Conversation the graph answers, when it came from a conversational channel.
How many continuation hops preceded this one, so a re-submitted graph can carry depth + 1.
The agent run that emitted the graph — the target of a reinvoke.
One-line human-readable roll-up.
Per-task outcome. Deliberately carries a summary and an outputRef, not inline payloads —
a ten-task graph's full outputs would swamp the continuation turn's context window, and the
agent can pull what it actually needs by task ID.
Everything a continuation needs to say what happened.