Prerequisites
- The deployment has conversations enabled.
- The deployment has a configured and consented model provider.
- You have the Editor or Admin role.
Start a conversation
Open a conversation with a subject, and optionally the incident it is about. A conversation started from an incident already knows the alert, so you do not have to describe it. Send your first question and OpenCluster starts investigating. Send another when the answer arrives, and it continues from where the last one stopped. Several people can take part in one conversation, and every message records who sent it. Two people looking at the same incident can also hold separate conversations, so one line of questioning does not derail another.What a follow-up already knows
What is not carried forward is another conversation’s messages. Two conversations about
one incident share the incident, never each other.
Ask something that is not an incident
A conversation is also how you ask an operational question in peacetime — “which version is currently deployed?” — and get a direct answer rather than a causal report. The answer still cites the reads that support it.Send a message while OpenCluster is working
You do not have to wait for a green light. A message sent mid-investigation is accepted and taken up as soon as the current one finishes. Send three and they are answered together, as one question in three parts, rather than as three competing investigations. The response tells you which happened: either the investigation it started, or that it was queued.Watch an investigation as it runs
Each investigation publishes an ordered stream of what it is doing: which read is running against which integration, what each read returned in one line, when one failed, and the answer as it is produced. You can close the page and come back. The stream resumes exactly where you stopped — nothing missed and nothing repeated. The stream never contains the model’s private reasoning. Every line is written by OpenCluster from what it did, not by the model describing itself.Long conversations
A conversation that runs for hours keeps working. When the record of it grows past what the model can hold, OpenCluster consolidates the older part into a structured summary: the goal, your instructions in your own words, established facts with their citations, what was ruled out, and what is still open. The full transcript is never edited or deleted — only what the model carries changes. If a single question is too large to answer inside that budget, the investigation concludes with what it established and says it stopped for context, exactly as it does for any other limit. A partial answer is never presented as a finished one.When something goes wrong
See Understand findings and citations for what a
finding’s type and confidence mean.