Chat: Streaming & Live Thinking
Understand the live streaming view: text, tool cards, and code execution steps.
The engineer assistant streams its work as it happens. Instead of waiting for one final block of text, you watch the assistant think, call tools, run code, and write its answer in real time. This page explains what each part of that live view means.
What you see as the assistant works#
A turn is rendered as an ordered sequence of steps. They appear in the order the assistant produces them and stay on screen so you can scroll back through the reasoning.
Live thinking#
Before answering, the assistant may surface that it's thinking through the problem. These markers are transient — once the turn finishes, the thinking step disappears, since the final text and the tool cards are what matter on review.
Tool cards#
When the assistant uses one of its ClearSpan tools (for example, listing your projects, creating a calc, or running a solver), a tool card appears:
- It starts in a running state when the tool is called.
- It fills in the inputs the assistant passed once they're complete.
- It updates to done (or error) with the result when the tool returns, along with how long it took.
Tool cards let you audit exactly what the assistant did on your behalf — which project it touched, what parameters it used, and whether the call succeeded.
Tool actions run against your live ClearSpan workspace. A tool card that shows a calc was created means that calc now exists in your projects.
Code execution steps#
For tasks like parsing a spreadsheet or crunching numbers, the assistant may run code in a sandbox. A code-execution card streams the code as it's written and then shows a short summary of the output when it finishes.
Answer text#
The assistant's written answer streams in token by token. You'll see it build up as it goes, rather than appearing all at once.
Why streaming can pause#
A turn may pause visibly between steps — for instance, while a tool call runs or code executes. That's expected; the assistant is doing work, not stuck. The view resumes as soon as the next result is ready.
If the stream stops with an error message instead of finishing, it's usually a transient issue or a usage limit. See Troubleshooting: Errors & Rate Limits.
Streaming and your transcript#
The streamed steps are saved with the conversation, so when you reopen it from Chat History, the same tool cards and text re-render. Thinking markers aren't re-rendered on reload — they only matter live.
Where to use this#
Streaming is part of the engineer assistant chat. For how to open the assistant, pick a model, and what it's capable of, see Engineer Chat Assistant.