Skip to main content
A passing run has shown that its change builds and that your repository’s own checks pass. It has not shown that the change does what you asked, that it has no bug your tests miss, or that it changes nothing you did not ask for. Review this asks a second agent, the reviewer, to judge those three things once the checks have passed. The agent that wrote the change is not the one that judges it.

​
Asking for a review

Review this appears in the header of a session once every check your repository declares has passed on the change as it stands. It is offered on sessions started from a typed prompt, to any member of your organisation.
The header of a session titled Add a page size to the items list, with the status Idle. Three buttons sit on the right: Review this, in the solid colour, then Stop session and View PR.

The header of a session whose checks have passed, offering Review this.

Press it and the conversation gets a muted line, Review requested, while the reviewer starts. The draft pull request opens when the checks pass, as Opening a pull request describes. Review does not hold it back. When a review leads the agent to change the code, the next passing change updates the same draft.

​
When Review this is not there

The button is missing when any of these is true:
  • A declared check is failing or has not run, or the agent has changed the code since the checks last passed. It comes back once the checks pass on the change as it stands.
  • The agent is working, or the session has been stopped or has failed.
  • A round has already judged this change. A Clean round makes the button go away, because there is nothing left to ask. It comes back only when the agent changes the code and the checks pass again.
  • The session is Waiting on you for Review findings. Send it a message instead; When the run waits on you says what happens next.
  • The session was not started from a typed prompt.
  • Your repository declares no checks. The session says No checks are declared in scaling.yml, and Checks says how to declare them.
  • The run has used all of its attempts. The header says Review is not available: this run has used all of its attempts.

​
What the reviewer judges

The reviewer judges three things, and only these:
  • Whether the change does what you asked.
  • Whether it has a bug, including one your tests do not cover.
  • Whether it changes something nobody asked for.
Security review is not part of this version. The reviewer works from your request and the change, and it reads your repository’s own AGENTS.md, CLAUDE.md and skills as the agent does. A change that breaks one of your rules is a finding that quotes the rule. Guidance the agent edited during the run is not what the reviewer follows. It never sees the agent’s conversation, its reasoning or its summary of its own work. It reviews the change in a new environment started from the same code and set up the same way, so nothing the agent left running, wrote to a database or created outside the change is there. Its environment is as closed to the network as the agent’s, and anything it changes is thrown away. The reviewer cannot push, and only the agent’s change can reach a pull request. The reviewer runs on your organisation’s own Anthropic key, the one in Adding your Anthropic key. If the key fails during a review, the session waits on you for the key, exactly as it does when the key fails for the agent, and the review is not counted as clean.

​
What a finding is

Every finding has exactly one severity, a one-line title, what is wrong, where (a file and line, or the behaviour), and at least one piece of evidence. Evidence is the lines of the change the finding concerns, or the output of a command the reviewer ran. A finding without evidence is thrown away and never shown. The agent receives the title, the place and the evidence of each blocking and should-fix finding, and nothing of how the reviewer reached it. A review is clean when no blocking finding stands.

​
Rounds and attempts

One pass of the reviewer is one round. Each round adds a line to the conversation: the round number, the attempts left, and how many findings it raised of each severity in lower case, such as 1 blocking, 0 should-fix, 2 notes. The line ends with how the round came out, and under it each finding’s title sits beside a badge with its severity: Blocking, Should-fix or Note.
One line of the conversation reading Review round 1, 2 attempts left, 1 blocking, 0 should-fix, 2 notes, followed by the muted words Sent back to the agent. Below it are three findings, each with a badge: a red Blocking badge beside The route ignores the page size it was asked to add, and two outlined Note badges beside shorter titles.

A round that raised one blocking finding and two notes.

A round ends in one of four ways, and the line says which:
  • Clean: no blocking finding stands. This is the normal end of a review, and there is nothing you need to do. The draft pull request opened when the checks passed, so it is already there for you to review as for any other session. If files in it waited for a member to confirm them, Files that wait for a member says how. Should-fix findings and notes stay recorded in the conversation.
  • Sent back to the agent: a blocking or should-fix finding went back to the agent, which works on it.
  • Out of attempts: the round found a blocking problem and the run has no attempt left to send it back.
  • Stuck: the same blocking finding stands and its files did not change: the reviewer raised the same blocking finding as in the round before, and the agent changed none of the files it names.
After a round that sends the change back, the agent works on the findings, the checks run again, and when they pass the reviewer reviews the new change in another fresh environment. It is told which of its earlier findings to look at, and says for each whether it is resolved or still stands. A run has the number of attempts your project’s Attempts setting gives it, 3 unless an admin changed it. A check that fails uses one, missing evidence uses one, and a review that sends the change back uses one. Rounds have no limit of their own: a review that finds nothing to send back costs no attempt. A review does not start when no attempt is left. Attempts and review says how to change the number and which model reviews. The last round is the one after which no attempt is left. On it, should-fix findings are recorded in the session and do not send the change back, so only a blocking finding can keep the run from being clean.

​
When the run waits on you

A session moves to Waiting on you with the reason Review findings in two cases, and the detail after the reason says which, as in Review findings: out of attempts:
  • out of attempts: the last round still found a blocking problem.
  • stuck: the reviewer raised the same blocking finding in two rounds in a row, and the agent changed none of the files that finding names in between. The session stops before its attempts run out, because another round would find the same thing.
The session lists the blocking findings that still stand under its title, each with where it is, and counts in the number beside Runs like any session that waits on you. There is no separate status for it.
The header of a session titled Add a page size to the items list, with the status Waiting on you and the line Review findings: out of attempts under it. Below is a heading, Blocking findings that stand, with one finding, The route ignores the page size it was asked to add, and the file and line it points to.

A session waiting on you because the last round still found a blocking problem.

You have two ways forward:
  • Send the session a message saying how it should go on. It goes to the agent. When the checks pass on the new change, the reviewer reviews once more, whatever is left of the attempts.
  • Stop the session, as for any other.

​
When a review cannot run

When the reviewer cannot run, the conversation says so with one of these lines, and the review is not counted as clean:
  • The review could not run: the change could not be reproduced in a fresh sandbox. Press Review this to try again.
  • The review could not run: the snapshot this session started from is no longer available. Press Review this to try again.
  • The review could not run: a fresh sandbox could not be started. Press Review this to try again.
  • The review could not run to an end: the reviewer recorded no result. Press Review this to try again.
When your Anthropic key cannot use the review model chosen for the project, the round does not run on another model. The session moves to Waiting on you with the reason Review model unavailable and the detail the review model is not available on your Anthropic key, followed by the model’s name as the picker shows it. Choose another model under Attempts and review, then press Review this. The session stays where it was, and Review this is offered again. If pressing it fails before the reviewer starts, the header says The review could not be started. Try again. When the run has used all of its attempts, the header says Review is not available: this run has used all of its attempts. instead.