Developer refactors AI agent orchestrator from spaghetti architecture to main-bus design inspired by Elm, separating decision logic from execution

The orchestrator originally mixed state transitions, conditional logic, and side effects in imperative code across multiple functions. Problems included silently dropped issues, stuck states, workers losing state during retries, and infinite ping-pong cycles between workflows due to mismatched conditions—burning precious Claude tokens.

The new architecture uses a pure function called a "reducer" (the bus) that takes current state and an event, then outputs the next state and a list of side effects as data. A separate interpreter performs those side effects. The bus makes decisions without checking the clock or calling external systems; the interpreter only executes what the bus prescribes.

Debugging became traceable: bugs are now either bus bugs (wrong decision) or interpreter bugs (wrong execution). The reducer is deterministic—given the same state and event, it always produces the same output—making issues reproducible by replaying an append-only event log against initial state.