---
title: "Starting and reviewing a task"
---

# Starting and reviewing a task

Start a session in the Agents app, choose the folder, isolation, engine, permissions and model, set your intent and context, then review the agent's changes, give feedback and apply them.

This page walks through one task from start to applied change.

## 1. Start a new session

Choose **New Session** (or **New Session in Folder...**). The new-session view opens with the chat at full width.

Above the input, choose:

| Picker | Choice |
|---|---|
| Workspace | The folder to work in. **Open Folder...** adds one; no workspace makes a quick chat. Before an agent works in a folder you are told: "An agent session will be able to read files, run commands, and make changes in this folder." |
| Repository | The repository, when the workspace holds several |
| Harness | The engine that runs the session, for example Claude or Codex |
| Isolation | **New Worktree** creates a separate copy for the session on a new branch; **Branch** works in the existing repository |
| Branch | For a new worktree, the base branch (the current branch's upstream when there is one, otherwise the default branch); for Branch, the branch to check out |

Inside the input, depending on the engine: **Add Context** (files, folders and other attachments), **Agent**, **Mode**, **Permissions** and **Model**.

Your choice of Permissions or Mode is a preference: your organisation's policy wins over it, and bypass modes are turned off under governance. Each action is still decided when the agent takes it.

### Set your intent

When your organisation's policy requires an intent for the repository, the new-session toolbar asks for one before the session starts: the work item from your tracker and the requirements the task addresses. The agent host refuses sessions in that repository until it is set, including sessions in a worktree. See [Intent, projects and releases](https://neuralinverse.com/docs/editor/intent-and-projects/).

### Choose organisation context

When your organisation uses the Context Platform, the new-session toolbar has **Context lens**, **Context snapshot** and **Context status** entries. They set the selection the first request of the next session takes. Without a selection, the space of the project the repository belongs to is used. See [Organisation context](https://neuralinverse.com/docs/editor/context/).

## 2. Describe the task and submit

Write the task in the input and submit. The session appears in the sessions list. While it runs:

- **Needs approval**: when an action needs approval, the session shows the rule and the reason and waits for you. See [Approvals and governance](https://neuralinverse.com/docs/agents-app/approvals/).
- **Needs input**: the session row and section show when an agent is waiting for you.
- The agent's terminals do not inherit your credentials, and tool results are redacted before they reach the model. See [Secrets and agent terminals](https://neuralinverse.com/docs/editor/secrets/).

You can start other sessions meanwhile. See [Sessions and parallel work](https://neuralinverse.com/docs/agents-app/sessions/).

## 3. Review the changes

The side pane opens on **Changes** when the session reports its first file changes.

- **Changes** is a multi-diff of every file the session changed, with a Branch Changes dropdown and diff statistics. **Collapse All Diffs** and **Expand All Diffs** help with long changes; the changes tree lists the files.
- The **Changes** pill in the session header opens the Changes tab from anywhere.
- **Files** shows a file or the explorer; **Browser** is the integrated browser. **+** adds a tab.
- The detail panel on the right shows the branch's changes and checks for Changes, or the file explorer for Files.

### Give feedback on a line

In the diff, choose **Add Feedback** on a line or range, write a comment, and **Add** or **Add and Submit**. Feedback comments are attached to your next message to the agent, so it can fix what you pointed at.

### Keep or discard

Keeping what an agent changed is itself an action your organisation's policy decides. When a control or approval rule covers the change set (for example two reviewers for safety-classified code), keeping it needs signed approvals from other enrolled people: choose **Request Approval** in the dialog, commit `.neuralinverse/records`, and try again once they have signed. See [Approvals and signatures](https://neuralinverse.com/docs/editor/approvals/).

## 4. Apply the result

| Action | When |
|---|---|
| **Apply Changes to Parent Repository** | The session worked in a worktree: merges the worktree's branch into the parent repository ("Applied changes to parent repository."). If the parent has diverged, the merge fails and you resolve the conflicts by hand. |
| **Create PR** | Open a pull request for the session's branch. See [Pull requests](https://neuralinverse.com/docs/agents-app/pull-requests/). |
| **Open in Editor** | Continue in the Editor with the session's folder and chat. See [Working with the Editor](https://neuralinverse.com/docs/agents-app/editor/). |

Once the git hooks are installed, commits made in the app and in its terminals carry your intent's `Work-Item:` trailer, and the commit gate runs as for any commit.

## Layout shortcuts

The side pane has three states (editor and detail, detail only, editor only) and can be closed for a chat-only view. On macOS:

| Shortcut | Action |
|---|---|
| ⌥⌘L | Toggle Details (when Changes or Files is active) |
| ⌥⌘E | Maximise or restore the editor area |
| ⇧⌘K B | Add a Browser tab |
| ⌘K S | Add a Search tab (workspace sessions) |
| ⌘K B | Add a Files tab |

**Hide Editor** and **Show Editor** in the tab strip collapse and restore the editor content while keeping the detail panel. **Toggle Side Panel** closes or opens the whole side pane. **Toggle Sessions List** collapses the list on the left.