Orchestrating Durable Workflows

Use one workflow per stage and chain them with an explicit auto flag:

• selection -> script -> asset -> render -> youtube_upload

Stage Payload Contract

Pass these fields in every child trigger:

{
  videoId,
  auto,
  triggerSource: 'workflow',
  parentWorkflowId: workflowId
}

Run Persistence

Persist each workflow instance in workflow_runs:

• workflow_id, stage, video_id, parent_workflow_id, trigger_source, status, config_json.
• Move status through: queued -> running -> complete|failed|terminated.

await createWorkflowRun(db, { workflowId, stage: 'script', videoId, triggerSource, config: { auto } });
await markWorkflowRunRunning(db, workflowId);

Chaining Guardrails

• If auto is false, run current stage and stop.
• If policy returns BLOCK, stop downstream stages and write blocking reason to video row.
• Do not overwrite previous stage ownership fields (for example script stage should not change selected format).

Retry Strategy

• Database and state updates: retry with constant backoff.
• AI calls: retry with exponential backoff.
• Expensive render/extract stages: disable retries when duplicate execution is harmful.