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

# Context Injection

> Attach files, diffs, selections, issues, and errors to an agent run

An agent is only as good as what it can see. Mains lets you attach precise context to a run -- files, diffs, terminal output, a highlighted element, a linked ticket, a Sentry error -- so the agent starts with exactly the material it needs instead of hunting for it.

Everything you attach appears as a **chip** in the workspace input and travels with your next message.

## **Adding Context**

The fastest way is the context picker in the workspace input. It groups everything you can pull in:

| **Section**  | **What it attaches**                                        |
| ------------ | ----------------------------------------------------------- |
| **Files**    | Any file from the workspace, by fuzzy path search           |
| **Issues**   | Tickets linked to the project, from your connected trackers |
| **Skills**   | A [skill](/skills) to run                                   |
| **Plugins**  | A [plugin](/plugins) capability                             |
| **Commands** | A provider [slash command](/claude/slash-commands)          |

Other sources attach from where they live:

| **Source**            | **How**                                                                        |
| --------------------- | ------------------------------------------------------------------------------ |
| **Diff**              | Send the workspace's current change set -- see the [Diff Viewer](/diff-viewer) |
| **Browser selection** | Highlight an element in the [Browser Panel](/browser) -- screenshot + selector |
| **Terminal output**   | Capture output from the built-in [Terminal](/terminal)                         |
| **Signals**           | A [Sentry](/sentry) error with its stack trace                                 |

## **Context Kinds**

Under the hood each attachment is recorded against the run with a kind, so the agent receives structured material rather than a blob of text:

| **Kind**    | **Holds**                                       |
| ----------- | ----------------------------------------------- |
| `file`      | A file path and its contents                    |
| `selection` | A selected region of code or a captured element |
| `diff`      | A change set                                    |
| `git`       | A commit or branch reference                    |
| `terminal`  | Captured command output                         |
| `note`      | Freeform text you typed                         |

Small items are stored inline; large ones are saved once and linked by hash, so re-attaching the same file doesn't duplicate it.

## **The Context Window Meter**

Every attachment -- and every turn of the conversation -- spends tokens. The circular **Context window** indicator in the workspace shows how full the agent's window is:

| **State** | **Meaning**                        |
| --------- | ---------------------------------- |
| Green arc | Plenty of room                     |
| Amber     | Nearing the auto-compact threshold |
| Red       | The window is nearly full          |

Hover it for the exact used / total token counts. Use it to decide when to start a fresh run instead of piling more context onto a saturated one.

<Note>
  Attaching context is additive and reviewable -- chips show exactly what the agent will receive, and you can remove any of them before sending. Nothing is sent until your next message.
</Note>

## **Related**

* [Diff Viewer](/diff-viewer) -- review and attach change sets
* [Browser Panel](/browser) -- capture page elements as context
* [Linked Tasks](/linked-tasks) -- pull issues from connected trackers
* [Sentry](/sentry) -- attach errors as signals