Stream LLM Output While Committing Reliable DAG Results

A workflow runs an LLM task whose output must appear incrementally in an interactive client. Downstream DAG nodes, however, require the complete structured result before they can run. Design a path that provides responsive token streaming while committing the final result reliably and unblocking downstream work only when that result is complete and valid. ### Constraints & Assumptions - Client-visible partial text is not equivalent to a committed task result. - A client can disconnect and reconnect while generation continues. - Generation may fail or produce an invalid structured result after partial text has already been displayed. - The source supplies no exact replay-retention, latency, or output-format requirements; identify those decisions explicitly. ### Clarifying Questions to Ask - Must a reconnecting client replay every fragment, or is receiving the latest snapshot sufficient? - Which schema or validation rule makes the final result usable by downstream tasks? - Can generation continue without a connected client, and what cancellation policy applies? - Must provisional output be durable, or only the final artifact and task outcome? ```hint Identify the commit signal Receiving the last visible token is not necessarily the same event as validating and durably storing the result required by the next DAG node. ``` ### What a Strong Answer Covers - Separate provisional streaming and durable completion semantics. - Run, task, attempt, and sequence identity for replay and duplicate handling. - A final artifact commit and readiness transition that cannot be inferred from a client's connection closing. - Bounded buffering, slow-client handling, invalid output, and retry visibility. ### Follow-up Questions - What should the UI show if a retry replaces a failed attempt that already streamed text? - How can reconnecting clients distinguish missing fragments from a new attempt? - Why should a slow browser not hold the task's completion transaction open?