[PRD] Operations HITL Spike

[joebudi]

Operations HITL Spike — PRD

Goal
Validate that we can support human-in-the-loop (HITL) execution using existing infrastructure (Bull) while preserving the request-driven model.

---

Problem
We need to confirm whether Bull can support:

• pausing execution
• escalating to a human
• resuming execution

…without turning the system into queue-driven automations or losing clear request state and lifecycle.

---

Solution
Build a minimal end-to-end flow:

• Trigger: Slack message creates a request
• Request enters Processing
• Agent evaluates and triggers escalation
• Request moves to Needs Input
• User responds (via Slack/UI)
• Request resumes → returns to Processing
• Request completes → Completed

---

Success Criteria

• Request state is explicit (Processing → Needs Input → Completed)
• Request is the primary object (not Bull job)
• Full history is reconstructable from request (not queue)
• Bull handles execution only (not lifecycle/state)
• Both happy + sad paths demonstrated

---

Out of Scope

• Final architecture decisions
• Scalability guarantees
• UI polish
• Production-ready durability

---

Timeline
~1 week spike, demo on Thursday (16th)