Human-in-the-Loop Approval Workflows #9
Description
Summary
Allow agents to pause execution and request human approval before proceeding. Use cases: destructive actions, financial transactions, sending external communications. Approval via dashboard notification or webhook callback.
Motivation
LangGraph's strongest differentiator — state checkpoints with human review gates. Critical for regulated industries (finance, healthcare). Google ADK supports agent callbacks that can interrupt flow — wire this into MATE's dashboard with a proper approval queue.
Scope
- New approval queue system (DB-backed)
- Dashboard approval page: pending approvals, context preview, approve/reject/modify
- WebSocket or SSE notifications for new approval requests
- Configurable timeout + fallback action (auto-reject, escalate, retry)
- Approval config per agent: which tool calls or conditions trigger approval
- Audit trail for all approval decisions
Acceptance Criteria
- Agent pauses and creates approval request for configured triggers
- Dashboard shows approval queue with full context
- Approve/reject actions resume/cancel agent execution
- Timeout with configurable fallback
- All decisions logged in audit trail