---
priority: 10
command_name: apm-loop-start
description: Start the APM Loop to automate prompt routing between Manager and Implementation Agents
---

# APM Loop Start Skill

This skill launches the **APM Loop** automated routing system. It takes Manager Agent output containing a Task Assignment Prompt and automates the routing loop between Manager and Implementation Agents.

---

## CRITICAL: Autonomous Execution Mode

**This loop MUST run without user input.** You are expected to:

1. **Never pause for user confirmation** - Make decisions autonomously based on the task context
2. **Never ask clarifying questions** - Use your best judgment and document uncertainties
3. **Continue through all tasks** - Do not stop the loop to ask the user anything
4. **Log issues for later review** - Any unresolved questions or issues must be written to the issues file (see below)

The user will review all logged issues AFTER the loop completes. This enables true autonomous operation.

---

## Issue Tracking

When you encounter unresolved issues, questions, or decisions that would normally require user input, **do not stop**. Instead, append them to:

.apm/loop-issues.json

### Issue File Structure

```json
{
  "loop_session": "2024-01-25T10:30:00Z",
  "issues": [
    {
      "id": "issue-001",
      "timestamp": "2024-01-25T10:35:22Z",
      "task_ref": "Task 1.2 - Setup Authentication",
      "agent": "Agent_Backend",
      "type": "question|decision|blocker|warning",
      "severity": "low|medium|high|critical",
      "title": "Brief description of the issue",
      "description": "Detailed explanation of what was encountered",
      "context": "Relevant code, file paths, or decisions made",
      "action_taken": "What the agent decided to do (or skip)",
      "needs_review": true,
      "resolved": false
    }
  ],
  "summary": {
    "total_issues": 1,
    "by_type": { "question": 0, "decision": 1, "blocker": 0, "warning": 0 },
    "by_severity": { "low": 0, "medium": 1, "high": 0, "critical": 0 },
    "unresolved_count": 1
  }
}

Issue Types

Type	When to Use
question	Ambiguity in requirements that you resolved with an assumption
decision	A design/implementation choice where multiple valid options existed
blocker	Something that prevented completion (you worked around or skipped)
warning	Potential problems the user should be aware of

Severity Levels

Severity	Description
low	Minor uncertainty, unlikely to affect functionality
medium	Notable decision that user should review
high	Significant assumption that could require rework
critical	Major blocker or risk that needs immediate attention

Always continue execution - even critical issues should be logged, not used as a reason to stop and ask the user.

---

What This Skill Does

When invoked, this skill will:

1. Save the Manager Agent's output to a temporary file
2. Launch apm-loop start to begin automated routing
3. The loop will:
   • Extract Task Assignment Prompts from Manager output
   • Route tasks to the correct Implementation Agent (by domain)
   • Capture Final Task Reports from Implementation Agents
   • Return reports to Manager for the next task
   • Handle handovers automatically when context limits approach
   • Continue until all tasks are complete or an error occurs

---

Instructions

The APM Loop can be started in two modes:

Mode 1: Bootstrap Mode (Recommended for New Projects)

Use this mode when starting a new project from scratch. The loop will automatically initialize the Manager Agent.

Step 1: Collect Bootstrap Prompt

Ask the User to paste the Manager Agent Bootstrap Prompt that includes:

• User intent and requirements
• Implementation plan overview
• Technology stack
• Success criteria

Step 2: Save to Temp File

Save the bootstrap prompt to:

.apm/loop-input.md

Step 3: Launch in Bootstrap Mode

Run the following command via Bash:

apm-loop start -f .apm/loop-input.md -b -v

The -b flag enables bootstrap mode. The loop will:

1. Initialize Manager Agent with /apm-2-initiate-manager
2. Auto-confirm when Manager asks to proceed
3. Extract first Task Assignment
4. Begin normal routing loop

---

Mode 2: Resume Mode (For Existing Manager Sessions)

Use this mode when you already have Manager output with a Task Assignment.

Step 1: Collect Manager Output

Ask the User to paste the full Manager Agent output that contains a Task Assignment Prompt.

The output can be in either format:

Format 1: Backtick-wrapped (original format)

```markdown
---
task_ref: "Task X.Y - Task Name"
agent_assignment: "Agent_Domain"
...
---
# Task Assignment content...

**Format 2: Raw YAML frontmatter (Claude Code output)**

---

task_ref: "Task X.Y - Task Name" agent_assignment: "Agent_Domain" ...

Task Assignment content...

Both formats are automatically recognized and parsed correctly.

**Step 2: Save to Temp File**

Once the User has pasted the Manager output, save it to:

.apm/loop-input.md

**Step 3: Launch in Resume Mode**

Run the following command via Bash:
```bash
apm-loop start -f .apm/loop-input.md -v

The loop will resume the existing Manager session and begin routing.

Command Options:

• -f .apm/loop-input.md - Read input from the saved file
• -b - Bootstrap mode (auto-initialize Manager from Bootstrap Prompt)
• -v - Enable verbose output to see detailed progress
• -t <ms> - Set timeout in milliseconds (default: 1200000 = 20 min)
• --debug - Enable debug output for troubleshooting
• -d - Dry-run mode to preview extraction without spawning agents (not supported in bootstrap mode)

---

Completion

When the loop completes:

• A summary will be printed to the terminal
• A summary file will be written to .apm/loop-summary.md
• Review .apm/loop-issues.json for any lo