Hosted demo lane
Seeded state, fake session storage, deterministic AI, and in-memory adapters keep the public demo credential-free.
OD-1025 is the concrete example: an urgent customer email becomes a durable ticket, worker jobs run triage and drafting, retrieval supplies evidence, Watchtower blocks risky auto-contact, and a signed CRM webhook records the review handoff.
sourceText on OD-1025Open artifact2Web actionReact Router appThe route validates the enquiry, creates the ticket, and hands slow work to the background lane.ticket tkt-wifiOpen artifact3Durable statePostgreSQL + DrizzleTicket, draft, follow-up, AI run, and audit rows share one stable ticket ID.tickets, ai_runs, audit_eventsOpen artifact4Worker lanepg-boss jobstriage.ticket and response.template_select run with retryable job boundaries.airun-003 + airun-004Open artifact5Evidence lookupRAG knowledgePolicies, playbooks, similar tickets, and document extracts can be retrieved with metadata filters.Wi-Fi playbook citationsOpen artifact6Model boundaryStructured AI outputThe model produces category, priority, missing info, template choice, confidence, and cost.demo-template-router-v1Open artifact7Safety gateWatchtower says reviewUrgent phone contact is useful, but it cannot auto-send; Priya owns the outbound action.approvalStatus=pendingOpen artifact8Contract eventai.draft.review_requiredThe decision is published with correlation ID, schema version, producer, and idempotency key.evt-1025-human-reviewOpen artifact9IntegrationCRM webhook deliveredSigned payload tells the CRM that Priya must review the urgent Wi-Fi draft.whd-wifi-001 HTTP 200Open artifactSeeded state, fake session storage, deterministic AI, and in-memory adapters keep the public demo credential-free.
Docker profiles turn on Postgres/pgvector first, then optional events, search, graph, and workers.
The same contracts map to Terraform, Kubernetes manifests, CI gates, observability exporters, and signed integrations.
pg-boss is the retryable queue boundary. Each job takes a small ID-shaped input, writes durable output, and leaves an audit trail that Watchtower and events can show later.
The schema separates operational work, agent design, retrieval knowledge, and audit evidence so production services can scale without hiding what the AI changed.