ORCH — CLI AI agent orchestrator: strict TypeScript state machine with layered DDD

[oxgeneral]

Hey Effect-TS community!

As a TypeScript developer who cares deeply about correctness, I built ORCH — a CLI runtime for orchestrating AI coding agent teams with a focus on type safety and clean architecture.

Architecture that Effect-TS fans might appreciate:

// Domain layer — pure types, no side effects
type TaskStatus = "todo" | "in_progress" | "review" | "done" | "failed" | "cancelled" | "retrying"

const VALID_TRANSITIONS: Record<TaskStatus, TaskStatus[]> = {
  todo:        ["in_progress", "cancelled"],
  in_progress: ["review", "retrying", "failed", "cancelled"],
  review:      ["done", "in_progress", "failed", "cancelled"],
  retrying:    ["in_progress", "failed", "cancelled"],
  done:        [],
  failed:      ["todo"],
  cancelled:   [],
}

// All transitions validated — no invalid state mutations possible
function canTransition(from: TaskStatus, to: TaskStatus): boolean {
  return VALID_TRANSITIONS[from]?.includes(to) ?? false
}

Design principles:

• Domain (models, state machine, errors) — zero dependencies
• Application (orchestrator, services, event bus)
• Infrastructure (adapters, YAML/JSON storage, process manager)
• CLI / TUI as thin outer layers

Type safety choices:

• strict: true + noUncheckedIndexedAccess: true
• No as any, no @ts-ignore
• Domain errors extend typed OrchestryError with exitCode + hint
• All IDs are prefixed strings: tsk_, agt_, run_, goal_, team_
• ESM-native: all local imports use .js extension

What ORCH does:
Manages AI coding agents (Claude Code, Codex, Cursor) as a team with task queue, state machine, auto-retry, messaging, TUI dashboard.

GitHub: https://github.com/oxgeneral/ORCH
npm install -g @oxgeneral/orch
1493 tests, MIT

Fellow type-safety obsessives — curious what you think of the state machine design. Have you used Effect for similar state management challenges?