From a Linear or Jira ticket to a pull request

Where the ticket comes from

Work can start in the tools the team already uses. However it starts, the change comes back as a pull request on GitHub.

• Linear. Start a session from an issue, or hand the issue to the AOSplane agent inside Linear. It answers in the issue, and a reply there keeps the session going.
• Jira. Connect one Jira Cloud site. Browse it with JQL, create issues, add comments, and start a session from an issue.
• The console. Write a task, or keep a backlog. The Needs you tab lists what agents are waiting on you for.
• The API. A project API key lets a CI pipeline or a script start and follow a session, in one project and one environment.
• Slack, rolling out. Mention @AOSplane or use /aos. A channel can have its own project, agent and rules. Slack may not be available to every organization yet.

Link repositories to a project with the GitHub App. The run stays tied to the issue it came from.

Plan first

Choose Plan first when you start the session, and describe the problem in your own words. The repository needs QA checks. Without them, the console shows Plan first switched off and links you to Repository setup.

1. The product manager searches the project’s GitHub repositories, reads the issues and, on Starter and up, your knowledge base. Every tool it has only reads.
2. It writes a spec: the problem, the goal and acceptance criteria, each with a check that proves it. Spec is the default level. Brief is shorter. PRD and Discovery go deeper, on Starter and Team.
3. You approve it in the console, or request changes. It can revise up to five times. A plan nobody decides on expires after 24 hours and the run ends.
4. Only then does the session’s agent, Claude Code or Codex, write the change. The approved spec is its brief. The work happens in that task’s private sandbox.

Loop until checks pass is the other mode that ends in the same kind of pull request. There, your agent writes the plan. Start now runs the agent on its own and does not wait for that loop. The product manager page has the levels and the monthly allowances.

Your agent, on your account

You choose the harness and the model. Claude Code uses an Anthropic account and reads CLAUDE.md from the repository. Codex uses an OpenAI account and reads AGENTS.md. Both can also use the other provider accounts the console connects: Amazon Bedrock, Claude Platform on AWS, Azure OpenAI, a custom provider, or a Claude or ChatGPT subscription where sign-in is available. The provider bills that account, with no markup.

Before the run starts, AOSplane reserves budget against the organization, the project and the agent. The tightest limit wins. If the reservation would exceed what is left, the run does not start. Runs already going are not cancelled by that check.

The sandbox is isolated from your other tasks and from other customers. A credential is issued for the run rather than stored in the sandbox. If the agent wants to do something sensitive, it stops and the request waits in the console until you approve or deny it.

Guides for Claude Code in the cloud and Codex, and a comparison of the two.

Review and QA

After the agent commits, a review reads the change. If it asks for changes, the work goes back. QA then runs the repository’s checks on that commit, and, for a Plan first task, one command per acceptance criterion. Exit code 0 passes. An agent saying the tests passed is not proof.

A failing check goes back to the agent with the command and the end of its output. You pick from 1 to 5 tries; 3 is the default. If the last try still fails, the session stops without a pull request. The QA and code review pages describe the checks and the review.

A person merges

When the checks pass, the gate’s approval opens a draft pull request with QA’s result in it. Agents cannot push to your repositories. Nothing in that pull request lands on the branch until a person merges it. You can review, ask for an AI review, and merge from AOSplane, or review it on GitHub.

The task keeps a record: what it was asked, the budget reserved, each approval, and the pull request. History is kept for 7, 30, 90 or 365 days depending on the plan.

What you need to start

1. An AI account the agent can call, and a GitHub repository linked to a project.
2. At least one QA check on that repository, if you want Plan first or Loop until checks pass.
3. The Free plan, which is self-serve: 2 projects, 2 tasks at a time per project, small machines, 7 days of history and a $5 monthly usage credit. Full pricing.

The same loop is on the home page. Linear’s own coding agent and Copilot in Linear also open a pull request from an issue. This page is AOSplane’s gate: scope, review, and QA before your team sees it, with a hard spend cap on the task. The tracker-specific guides are Linear and Jira.