AI review
Have a second agent review the work on a ticket, and send the findings you agree with back to the agent that did it.
An AI review puts a second agent on the work an agent has done for a ticket. The reviewer reads the branch's diff against its base branch, edits nothing, and hands in its findings with a verdict: approved, or changes needed. You decide which findings go back to the agent that did the work.
Reviews follow a ticket. They are available for sessions started from a GitHub or Jira issue, in a worktree whose branch carries the ticket's key. A session started without a ticket has no review.
When work is ready for review#
When reviews start automatically for the repository, Tegment adds a line to the prompt of an agent you start on a ticket, asking it to call request_review on the MCP server once the work is done and committed. That call is how the agent says the work is ready. With automatic reviews off, the prompt doesn't ask for it, and you start reviews yourself.
Tegment treats the work as up for review when the agent has asked for a review, when a reviewer is on it or has handed in, or when the branch has an open pull request that isn't a draft. This is Tegment's own state: nothing moves in GitHub or Jira.
Starting a review#
Start one yourself with Start review, from any of these:
- the ⋮ actions menu of the agent's terminal,
- the right-click menu of the session in the sidebar,
- the Review tab of the session panel.
After the first review the item reads Review again. It is offered while nobody is reviewing the work, or once the last reviewer has handed in.
To have reviews start on their own, turn on Start reviews automatically under Settings → Review. A reviewer then starts when the work becomes ready for review and nobody is reviewing it. Each repository can choose differently on its Workflow settings page, and the Review tab has the same switch for the repository you are in.
The reviewer#
The reviewer runs in the same worktree as the work, with the agent, model and effort set under Settings → Review. Its terminal is named after the work with Review · in front, and by default it starts in the background, under Background in the sidebar, so it doesn't take a tab next to your terminals.
When the reviewer hands in, Tegment closes its terminal a few seconds later, unless you have brought it forward. Stop review withdraws the review and closes the reviewer.
The reviewer hands in through the MCP tool submit_review, and Tegment always tells it to, whatever your prompt says. The review is saved as a Markdown file in the worktree, under .tegment/reviews/, named after the ticket key and the commit it reviewed. Tegment keeps that folder out of git, so it never shows up among your changes. A review of a later commit gets a new file.
Following a review#
While a review runs, a bar over the agent's terminal says AI reviewing, with Go to reviewer and Stop review. The reviewer's own terminal carries a bar too, with Go to agent and Move to Background. If the reviewer stops to ask you something, the bar says Reviewer is asking.
Once the review is in, the bar shows the verdict, Approved or Changes needed, with Open the review, Send to the agent and Put the review aside. The bar goes away when the agent commits again, since the review then describes an older state of the branch, or when you put it aside.
The Review tab in the session panel holds the whole review: the verdict, which automatic round it was, the findings and the review itself. Open large shows it in a window of its own, with Open in editor to open the file. When the agent has committed since the review, the tab says how many commits have landed and offers Review again or Done anyway.
Sending findings to the agent#
The reviewer lists its findings one by one, each marked Should fix or Suggestion with the file and line it is about. In the Review tab every finding starts ticked. Untick the ones you disagree with and choose Send N findings to the agent: Tegment types a message into the agent's terminal naming the review file and the findings to work on.
Send to the agent on the bar sends the whole review instead, with the message for its verdict (see below).
Settings#
Settings → Review has three parts.
Reviewer
| Setting | What it does |
|---|---|
| Prompt | What the reviewer is asked to do. It can use {key}, {title}, {branch} and {baseBranch}. |
| Agent | The agent that runs every review, whether you start it or it starts automatically. |
| Model and effort | For an agent that offers them. Default uses the agent's own setting. |
| Auto-approve | Lets the reviewer run commands without asking. When off, a reviewer in the background can sit waiting on a permission question you don't see. Off by default. |
| Start reviewers in the background | On by default. |
Automation
| Setting | What it does |
|---|---|
| Start reviews automatically | Starts a reviewer when the work is ready for review. Off by default. |
| Automatic rounds | How many rounds start on their own, from 1 to 5. A round is a review and the fixes that follow it. After the last one, you start further reviews yourself, and an approval ends the rounds early. |
| Send the review to the agent automatically | Waits until the agent has finished what it is doing, then sends the review with the message below. Off by default. |
Back to the agent
The two messages sent to the agent that did the work: one for a review that asks for changes, one for an approval. Both can use {reviewFile}, the path to the review. By default, a review that asks for changes tells the agent to fix what it agrees with, commit, and ask for review again, which is what lets automatic rounds carry on. An approval asks the agent to say which notes are worth doing, without changing anything yet.