The problem
A request that touches your calendar, your e-mails and the weather can take ten or twenty seconds. Meanwhile, a plain “LIA is thinking…” tells you nothing. Did she understand? Is she stuck? And once the answer arrives, there's no way to know what produced it.
What I wanted
I wanted you to see what LIA is really doing, as she does it, in your language. And I wanted that account to stay attached to the answer, so you can come back to it later.
While LIA works
The first line is a little wink while LIA analyzes your request, something like “Brain synchronization…”. After that, each line matches a real step: “Preparing action plan…”, “Verifying plan…”, “Retrieving events…”. In ReAct mode, each tool call appears in turn, with a preview of the reasoning in progress.
- What the step did. Each line ends with numbers: ×2 if the tool was called twice, the number of results found, how many LIA read when there were too many to read them all, and the failure on the line where it happened.
- A live display. Past steps fade, the current step breathes, and a cursor blinks while the answer is being written. If you switch tabs, the tab title alternates with “✦ LIA is writing…”. All of this respects your system's reduced-motion setting.
- A very long turn keeps the latest steps on screen and shows how many earlier steps are hidden.
What stays afterwards
As soon as the answer starts, the steps fold into a discreet line under the message, “⚙ N steps · X s”. One click unfolds it: how your request was understood, which tools were called, and the reasoning sorted by category. This account is saved with the message: you find it again after a reload and on your other devices. On a very long turn, it keeps the first steps as well as the last ones, shows the real total and says how many steps in between aren't shown. Only the live reasoning stream isn't kept.
When the connection goes silent
On a phone, putting the tab to sleep can cut the connection without warning: no end, no error. The screen would stay frozen on the last step while the answer is finished and saved on the server. LIA recognizes this silence: after a certain time with no sign of life, she drops the dead connection and reattaches to the turn that kept going in the background.
A label for every tool
A step without a label falls back on a generic “thinking…”, and you no longer know what's happening. So each tool declares its own label, translated into every interface language, and a test blocks any new release if one of LIA's tools lacks one.
The limits
These steps say what LIA does; they don't explain each of her choices. They only show numbers, without the names of the items read: a message's trace must not become a copy of your e-mails. And the reasoning shown is the model's prose, useful for understanding, not proof. To know what really happened, rely on the answer's report and on the registers that list what LIA did and consulted.
Read more: A plan before acting: how LIA prepares and checks your multi-step requests; Seeing what's going on: how LIA watches over itself.