# The live status line

While the Assistant is working you get one status line, not a row of stage pills. It reads `Understanding…` while the model decides what to do, then settles to `Done · N steps` when the answer is finished. N is how many tools ran, and the folded details say which ones. A compact Stop sits beside it and ends the run where it stands.

- **Status:** Available
- **Audience:** both
- **In the app:** #/home
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/assistant/live-stream/

## What the line says, and in what order

The line has one job: tell you whether the Assistant is thinking, working, waiting on you, or finished. It replaces itself in place rather than stacking, so the conversation above it never moves while you read.

| You see | What is happening | What ends it |
|---|---|---|
| `Understanding…` | The model has your question and the screen context and is choosing tools | The first tool call, or an answer that needs no tools |
| A working line with a live step count | Tools are running one after another; each completed one increments the count | The last tool returns |
| `Done · N steps` | The answer is complete and N tools ran to produce it | Nothing — it stays as the record of that turn |
| A proposal instead of a final line | A writing tool wants to run and is waiting for you | You confirm or discard the [proposed action](/docs/assistant/action-cards/) |
| A stopped line | You pressed Stop | Already ended; the steps that finished stay finished |

`Done · 0 steps` is a real and healthy result. It means the model answered from what it already had — the conversation, the screen context, the record you pinned — without querying anything. A short factual question about the record already on screen often costs no steps at all.

## The details, folded

The step list is collapsed by default and opens to show the tools that ran, in order, with what each was asked for. It is deliberately not the same thing as [where an answer came from](/docs/assistant/evidence-and-citation/): the citations name the records an answer rests on, while the step list says what the Assistant did to find them. An answer can cite two records and have taken five steps, three of which found nothing useful.

Opening the details is the quickest way to diagnose an answer you distrust. If `find_anything` ran and returned nothing, the answer came from context and should be read as such. If a read tool ran against a record you were not asking about, the [screen context](/docs/assistant/screen-context/) was wrong, not the model.

## Why one line and not five pills

The Assistant shipped on 2026-09-10 with a single status line in place of the five stage pills it used to show. Five pills implied five phases that each meant something to a person, and they did not — most turns lit two of them for well under a second, which reads as flicker rather than progress. One line that names the current state, keeps a count and folds its detail away carries the same information in a fifth of the vertical space, which matters most on the layout where the panel is the whole screen.

> **Note** The step count counts tool calls. It is not a count of model turns and not a count of records touched: one `data_query` step can read a whole sheet, and three steps can change nothing at all.

## Stop, and what it leaves behind

Stop ends the run at the current step. It rolls nothing back. A writing tool that has already completed has already written — the follow-up exists, the memory is stored — and stopping the run afterwards does not remove it. What Stop reliably prevents is the next step.

Because every write the Assistant performs goes through the service that owns the record, a stopped run leaves the same audit entries a completed one would have left for the steps that ran. Nothing is half-written and nothing is quietly discarded. A completed step that needs reversing is [undo](/docs/assistant/undo/), a different operation with its own record.

> **Careful** Stopping during `Understanding…` is safe in every case, because nothing has been called yet. Stopping mid-run is safe for the work that has not started and irrelevant to the work that has finished.

## Questions

### The line has said `Understanding…` for a long time. Is it stuck?

It can be. A long `Understanding…` with no step count means no tool has been called yet, which points at the model call rather than at your workspace — a stream that dropped, or an AI budget that is spent. [The Assistant stopped responding](/docs/troubleshooting/assistant-stuck/) separates the three causes.

### Does a higher step count mean a better answer?

No. It means more tools ran. A question answered in one step from the record you pinned is usually more trustworthy than the same question answered in six, because fewer steps means less searching and less inference between the source and the sentence.

### Can I see the steps for an older turn?

Yes. The line and its folded details stay with the turn in the conversation, so scrolling back to an answer from an hour ago shows what produced it. Each tab keeps its own scroll, so this survives switching tabs and coming back.

## Related

- [Connect Assistant](https://connectbyjbrh.com/docs/assistant/)
- [Running an action](https://connectbyjbrh.com/docs/assistant/executing-actions/)
- [Where an answer came from](https://connectbyjbrh.com/docs/assistant/evidence-and-citation/)
- [The Assistant stopped responding](https://connectbyjbrh.com/docs/troubleshooting/assistant-stuck/)
- [Deriving control state from the log instead of storing it](https://connectbyjbrh.com/research/derived-control-state/)

## What this page is based on

- `docs-source/sources/GENERAL.md` §8 — the Assistant, shipped 2026-09-10
- Connect capability registry (docs-source/facts.py) — `assistant`, `assistant_tools`
- `backend/app/manager.py` — the Assistant turn and its tool loop