A Model Builder session builds a semantic model in rounds, and reports its progress as cards in your chat with the Warehouse Builder Mind. This page shows how to read those cards. To start, stop, continue or restart a session, see Model Builder.
The session card is the live frame of one build. Its parts follow the work in this order:
The card also states Notes passed to the Model Builder: {n} characters, the size of the notes that went to the session.
While the card loads the live status, it says Reading the live status. If it cannot read the status at all, it says The live status of this Model Builder is not available to you right now. and shows the last stored status.
Two meters track the session's work:
Hover a dot to read it. In a closed session the dots label the finish:
In an open session a dot reads Round {n} · {kind} · {status} · {duration}, or one of the plain lines Round {n}: finished, Round {n}: waiting for the Model AI chat to be free and Round {n}: not started. A round that waits for the Model AI chat is dashed and pulsing. A closed session draws only the rounds it used, so no empty slots remain.
Each round reports back as a card headed Model AI report to the Warehouse Builder:
When only the summary of the round has arrived, the card shows a compact stub and a Load the report button; select it to load the full report.
The The Model AI asks the warehouse to change section lists what the round wants changed in the warehouse. Each request shows its target as {object}.{column}, its kind, a Blocking chip when it blocks, a Proposal: {proposal} line and evidence chips as {toolName} · round {round}. The status chip says where the request stands:
Chart timings lists the charts a round touched. A row shows the chart's name, or Untitled chart, its time, its rows and its peak, with a Measured or Provisional chip. A Provisional timing comes from a single first run of the chart, so the result cache can make it look faster than it is; such rows show Estimated {n} rows instead of a peak. When the set is mixed, the provisional rows are marked one by one. The verdict words are Fast, Slow and Too slow; a chart that reads too slow is noted as Slow: recorded as an open item.
A round hands its work to the Model AI chat, and the handoff card, headed Hand off to Model AI, records the exchange:
The card's details open a drawer with What the Model Builder briefed and What the Model Builder replied, line by line: Summary, Measure, This round, Question, Proof, Answer, Relayed, Chart to time and Notes.
The brief card, headed Brief from the Warehouse Builder, shows what a round was asked to build. It carries Round {n} of {limit} and a for {requester} · fact {name} line, then these sections, each shown only when it has content:
A round that had to wait adds This round waited for the chat to be free, so the Model AI re-reads the model first. The summary is plain text with Show more and Show less; Full brief expands the full brief, and Copy the brief copies it.
When the build asks for more history than the sample window, the full history backfill card appears, headed Full history backfill, and the session card adds a Full backfill section. A load moves through Queued, Loading, Loaded, Failed or Cancelled.
When the session ends, the done card, with the Model Builder report eyebrow, gives the outcome:
While the session still works the card shows the round's progress instead, and after a continue it reads Working again · {round text}. The card also shows Updated {time}, Owner {ownerRole} and the model it built. The scorecard carries four tiles: Measures (all built, or {n} not built), Tests passed ({passed}/{written}, with none failed or {n} failed), Model check (Clean, or the error count, with the warning count beside it) and Rounds ({used}/{limit}, labelled Model AI rounds used).
Below the scorecard:
The footer buttons are Open the session card (while the session still works), Continue with 2 more rounds, Start again, Open Model AI chat and Open model, which opens the semantic model itself (Semantic models). For what each action does, and who may use it, see Model Builder.
A running round can also show a compact Model Builder status card, which reads Reading the session while it loads. It lists the facts: {used} of {limit} rounds, {used} of {budget} tokens, {relayed} questions with the user, {decisions} waiting for a decision, round {n} {status} and backfill {state}. Live data wins over the stored snapshot.
Deliveries in the conversation appear as collapsible rows, titled with what they carry: Model Builder task, Model AI report, Answers from the Warehouse Builder chat, Full backfill finished or Session notice, or Model Builder delivery when no kind is known. A Round {n} chip and, on a full screen, a timestamp sit beside the title. Expand a row to read its content. A note can tell you who may continue the session, for example Only {name} can give it 2 more rounds.
The session card tracks the build while it runs: rounds, tokens, what needs you, the latest report and any backfill. The done card is the final report when the session ends, with the scorecard, the open items and the follow-up actions.
The card could not read the session's live status, so it shows the last stored status. Open it again later; the live view returns when the session can be read again.
Each dot is one round: hover it to see the round's kind, status and duration. Closed sessions label the finish, such as Done in 3 rounds or Failed in round 2. The token bar compares used and budget tokens, and turns warning-coloured at 90% or more.
That round is parked because the Model AI chat was busy with other work. It is dashed and pulsing, and it starts by itself when the chat is free.
The questions relayed to you plus the decisions waiting for you. Answer them on the card; when nothing waits, the section folds to Recently decided.
That is the stub: only the summary of the round is loaded. Select Load the report to load the full report.
Open means the request is recorded with no decision yet, Waiting for you means it needs a decision, Decided means it is settled, and Withdrawn means it was taken back.
Measured comes from real runs of the chart. Provisional comes from a single first run, so the result cache can make the chart look faster than it is.
The history the design approved is the same as the sample window you already have, so there is nothing more to load.
At least one chart was too slow to count. Open Chart timings on the report; the slow chart is recorded as an open item.
They are what is still outstanding, such as questions, change requests, chart timings, backfill, tests and checks. Blocking items are emphasised.