What is a Session?
A session is a conversation with an agent that can read your code. Ask it how something works, where a behaviour lives, or what a change would involve — and when you decide you want that change made, it writes up a spec you can hand straight to a Task.
Sessions live under Assistant in the sidebar.
Session or Task?
| Session | Task | |
|---|---|---|
| Use it to | Decide what to do | Get it done |
| Interaction | Back-and-forth chat | Fire and forget |
| Your code | Read only | Writes code, opens a pull request |
| You end up with | Answers, plus a spec | A pull request to review |
Start a session when you're still working out what needs to change — "how does auth work here?", "what would it take to add SSO?". Create a Task when you already know.
One button connects them: Create task turns the session's spec into a Task, carrying everything across.
It won't touch your code
A session only reads. It doesn't edit files, run tests, or push anything, and the copy it works from disappears when the session ends.
Ask it to run your test suite and it will decline — and usually add the check to the spec instead, so it gets run later when the Task executes.
With Claude Code this is enforced: the agent physically can't do anything but read and search. Codex sessions rely on the agent following instructions. Either way nothing reaches your repository, but pick Claude Code if you want the stronger guarantee.
Before you start
Four things, all checked for you when you open New session — anything missing is named on screen with a link to fix it.
| You need | Why |
|---|---|
| A runner in your AWS account | Sessions run on your own infrastructure. See Runner Installation. |
| It to be online | An offline runner can't pick up the work. |
| An LLM key for the agent you pick | Anthropic for Claude Code, OpenAI for Codex. See LLM Keys. |
| At least one connected repository | The agent needs something to read. See Adding a Repository. |
Usage is billed by Anthropic or OpenAI on the key you configured.