AI code review and QA before your team sees the pull request

Review reads. QA runs.

They are different proofs, and mixing them is how an agent gets a pass it did not earn. Review reads the diff and the AGENTS.md guidance that applies to the files. It does not run the code. A finding says what breaks, on the line, and why. P1 should block the merge. P2 should not. The review ends in a confidence score out of 5, a risk rating, and a verdict: approve, or request changes. Where your guidance is silent, it does not invent a rule.

QA does not read a verdict out of the model. Each check runs with bash, exit on error, and a pipefail option. Exit code 0 passes. Anything else fails. The agent saying the tests passed is not evidence. Checks keep going after one fails, so the agent hears every failure in that try. A failure goes back with the command and the last 200 lines of output, secrets scrubbed. You set 1 to 5 tries. 3 is the default. The last failure ends the session without a pull request.

The order inside a task

1. Plan first, if you chose it: a spec you approve, each criterion tied to a check, before any code.
2. The agent, Claude Code or Codex, commits in that task’s sandbox.
3. Review reads the commit. A request for changes sends the work back.
4. QA runs every check on that commit, and, for Plan first, one command per acceptance criterion.
5. The gate’s approval opens a draft pull request with the proof on it.

What the team opens is that draft, plus the spec, the review notes, QA’s results, and the cost of the task. A person merges. Agents cannot push. The same order runs whether the task started in Linear or Jira.

Reviews of pull requests the team opened

There is a second review, on pull requests as they open and change, including ones a person opened. A push starts it. Pushes within a minute of each other become one review of the latest commit. Documentation, renames that change no lines, and version bumps end before a model runs. Changes to agent instructions, such as AGENTS.md, are always reviewed. A skipped review costs nothing. Past the month’s included reviews, a larger pull request costs more than a small one, and reviews can pay from the balance up to a monthly cap, or pause with a comment that says why.

That second review is closer to what Copilot, Devin Review, and Code Reviewer do: comments on a pull request a human is already in. It is useful, and it is not a substitute for the gate inside the task. The product pages have the rest: code review and QA. Recordings, on paid plans, can be attached. They never decide the verdict. The exit code does.

The cost is the task’s

A shared credit pool makes the cost of one pull request a guess. The task cap does not. Organization, project, and agent limits are checked before the run. The tightest one wins. Over it, the task does not start, so the team is not reviewing a change that already overspent. The cost of a task that ran is on its record.

Linear’s agent, Copilot’s coding agent, Devin, and Rovo Dev each review or test in their own way. Read those pages before treating “we review code” as a distinction. Plans include reviews on the free plan.