How AI CodeOps works

A session takes one task from a sentence to a reviewed pull request. This page follows a session through each stage, in the order AI CodeOps runs them.

1. You describe the task

Open Develop → AI Sessions and choose New session. Describe the task, pick a Workflow and a Project, and choose Start session. Every session belongs to a project.

Before anything runs, AI CodeOps checks two things:

• Your budget. If the workspace is over its budget or out of allowance, the session doesn't start and says why. See Cost controls.
• Your workspace policy. A deny rule stops the session. An ask rule holds it until someone approves it on Review & Gate → Approvals. See Approvals and policies.

The session is then queued with the status pending until a worker picks it up.

2. The agents work through the workflow

A workflow is a fixed list of steps. Each step is run by one agent, and the steps run one after another. The default workflow, Plan, Build, Review & Test, runs:

1. Plan, by the Architect.
2. Build, by the Builder.
3. Review, by the Reviewer.
4. Test, by the Tester.

Each step sees the task and what the earlier steps wrote. The budget is checked again before every model call. A session that goes over budget mid-way stops with the reason. Other workflows are listed in Workflows.

3. The change is written

Once the steps finish, AI CodeOps turns the team's plan into complete files. It applies them to a fresh copy of your repository and produces a diff. A project without a repository gets a new codebase instead.

A diff over 60 KB is refused whole: "Split the task into smaller changes." If no usable file changes come back, the session fails with the reason.

4. Your tests run

AI CodeOps rebuilds the change in an isolated sandbox. It installs dependencies, cuts the sandbox off from the network, and runs your project's test command. The result is recorded as the session's test run: passed, failed, skipped and coverage.

The test command is the one set on the project. Without one, AI CodeOps looks for a package.json test script, then pytest, then go test. If tests can't run at all (no command, or dependencies that won't install), nothing is recorded and the run page shows that the change wasn't verified.

5. The diff is reviewed

A review pass reads the actual diff, with the review and security steps' notes as context. It records findings, each with a severity, a category and usually a file. See Multi-agent reviews.

If that review can't finish, the change is held: its findings are unknown, so it can't be treated as clean.

6. The change waits for a person

When the session finishes with a diff, AI CodeOps:

• marks the session completed;
• moves its change to in review under Develop → Changes;
• opens a review request on Review & Gate → Approvals. It expires after 7 days.

Open the session page to see the summary, the Proposed changes, the Quality gates panel and the Findings.

7. A person approves

Choose Accept changes on the session page, or Approve on the Approvals page. The person who started the session can't approve it while someone else in the workspace can. A one-person workspace can approve its own change. One rejection ends the request.

8. The pull request opens

Approving starts the pull request. Before opening it, AI CodeOps checks, in order:

1. the workspace policy;
2. the approval;
3. the quality gate: no open high or critical findings, no failing tests, and a completed review.

If all three pass, AI CodeOps pushes a branch named codeops/run-<id> and opens a pull request into the project's default branch. It uses the Git connection of the person who approved, so they need a connected Git account for that host. If the pull request couldn't open, choose Create Pull Request on the session page. The button says what's missing.

9. Merge and deploy

Merge the pull request when you're ready. If you merge it from the console, the quality gate is checked again first. If the project has a deploy target, AI CodeOps then calls your deploy webhook and records the deployment. See Gated deploys.

Time limits

A session that times out or stalls ends as failed, with the reason on its page.

Next steps

• Core concepts
• Sessions
• Quality gates