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

# Interactions

> Mentions, file attachments, session management, and keyboard shortcuts

Beyond typing messages, Craft Agent supports several ways to interact with your conversations — from mentioning sources to attaching files and managing sessions.

## Mentions

Use `@` mentions in your messages to activate sources and skills inline.

### Source Mentions

Type `@` followed by a source name to tell the agent to use that source:

```text theme={null}
"Check @linear for my open issues"
"Search @slack for messages about the deployment"
"Read my latest emails from @gmail"
```

When you mention a source, it's automatically activated for the conversation — the agent gets access to that source's tools without you needing to configure anything.

### Skill Mentions

Similarly, mention skills with `@` to invoke them:

```text theme={null}
"Run @daily-standup"
"Use the @code-review skill on this PR"
```

<Tip>
  As you type `@`, an autocomplete menu appears showing available sources and skills. Select one to insert the mention.
</Tip>

## File Attachments

You can share files with the agent by:

* **Drag and drop** — Drag files directly into the chat input
* **Paste** — Paste images or files from your clipboard
* **File picker** — Click the attachment icon to browse and select files

The agent can read and process most common file types:

| Type          | Examples                  | What the agent can do               |
| ------------- | ------------------------- | ----------------------------------- |
| **Code**      | .py, .ts, .js, .go, .rs   | Read, review, refactor, explain     |
| **Documents** | .pdf, .docx, .xlsx, .pptx | Extract text, summarise, convert    |
| **Images**    | .png, .jpg, .svg, .webp   | View, describe, resize, convert     |
| **Data**      | .json, .csv, .xml, .yaml  | Parse, analyse, transform           |
| **Notebooks** | .ipynb                    | Read cells, outputs, visualisations |

## Sending While the Agent Is Responding

You can send a follow-up message at any time — even while the agent is still streaming a response or running a tool. How that message lands is controlled per **AI connection** with two modes:

| Mode      | What happens                                                                                 | When to use                                                         |
| --------- | -------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- |
| **Steer** | Your message is delivered into the in-flight turn so the agent picks it up before finishing. | Course-correcting an ongoing task ("actually, use Python instead"). |
| **Queue** | The current turn finishes naturally; your message is replayed as a new turn afterwards.      | Adding a follow-up without disturbing the current work.             |

### Visual cue

When you send a message that gets queued, the user bubble shows a small **Queued** chip until it's picked up. Steered messages are delivered immediately and don't show the chip.

### Defaults

The default depends on the backend:

* **Anthropic (Claude Agent SDK)** → `queue`. Claude's emulated steer relies on the model invoking a tool before the turn ends; if it doesn't, the steer falls back to a re-queue and you've paid for the original turn's tokens for nothing. Defaulting to queue is more predictable.
* **Pi-backed providers** (OpenAI, Gemini, OpenRouter, Ollama, custom endpoints) → `steer`. Pi's native steer is non-destructive — your message is delivered after the current tool call finishes, full context preserved.

### Changing the mode

Open **Settings → AI**, scroll to the **Connections** section, click the **`…`** menu on the relevant connection (or right-click the row), and choose **Mid-stream sends → Steer immediately** or **Queue until ready**. The choice is saved on the connection and applies to every session that uses it. The setting can be flipped at any time and takes effect on the next mid-stream send.

<Tip>
  If you're unsure, leave the default. The two modes deliver the same end result for a calm follow-up — the difference only matters when you're actively redirecting an in-flight task.
</Tip>

For the schema-level field, see [LLM Connections → midStreamBehavior](/reference/config/llm-connections#mid-stream-behavior).

## Session Management

Every conversation in Craft Agent is a **session**. You can manage sessions from the sidebar:

### Flagging

Flag important sessions to mark them for follow-up. Flagged sessions appear with a flag icon and can be filtered in the sidebar.

### Archiving

Archive completed or inactive sessions to keep your sidebar clean. Archived sessions are hidden from the default view but can be accessed via the filter.

### Batch Operations

Select multiple sessions to perform bulk actions:

* **Archive** multiple sessions at once
* **Delete** sessions you no longer need
* **Apply labels** across sessions

### Read / Unread

Sessions are marked as unread when the agent produces new output. Click a session to mark it as read.

## Keyboard Shortcuts

Craft Agent includes keyboard shortcuts for common actions. Open the shortcuts reference with:

* **macOS:** `Cmd + /`
* Or click the **?** icon in the bottom corner

Key shortcuts include:

| Shortcut      | Action                            |
| ------------- | --------------------------------- |
| `Cmd + N`     | New conversation                  |
| `Cmd + K`     | Quick search / command palette    |
| `Shift + Tab` | Cycle permission modes            |
| `Cmd + Enter` | Send message                      |
| `Escape`      | Stop the agent's current response |
| `Cmd + /`     | Show keyboard shortcuts           |

<Note>
  Keyboard shortcuts can be customised. Ask the agent: "Help me customise my keyboard shortcuts" to modify bindings.
</Note>