Concepts
The mental model behind ScaiFlow.
Read these in order on first contact:
- Architecture — the pieces (canvas, backend, compiler, ScaiGrid, ScaiCore) and how they fit together.
- The flow graph — nodes, edges, ports, the JSON shape on the wire.
- The rigidity spectrum —
@rigid,@guarded,@flexible, and the "adaptive approval pattern" that composes the third with@checkpoint. - Node kinds — the 24 building blocks, grouped by category.
- Triggers — entry nodes compile to top-level
IRTriggerdeclarations, not to in-flow blocks. The closed set is{api, webhook, schedule}. - Models and the registry — declare every model your flow needs by role; LLM nodes pick a role.
- Human-in-the-loop and checkpoints — pause execution, route to a review queue, resume on response.
- Sandboxes (ScaiBunker) — provision, exec, file I/O, destroy.
- Queues (ScaiQueue) — topology, publish/consume, escalation.
- Sub-flows — compose flows from other deployed Cores.
- Publish as model — turn a deployed Core into a chat model reachable over OpenAI-compatible chat completions.
- Multi-tenancy and auth — ScaiKey OIDC, per-tenant credentials, per-flow ACLs.
- Compilation targets — YAML for ScaiGrid (default), SCIR MessagePack for verification,
.scaitext for inspection.