approval-workflow
Manages Human-in-the-Loop (HITL) approval workflows for sensitive actions. Use when creating approval requests, processing approved items, or implementing safety controls for autonomous actions.
When & Why to Use This Skill
The Approval Workflow skill implements a robust Human-in-the-Loop (HITL) framework designed to provide human oversight for sensitive autonomous agent actions. By utilizing a structured folder-based state management system, it ensures that high-risk operations—such as financial transactions, external communications, and critical file modifications—are reviewed and authorized by a human before execution, serving as an essential safety guardrail for AI deployments.
Use Cases
- Financial Oversight: Automatically flagging and pausing any payment requests over a specific threshold or to new payees for manual human approval.
- Communication Safety: Reviewing AI-generated social media posts or emails to new external contacts to ensure brand alignment and accuracy before they are sent.
- Data Protection: Preventing unauthorized or accidental deletion of files by requiring a human sign-off for any destructive file operations.
- Risk Management: Implementing a standardized audit trail and risk assessment for autonomous actions, allowing teams to track the lifecycle of an action from 'pending' to 'executed'.
| name | approval-workflow |
|---|---|
| description | Manages Human-in-the-Loop (HITL) approval workflows for sensitive actions. Use when creating approval requests, processing approved items, or implementing safety controls for autonomous actions. |
| allowed-tools | Read, Write, Edit, Glob, Grep |
Approval Workflow Skill
This skill implements the Human-in-the-Loop (HITL) approval system that ensures human oversight for sensitive autonomous actions.
Core Concept
The approval workflow uses folder-based state management:
/Pending_Approval/ → Human reviews → /Approved/ → Execute
→ /Rejected/ → Archive
Approval Request Format
---
type: approval_request
action: [action_type]
created: [ISO timestamp]
expires: [ISO timestamp]
status: pending
priority: [critical|high|medium|low]
requestor: [agent_name]
risk_level: [low|medium|high]
reversible: [true|false]
---
## Action Summary
[Brief description of what will happen]
## Details
[Full action parameters]
## Risk Assessment
- **Reversible**: [Yes|No|Partial]
- **Impact**: [Description]
- **Sensitive Data**: [Yes|No]
## Instructions
- **Approve**: Move to `/Approved/`
- **Reject**: Move to `/Rejected/`
- **Edit**: Modify, then approve
Approval Thresholds
| Action Type | Auto-Approve | Human Required |
|---|---|---|
| Email (known contact) | Reply only | New recipients |
| Payment | < $50 recurring | New payee, > $100 |
| Social post | Scheduled | Immediate, replies |
| File operations | Create/read | Delete |
| Greetings | Business msgs |
Workflow States
- pending - Awaiting human decision
- approved - Cleared for execution
- rejected - Denied by human
- expired - Timed out without decision
- executed - Action completed
- failed - Execution error
Reference
For detailed implementation, see reference.md
For usage examples, see examples.md