# What the AI receives

Every task carries its own context: the brief, and the accepted knowledge that work needs. Legevio puts it together when an AI reads the task, so a new chat or agent does not depend on what an earlier one remembered.

## See it before the AI does

There are two places to look:

- **In a task:** the **Task Context** tab shows the **Agent Brief preview**, with what is **Included in this brief** and what the AI **Will NOT receive**.
- **For the whole picture:** open **Context Pack** in the sidebar and choose **What AI receives**. Pick a task, and two columns show **This AI will receive** and **Will NOT receive**.

Every source Legevio selects for the task is listed. There are no hidden sources. What you see is what an AI gets when it reads the task, whether it is a role inside Legevio or the AI you connected over MCP.

## What goes in

- **The brief.** What to do and what done looks like.
- **Accepted Project Memory.** Accepted records, the project's rules and standards, the accepted text of the project documents, and filled collections.
- **What you attached for the AI.** Files, outside references and board cards added to the task for the AI to see. A file goes in as its text; the original file is not sent. A file with no text, such as an image, goes in by its title only.
- **Ground rules.** Every context reminds the AI to prefer accepted Project Memory, to treat drafts as proposals, and not to use rejected or deprecated material.

In a large project (more than about 40,000 characters of accepted records), only the records the task names or best matches travel in full. The rest come as an index: title and one line each, shown as **In the index — title only**. The AI fetches the full text of the ones it needs.

## What stays out

- Records that are not accepted: ideas, drafts, and anything still in review. They are listed as excluded, not approved for AI context.
- Rejected and deprecated records.
- Document text nobody accepted.
- Attachments meant for people only.
- Legevio's own bookkeeping, such as where a record came from.

The project state page is not part of a task's context. Your AI reads it separately at the start of a session.

## When the inputs change

When your AI reads a task over MCP with permission to hand in results, Legevio keeps a receipt of what it handed out (for 30 days). If one of those inputs changes before the result is reviewed, the result says **Made from an older version of:** and names the record. You can still accept it; it then stays marked as affected work until someone reviews it.

When an accepted record changes later, the work that was built on the old version is listed as affected: the record shows **Inputs changed** and the task **Prior run inputs changed**. **Review affected work** gives three choices, in this order:

1. **Compare source data**: what the work used against what it would get now.
2. **Recreate selected tasks**: new linked tasks in Backlog. Existing tasks keep their status, and no AI run starts.
3. **Leave selected as is**, with a reason.

Legevio never reruns work on its own, because a run costs your AI's time or your own key. A missing link does not prove that work is unaffected, so the list says what it knows and no more.

## For your AI

- `get_task` returns the brief, the context and a receipt to send back with the result.
- `get_task` with `includeContextPack: false` rereads a task's status or revision without the context. It is many times smaller.
- `get_project_memory` with `ids` fetches the records the index lists.

See the [MCP tools reference](/reference/mcp-tools/).