Tuning Deep review
⚡ Deep review looks for problems in one layer of a review. Up to three Reviewers each read the whole change, in a read-only checkout of its head, and report the problems they find. One Refuter then merges reports that describe the same problem and tries to disprove each one, citing a file and line. Each finding becomes a thread on its line.
Deep review never runs on its own. Select ⚡ Deep review on a layer in the Review Room when you want it. See Reviewing a PR stack. In the round view of a PR you reviewed before, Other changes has its own ⚡ Deep review · round changes. It reviews only those changes.
Run it
With more than one Reviewer, ⚡ Deep review is a split button:
- The main part runs every Reviewer you set up.
- The menu runs fewer: only the first one, or the first two. Each entry names the models it uses, for example 2 · Opus ×2.
While it runs, there is a pill for each Reviewer and one for the Refuter, with its model and progress. Select Stop to end the run. The pills stay after the run: each Reviewer says how many reports it found, and the Refuter says how many findings it left open and how many it disproved. If an agent fails, its error is shown under the pills.
Read the findings
Each finding says how many Reviewers found it, for example found by 2 of 3. This is a label, not a filter: a problem only one Reviewer found still shows.
When the Refuter disproves a finding, the finding stays open but starts folded, with the Refuter’s proof. Read the proof and decide. Only Dismiss removes a finding.
Pick the Reviewers
- Open Settings → Deep Review.
- Pick a model for each Reviewer. Select + Add Reviewer for a second or third one. It starts as a copy of the last.
- Put them in order of preference. Drag a row by its handle, or focus the handle and press Alt+↑ or Alt+↓. A run with fewer Reviewers uses the top ones.
- Select ✕ to remove a Reviewer. One always stays.
The same model can run more than once. You can also mix models, for example Claude next to Codex or OpenCode. Codex and OpenCode Reviewers read the checkout too: SimpleEdit locks them to read-only tools and switches off your MCP servers, plugins and hooks for the run.
A local model through Ollama can’t read the checkout, so it reviews the diff only. Its row says diff only, and so does its pill during a run.
Pick the Refuter and the time limit
- Refuter picks the model that checks the reports. By default it uses Reviewer 1’s model, or a built-in Claude model when Reviewer 1 is diff only. Ollama models aren’t offered here.
- Time limit per agent stops a Reviewer or the Refuter still running after that many minutes. Leave it empty for 10 minutes. A stopped agent says Timed out after N min.
Rewrite the instructions
- Open Settings → Prompts.
- Find Reviewer or Refuter and select Customize.
- Edit the instructions and select Save.
You change only the instructions. SimpleEdit still adds the output format and the change itself, so an override can’t break a run. There are no placeholders to fill in. If your file is empty or can’t be read, SimpleEdit uses the default instructions for that run.
To go back, select Reset. When SimpleEdit ships new default instructions, a customized prompt is marked Outdated. Select Compare to see the difference.
If you customized the instructions of the earlier review lenses, SimpleEdit keeps those files. They show read only as Earlier lens instructions beside the Reviewer prompt, so you can carry over what still applies.
Give it your team’s rubric
Put a REVIEW.md in the repository: at its root, or in a folder the change touches. Deep review, the overview and the Companion read it as a rubric. Write what your team cares about in a review, for example “Every public function needs a test.” SimpleEdit never creates or edits this file.