HITL (Human-in-the-Loop) Approval Workflow

Multi-step workflow for human approval integration in ccswarm agent operations.

Overview

This skill guides you through implementing and using human-in-the-loop approval mechanisms for high-risk agent operations.

When HITL is Required

High-Risk Operations

• File deletions
• Database modifications
• External API calls
• Configuration changes
• Deployment actions

Risk Levels

Implementation

1. HITL Configuration

// ccswarm.json
{
  "hitl": {
    "enabled": true,
    "risk_threshold": 3,
    "auto_approve_patterns": [
      "*.md",
      "docs/*",
      "*.test.rs"
    ],
    "always_require": [
      ".env*",
      "Cargo.toml",
      "production/*"
    ],
    "timeout_seconds": 300
  }
}

2. Approval Request Format

pub struct ApprovalRequest {
    pub id: Uuid,
    pub agent_id: String,
    pub operation: OperationType,
    pub target: String,
    pub risk_score: u8,
    pub context: String,
    pub timeout: Duration,
}

pub enum ApprovalStatus {
    Pending,
    Approved { by: String, at: DateTime<Utc> },
    Rejected { by: String, reason: String },
    TimedOut,
}

Workflow

1. Request Approval

# Agent requests approval
ccswarm hitl request \
  --operation "delete" \
  --target "src/legacy/old_module.rs" \
  --reason "Removing deprecated module" \
  --risk-score 4

# Output: Approval request ID: abc123

2. Review Pending Requests

# List pending approvals
ccswarm hitl list

# View details
ccswarm hitl show abc123

3. Approve/Reject

# Approve with comment
ccswarm hitl approve abc123 --comment "Verified module is unused"

# Reject with reason
ccswarm hitl reject abc123 --reason "Module still referenced in tests"

4. Check Status

# Check approval status
ccswarm hitl status abc123

TUI Integration

The HITL system integrates with ccswarm's TUI:

┌─────────────────────────────────────────┐
│ Pending Approvals (2)                   │
├─────────────────────────────────────────┤
│ [!] abc123 - Delete old_module.rs       │
│     Risk: 4/5  Agent: backend-agent     │
│     Requested: 2 min ago                │
│                                         │
│ [!] def456 - Modify Cargo.toml          │
│     Risk: 3/5  Agent: devops-agent      │
│     Requested: 5 min ago                │
├─────────────────────────────────────────┤
│ [A]pprove  [R]eject  [D]etails  [Q]uit  │
└─────────────────────────────────────────┘

Best Practices

For Agents

1. Always provide clear context
2. Calculate accurate risk scores
3. Break large operations into smaller requests
4. Handle timeout gracefully

For Reviewers

1. Review operation context thoroughly
2. Check affected files/systems
3. Verify agent has appropriate scope
4. Document approval/rejection reasons

Notifications

Slack Integration

# Configure webhook
ccswarm config set hitl.slack_webhook "https://hooks.slack.com/..."

# Notifications sent for:
# - New high-risk requests
# - Approaching timeouts
# - Auto-approved operations

Email Notifications

ccswarm config set hitl.email "team@example.com"

Audit Trail

All HITL decisions are logged:

# View audit log
ccswarm hitl audit --last 24h

# Export for compliance
ccswarm hitl audit --format json > hitl_audit.json

Emergency Override

For critical situations:

# Emergency approval (requires sudo/admin)
ccswarm hitl emergency-approve abc123 \
  --reason "Production incident mitigation" \
  --admin-key $ADMIN_KEY

Configuration Options

timeout_seconds

risk_threshold

max_pending

auto_reject_timeout