Orbit v1 - Core Platform

Orbit v1: Core Platform (Phase 2)

Scope

• orbit-api (Spring Boot 4, Spring Modulith): modules tenancy, workflows, prompts, tools, runs; versioned workflow definitions validated against JSON Schema; Flyway; ProblemDetail; OpenAPI
• Multi-tenancy: Keycloak (orgs/realms), JWT resource server, API keys (hashed, prefix lookup), Postgres RLS for tenant isolation
• Idempotency-key starter (custom Spring Boot starter backed by Redis)
• orbit-llm-gateway v1 (Go): provider adapters (Anthropic, OpenAI, Ollama) behind a Strategy interface; streaming SSE passthrough; timeouts, retries, fallback; token counting; per-tenant token bucket in Redis Lua; exact-match cache
• orbit-worker v1 (Go): executes llm, tool, http, branch steps from orbit-dag; the tool-calling loop (model → tool_use → execute → tool_result → model) with max iterations, per-tool timeout, output truncation
• Tool registry: JSON-Schema tool definitions; built-in tools: http_request, sql_readonly, calculator, github_create_issue
• gRPC between orbit-api (Java) and engine/worker (Go) with buf
• GraphQL for the UI: workflows → versions → runs → steps (DataLoader / @BatchMapping)
• orbit-stream (Go): SSE for tokens + run events via Redis pub/sub; Last-Event-ID resume
• orbit-web: playground (chat + streaming), run timeline, workflow JSON editor
• Docker Compose for everything; GitHub Actions CI (Java + Go matrices)

Use cases shipped

UC1 Support triage · UC2 v0 (chat + memory, no RAG yet)

ADRs

ADR-001 Modular monolith control plane · ADR-002 Go data plane · ADR-003 Tenant isolation via RLS · ADR-004 Workflow definition format (JSON DSL vs code) · ADR-005 Gateway quota algorithm

Definition of done

• 200 concurrent streaming chats without errors; p99 time-to-first-token overhead of the gateway < 30 ms
• Cross-tenant access attempts all fail (automated test suite)
• Blog: “Building an LLM gateway in Go: streaming, quotas, and fallbacks”