Code review

A second agent reads what the first one did, with fresh eyes and without its conversation. Its findings go back to the agent that did the work, which fixes them; then the reviewer checks again. Codex reviewing Claude’s work, or Claude reviewing Codex’s, catches mistakes neither would catch alone.

Starting a review

In the pane of the agent whose work is to be reviewed, click +review or press ctrl+q t r (or ctrl+q a c). A review pane opens beside it, listing the agents to review with; the configured reviewer is first. Press enter, its number, or click to start.

• The button beside an agent shows the model it reviews with; click it, or press m, to pick another from the models the agent itself lists and those it has been started with before. Type to narrow, or type one it does not list. The model picked is kept in [review.models] for the reviews to come; its own goes back to the agent’s default.
• - / + change the number of rounds.
• esc closes the pane.

A review is always of one agent’s work, and its findings go back to that agent. Several agents in a tab can each be reviewed.

When the agent says it is ready

With the nerdr-request-review skill, the agent asks you when it has finished whether you want a review, and starts one if you say yes. Name the reviewer you want, “have codex with astra review it”, and it passes that on.

Automatic reviews

Set [review] auto = true to review without asking. By default the review starts when the agent says its work is ready, with what it wants looked at most; when = "turn" reviews after every turn that changes the worktree. See Configuration.

What the reviewer does

The reviewer runs in the same worktree. It reads the branch’s commits and the uncommitted changes, is told to change nothing, and sends a verdict, a summary and findings on specific lines.

While it reviews, tasks queued for the agent wait, so nothing changes under the reviewer. The status panel shows rev: with the round and what the reviewer is doing.

Findings

Findings appear beside the code in F1, marked ⚑, and in F3. When a round ends, nerdr asks in the tab whether to send them to the agent now or keep them in F3; in another tab, it asks when you come back, also if nerdr has been restarted since. Until you answer, the tab is marked ⚑ in the tab row and the sidebar, or ✓ if the reviewer approved. When you step through the tabs with ctrl+q →, a small card shows the result as you pass the tab; press enter to stop there and answer. With send_findings = "auto" they are sent without asking.

After the agent has addressed them, the reviewer checks again, in its own conversation. This goes on until it approves, or the rounds limit is reached (2 by default); remaining findings then wait for you in F3.

Several reviewers

Select + reviewer while a review is underway, or open another review pane from the agent. You can run one reviewer of each agent kind. A round ends when all of them have answered, and their findings are sent together in one numbered message. Reviewers that approve stop and their panes close; the others check the fixes.

The review pane

The reviewer runs in its own pane, under a line on how the review stands. Go to the pane to answer its questions. A review that starts by itself opens beside the agent without taking the keyboard, and is the first to be minimised when room runs out. Stop it with the stop button, ×, ctrl+q t x in its pane, or ctrl+q a x from the agent. The pane closes when the review ends.

Reviews are kept in reviews.jsonl beside nerdr’s notes. The latest findings remain after a restart, but reviewers are not restarted: a review still going on when nerdr quits ends there, without being told of.

Read-only reviewers

Reviewers are told not to change files, but run with the agent’s normal settings. To enforce it, give them arguments of their own, such as [review.args] codex = ["--sandbox", "read-only"]. See Configuration.