Read what previous nights concluded about this dataset.

Call this FIRST, before walking any lineage. If a previous incident already explains today's symptom, you can skip the investigation entirely.

Read Nightshift memory for several assets at once, e.g. a lineage path.

Find every asset in the graph that already suffered a given failure mode.

Failure modes are slugs such as silent-schema-change or late-upstream.

Open an incident on an asset in DataHub, so the break is visible to humans.

Resolve an incident with a message explaining what was actually done.

Leave an assertion watching a column, so this break cannot recur silently.

Call this after a fix. It is what turns one repaired incident into a permanent guarantee for everyone downstream.

Write tonight's conclusion into the graph, for humans and for the next agent.

This is the last thing you do on any incident, and it is not optional. Write a conclusion, never a status: the next night reads this instead of starting from nothing.