Drop-in replacement for Oban.Worker with DAG dependency gating, result passing, PubSub broadcasting, and retry idempotency.

Retry idempotency

Before running dep checks or calling perform_workflow/1, the worker checks whether this job already stored a result on a previous attempt. If it did, it returns :ok immediately — no LLM call, no side effects, no token spend.

This handles the case where:

• The LLM responded successfully
• store_result wrote to the DB
• The node crashed before Oban marked the job completed
• Oban rescues the job and retries it

Without this guard, the retry would re-call the LLM unnecessarily.

For LLM workers, the only additional thing you need to ensure is that your perform_workflow/1 has no side effects that happen before the LLM call that would be harmful to repeat. Side effects after a successful LLM call (e.g. writing to another table) should be guarded separately if needed.

Lifecycle broadcasts

Each step broadcasts to "workflow:{workflow_id}" and "workflow:all":

• :executing — immediately when perform/1 is invoked
• :snoozed — when deps aren't ready yet (no retry consumed)
• :completed — after perform_workflow/1 returns :ok / {:ok, result},

              or when a cached result is returned on retry

• :retryable — after performworkflow/1 returns {:error, } with retries remaining
• :discarded — after performworkflow/1 returns {:error, } with no retries left
• :cancelled — when a dep was cancelled/discarded/pruned

Usage

For general workflow steps:

defmodule MyApp.Workers.SomeStep do
  use Baton.Worker, queue: :default, max_attempts: 3

  @impl true
  def perform_workflow(job) do
    {:ok, %{result: "something"}}
  end
end

For LLM steps, prefer Baton.LLMWorker which sets the :llm queue, jittered backoff, and a configurable timeout/1 automatically.