# Coding sessions and the task-to-PR loop

Coding in Volobox happens in **coding sessions**: a sandbox where a coding agent and your team work on one branch of a repository, in view of everyone in the session. A developer can work with the agent directly. A team can also assign a task to an agent and get a pull request back, connected to the task and its reviewers.

![Fieldwork task #142, assigned to the Software Developer agent with coding on, and the Start coding work dialog over it: a note for the agent, OpenCode as the coding agent with Auto model choice in Agent mode, the meridian-works/fieldwork-app repository on main, Cloud sandbox, the task as context, and Start agent work.](https://volobox.io/docs/images/task-start-coding.light.webp)

*The Start coding work dialog on a task: a note for the agent, the coding agent, model and mode, the repository, base branch and Cloud sandbox, and Start agent work.*

## Coding sessions

A coding session runs in a **sandbox**, an isolated computer with the repository checked out. Each session works on its own **working branch**, separate from the base branch, and has at most one active pull request.

In the session you:

- chat with the coding agent
- follow each step as it works: the commands it runs and the files it edits
- read the diff of every changed file
- use a live terminal, and start the app

Several people can join one session, and its **Team chat** is a discussion for them beside the agent conversation.

> **Note:** A session that runs on a person's own machine is theirs alone. Others cannot join it.

The composer has three modes. **Ask** answers questions about the code without changing files, **Plan** has the agent write a plan you approve first, and **Agent** changes files.

After each Agent turn, Volobox commits and pushes the working branch by default, so the work survives a stopped sandbox.

## From a task to a pull request

On a project with coding turned on, you assign a task to a team agent that writes code and start its work. The agent works under its own identity and access.

It reads the task, its discussion, a note you add when starting, and the repository's own agent instructions. It asks when something is unclear instead of guessing. The task discussion shows the run as milestone cards (a plan to approve, a question, a preview ready, changes ready for review), and you steer the run by replying there.

Before the pull request opens, Volobox runs the repository's **verify** steps (build, test, lint) declared in its preview manifest. A failing check gives the agent a turn to fix it.

> **Note:** If a check still fails after the agent's fix, the pull request opens with the failure marked.

The pull request shows a Verification section with each check's result, and the agent's own **Agent confidence** (high, medium or low).

The guide [From a task to a pull request](/docs/task-to-pull-request) walks through the loop, and [Guide coding agents](/docs/guide-coding-agents) covers the instructions agents read.

## Shipping and review

One **next-step button** in the session names what waits on you: push the changes, open the pull request, or mark it ready for review. Commit messages and pull request text are written from the work, so one click is enough, and you can still write them yourself. A new session opens a **draft pull request** after its first automatic push.

Before pushing, you can review the diff in the session, comment on lines, and send **Request changes** back to the coding agent.

Pull requests from every connected code host can be read, reviewed and merged in Volobox, whoever opened them. For one linked to a session, your requested changes can go back to the agent so it keeps working. Approvals and branch protection on the code host decide what merges.

## Branch previews

A repository describes how to run its app in a **preview manifest**, which a coding agent can write for you. With it, Volobox starts the app from any branch in a sandbox, a **branch preview**, so reviewers try the change before it merges.

The same file lists the **verify** steps. See [Write a preview manifest](/docs/write-preview-manifest) and the [preview manifest reference](/docs/preview-manifest).

## AI review

A reviewer agent reads a pull request and reports findings anchored to lines of the diff, whether each acceptance criterion is met, and a verdict.

> **Note:** AI review is advice. It never approves, never requests changes and never blocks a merge.

Anyone allowed to can select **Request AI review**. Teams that want reviews automatically create workflows from the AI delivery templates (see [Workflows](/docs/workflows)). A finding can go back to the coding agent, become your own review comment, or be dismissed.

![A pull request opened by the Software Developer agent, on its Reviews tab: your review is requested, checks 4 passed, agent confidence High, and the AI review asking to fix one thing first, with acceptance criteria 2 of 3 met and an Important correctness finding about the retry counter.](https://volobox.io/docs/images/pull-request-review.light.webp)

*A pull request on its Reviews tab: an AI review that asks to fix one thing first, the acceptance criteria it checked, and a finding on a file and line.*

## Continue in another repository

When the work needs a change in another repository, you continue it in a new session there. The new session starts from a handoff of the work so far, links back to the one it continues, and ships its own branch and pull request.

## Coding agents

- **OpenCode** runs on the organization's models, and its cost counts toward the organization's AI spend.
- **Cursor** and **Codex** run on a person's own account (a Cursor API key, a ChatGPT sign-in), billed to that person's plan with the vendor.
- **Claude Code** runs only on a person's own machine signed in to their Claude subscription, for their own sessions. It never runs in the cloud sandbox.

A team agent can hold its own Cursor or Codex account. Every coding agent gets the organization's and the project's guidance, and can read Volobox's own documentation inside the sandbox through the read-only [sandbox command](/docs/sandbox-cli).

## Where sandboxes run

Sandboxes run in the cloud by default. A person can add a computer they own: it connects to Volobox outbound, opens no port, and runs only that person's private work. See [Run coding sessions on your own machine](/docs/own-machine).

## Code hosts, issues and deployments

Volobox works with **GitHub** (through a GitHub App your organization owns), **GitLab.com** groups and **Azure Repos**. Repositories are registered once and linked to projects.

GitHub and GitLab issues can be mirrored as tasks or requests, and closing a mirrored item in Volobox closes its issue.

A project can link **Azure DevOps** pipelines and start runs from the project, and a **Deployments** page shows those runs across every project you can open. Azure DevOps work items can be synchronized with a project too.

To set up a code host, see [Connect GitHub](/docs/connect-github), [Connect GitLab](/docs/connect-gitlab) or [Connect Azure DevOps](/docs/connect-azure-devops).

## Related

- [Projects](/docs/projects)
- [Agents](/docs/agents)
- [Workflows](/docs/workflows)
- [Integrations](/docs/integrations)
- [Deployment and models](/docs/deployment-and-models)
- [AI Coding](/product/ai-coding)
