> ## Documentation Index
> Fetch the complete documentation index at: https://docs.attensira.com/llms.txt
> Use this file to discover all available pages before exploring further.

# The set-up run

> The one session that runs itself once your first readings are in: it reads your site, groups what is wrong by the source that decided it, and files a starting plan.

## What it is

Once your [first reading](/start/first-reading) reaches its floor — **at least half the reads back** — one session starts on its own, without you asking for it. It is the first thing the product does with your money, and the only run allowed to look at a workspace it has never seen before.

It runs **once, ever**, per workspace.

## What it does

1. **Reads your site.** The home page and whatever pages answer a buying question — never a claim about a page it did not actually fetch.
2. **Reads the losing prompts** from your first reading: where you were not named, on which engine, against which rival.
3. **Groups what it finds by deciding source** — the same grouping [Who decides an answer](/measure/deciding-sources) uses — so a domain behind four losses is one piece of work, not four.
4. **Files at most three rows** in your [Inbox](/agent/inbox), each naming one source and the decision that is yours to make, backed by a verbatim excerpt from the answer.
5. **Proposes one fix as a pull request**, when a repository is [connected](/agent/github) — the single best on-site change among the sources it found. Never more than one; a first session that opens four pull requests is a first session nobody merges.

## If it finds nothing wrong

A workspace with nothing to fix does not end with an empty run. Instead, the session takes your **worst prompt** — lowest rate, largest sample — and files one plan for winning it: the page that should answer it, the claim that page is missing, and the evidence behind each part of that plan.

A set-up run never ends with nothing filed.

## If credits cannot cover it

The balance is checked before anything is read or dispatched. If your workspace cannot afford the run, nothing is read and nothing is charged — but the refusal is never silent. It is written down three ways: a **blocked** row in your Inbox, an email, and a session you can open that says exactly why it did not start. Top up, and the set-up run can be started again.

## What it never does

* It never reports a trend. There is one day of readings and nothing to compare it to yet.
* It never lists every page it read or every prompt you are named on — only what needs a decision.
* It never files more than three Inbox rows, and never fewer than one.

<Note>
  This is a once-ever run. The standing job that keeps watching after it is the hourly pass described in [What the agent notices on its own](/agent/findings).
</Note>