaner session boundaries
•Better context management
•Easier resume from checkpoints
•More efficient token usage

Initializer Responsibilities

Phase 1: Project Detection

bash

# Run project detector
${CLAUDE_PLUGIN_ROOT}/scripts/detect-project.sh

Output: .claude/project-profile.yaml

Phase 2: Work Type Classification

Analyze user request and classify:

yaml

# .claude/auto-context.yaml
work_type: FRONTEND | BACKEND | FULLSTACK | RESEARCH
execution_mode: two_agent
initialized_at: "2025-01-18T10:00:00Z"
initializer_complete: false

Phase 3: Task Decomposition

Use task-decomposer skill to break down into atomic tasks:

markdown

# .claude/plans/auto-{timestamp}.md

## Execution Plan

### Tasks
- Task 1: ...
- Task 2: ...

### Parallel Groups
- Group 1: [Task 1, Task 2]
- Group 2: [Task 3, Task 4]

### Dependencies
- Task 3 depends on Task 1
- Task 4 depends on Task 2

Phase 4: Feature List Creation

Generate feature list with failing status:

yaml

# .claude/plans/features-{timestamp}.yaml
features:
  - id: task-001
    description: "..."
    status: failing
    verification_command: "..."
    expected_output: "..."
    evidence_required: true
    can_be_removed: false

Phase 5: State Machine Setup

Initialize state machine:

yaml

# .claude/auto-state-machine.yaml
session_id: "auto-20250118-100000"
current_state: EXECUTE
work_type: FRONTEND
initialized: true
initializer_complete: true
current_group: 1
total_groups: 3
checkpoint_files: []

Phase 6: Write Initial Checkpoint

bash

${CLAUDE_PLUGIN_ROOT}/scripts/checkpoint-manager.sh write INIT \
  "Initialization complete. Plan created with X tasks in Y groups."

Phase 7: Exit with Handoff

Output handoff message and exit:

code

┌─────────────────────────────────────────────────────────────┐
│ Initializer Agent Complete                                  │
│                                                             │
│   Project: Next.js application                              │
│   Work Type: FRONTEND                                       │
│   Tasks: 8 total                                            │
│   Parallel Groups: 3                                        │
│                                                             │
│   Files Created:                                            │
│     ✓ .claude/project-profile.yaml                          │
│     ✓ .claude/auto-context.yaml                             │
│     ✓ .claude/plans/auto-20250118-100000.md                 │
│     ✓ .claude/plans/features-20250118-100000.yaml           │
│     ✓ .claude/auto-state-machine.yaml                       │
│                                                             │
│   Next Steps:                                               │
│     Run /auto-continue to start execution                   │
│                                                             │
│   Coding Agent will execute parallel group 1:               │
│     - Task 1: Create UserForm component                     │
│     - Task 2: Create Header component                       │
│     - Task 3: Create Sidebar component                      │
└─────────────────────────────────────────────────────────────┘

<promise>INITIALIZER_COMPLETE</promise>

Initializer DO NOT

The Initializer Agent MUST NOT:

•❌ Write any production code
•❌ Modify existing source files
•❌ Run tests (no code to test)
•❌ Create worktrees (Coding Agent does this)
•❌ Continue beyond handoff

Token Budget

Initializer should use minimal tokens:

Phase	Max Tokens
Project Detection	~500
Work Classification	~300
Task Decomposition	~1500
State Setup	~200
Checkpoint	~200
Total	~3000

Handoff Validation

Before signaling INITIALIZER_COMPLETE, verify:

• .claude/project-profile.yaml exists
• .claude/auto-context.yaml exists with execution_mode: two_agent
• Plan file exists at .claude/plans/auto-*.md
• Feature list exists at .claude/plans/features-*.yaml
• State machine initialized with initializer_complete: true
• Initial checkpoint written

Error Handling

If initialization fails:

code

<promise>INITIALIZER_FAILED</promise>

## Failure Report

### Error
[What went wrong]

### Partial State
[What was completed before failure]

### Recovery
[How to recover or retry]

Integration

The Initializer Protocol integrates with:

•auto.md - Invoked when execution_mode: two_agent
•auto-continue.md - Reads Initializer output
•checkpoint-manager.sh - Writes handoff checkpoint