Coordinator Interaction — Steer Document¶

Date: 2026-04-10 (Session 35) Status: Spec only — Phase 1 (streaming) enabled, Phase 2 deferred ADR: ../../adr/ADR-0015-coordinator-interaction-model.md

Problem¶

Curaway's conversational UI only supports AI ↔ patient interaction. Real-world medical travel coordination requires human coordinators for visa logistics, hospital scheduling, payment processing, and clinical clarifications the AI shouldn't handle.

Phases¶

Phase 1: AI streaming (NOW — enabled)¶

SSE token streaming via enable_response_streaming flag
sender_type column on Message model (ai | patient | coordinator)
Zero coordinator features — just future-proofing the data model

Phase 2: Coordinator chat (POST-SEED)¶

WebSocket upgrade for bidirectional real-time messaging
Coordinator dashboard (admin.curaway.ai or /admin route)
Escalation flow: AI detects → emits event → coordinator accepts
AI pauses during coordinator takeover, resumes on release
Internal notes (coordinator-only, patient doesn't see)

Phase 3: Multi-channel (POST-SERIES A)¶

Phone (Twilio) → call summary logged to case timeline
WhatsApp (Twilio) → webhook → message in conversation
Email forwarding → attachment triggers OCR pipeline
All channels unified in case timeline with message.source field

Design Decisions¶

Why SSE now, WebSocket later¶

AI chat is request-response. Patient types, AI responds. No simultaneous typing. SSE handles this perfectly with less complexity. WebSocket becomes necessary when two humans need real-time messaging in the same conversation.

Why `sender_type` now¶

Adding a column with server_default is free — no migration on existing data, no code changes needed. Waiting until Phase 2 means a schema migration under production traffic.

Why not Intercom/Zendesk¶

Clinical context (EHR, match results, documents, consent forms) must render inline in the conversation as rich cards. Third-party chat SDKs don't support custom clinical UI components.

Capacity Planning¶

Tier	Cost/mo	Concurrent users
Current (free tiers)	$20	~7 (Upstash Redis limit)
MVP launch (Redis Pro)	$30	~500-1000
Growth ($100/mo)	$100	~2000-5000
Scale (multi-instance)	$300+	~10,000+

Upstash Redis free tier (10K commands/day) is the immediate bottleneck for streaming. Pro tier ($10/mo) removes this constraint.

References¶

ADR-0015: Coordinator Interaction Model
Session 31: Streaming implementation
Session 35: Architecture assessment + flag enablement