Artifacts panel
Open Artifacts from the right-panel launcher. When a tool action in an activity trace produced an artifact, its link opens this panel and focuses that artifact. The panel has three sections:
CPython cells open in the REPL panel, and dashboards open in the Dashboards panel, so neither appears under Technical artifacts.
Dashboards
Ask for a dashboard in the chat, for example “Make a dashboard of the test results you just found.” Prime Agent researches the request with its normal tools and states its findings in the reply first. The dashboard only lays out figures that reply already states. It does not search, browse, or fill gaps. The reply ends with a dashboard attachment. Select it to open the board in the Dashboards panel, which lists every dashboard in the session. Select a board in the list to draw it.A message creates at most one dashboard. If a reply asks for more than one, Fleet opens the first and shows “This message asked for more than one dashboard. Fleet opens the first.”
Who builds the board
How Fleet builds the board depends on whether a Thesys API key is set in Settings → Dashboards or in the server environment:
A board keeps the mode that created it. Adding a key later does not send existing local boards to OpenUI Cloud. Removing the key does not delete boards OpenUI Cloud already built, but you need to add the key back to change them.
See Configure dashboards to set the key and model.
Where the figures come from
The assistant copies the figures it stated into a small table that travels with the dashboard. Every chart and table on the board reads that table, so the board never shows a number the reply did not state. Numbers are not baked into the layout: the board reloads its data every time it opens. The table has these limits:- At most 12 columns and 200 rows.
- Column keys start with a letter and use letters, digits, and underscores.
- Each column is text or a number. Text cells are cut at 500 characters.
session_outputs data tool instead of a copied table. See Data for generated UI.
Attachment states
The attachment in the transcript shows the board’s state:
A failed build stays failed. Fleet does not substitute a different board.
Interact with and change a board
Buttons on a board send their message to the chat, the same as if you had typed it. Use them to ask follow-up questions from the board. To change a board, ask the assistant in the same session, for example “Add a chart of failures by package.” The assistant updates the latest dashboard instead of creating a new one, unless you ask for a new one. It keeps the stored figures unless the change needs different ones. With OpenUI Cloud, an edit sent while the board is still building waits briefly for the build to finish instead of failing. Editing a Cloud board without a key fails with “This dashboard was built by OpenUI Cloud. Add the Thesys key back to change it.”Configure dashboards
Dashboards work without any setup. To have OpenUI Cloud build them, open Settings → Dashboards:- Thesys API key lets OpenUI Cloud build and edit dashboards. Fleet stores it on the server in
~/.prime/agent/thesys.jsonand never sends it back to the browser. A key saved here overridesTHESYS_API_KEYfrom the server environment. Select Remove key to fall back to the environment, or to the assistant writing dashboards when no key is set. - Dashboard builder model is the model OpenUI Cloud builds with. It is separate from the model you chat with. Use a provider-qualified id such as
openai/gpt-6-luna. The provider must be one you added a key for on the Thesys BYOK page; any other provider is billed to Gateway managed credits instead, which fails once those run out. The model comes from this setting, thenFLEET_DASHBOARD_MODEL, then the defaultanthropic/claude-sonnet-5.5. Select Use default to clear a saved model.
VITE_OPENUI_CLIENT_KEY. Events are sent without response text. See Configuration.
Data for generated UI
Generated OpenUI programs, including dashboards, read Fleet data throughQuery(...) calls. Fleet routes each call to a read-only tool on its own server, so generated UI never reaches the runtime directly and cannot write anything.
Each tool returns at most 200 rows, 50 by default.
session_outputs always reads the session that is open, not one named in the arguments. It leaves out reasoning, and each preview is cut at 240 characters. Its total and toolCalls count the whole session, not only the returned page.
Dashboards also read dashboard_rows, the copied figures described in Where the figures come from. It returns columns, rows, and columnValues, one array per column key.
For the endpoints behind these tools, see the HTTP API reference.