Reviewer and Auto-review

The built-in Reviewer checks a completed response against the request and the evidence available to it. It is separate from a custom Specialist whose name contains “Reviewer,” and separate from permission approval.

For execution-based output comparison, use Reproducibility. A Reviewer assessment and a reproduced output are separate records.

Session review versus artifact review

A conversation review and the Review tab of an artifact's provenance panel are different records. Inspect the exact artifact version you intend to share. If it says No review for this version, preserve that label even if another response was reviewed. Likewise, partial environment capture and bounded evidence remain partial and bounded after a model expresses confidence.

For input errors, attach an accessible current input or resolve its actual version through the application. A local file's existence does not guarantee that every child/reviewer kernel can read it. See Notebook, Delegation and Troubleshooting.

Inspect the selected artifact version when reopening a historical review. After cancelling a review or correction, read the final state and retained findings before deciding whether to rerun. Cancellation does not create a successful review.

Choose a result you can check

For a first review, complete the inline table check: provide the complete sample-QC table and ask for one arithmetic result per sample. This supplies a precise criterion: twelve unique sample identifiers, twelve rows, and zero-count plus detected genes equal to the gene total for every row.

Before review, open the child result and its Notebook output yourself. Then request review of that response. Compare the review's checks with those criteria; if a row or executed result is missing, resolve that finding before using the result. Review outcomes depend on the response and available evidence; this exercise does not promise a zero-finding badge.

Request a review

1. Complete a conversation turn with a working model.
2. Open the composer + menu → Request review. The menu changes to Reviewing… while review runs.
3. Open the resulting Reviewer card. Read the number of findings and checks, then expand each check's explanation.
4. Select Go to transcript to open Session Reviewer. Check the model, timestamp, PASS/FAIL statements, evidence references and Reviewer log.
5. If corrections are requested, inspect Main Agent's follow-up and any child permission requests. Review does not grant those operations automatically.
6. Use Re-run review after addressing the identified issue. Preserve unresolved findings if a required input or operation is still unavailable.

Worked example Read a review with unresolved findings

View the checks and unresolved findings

Using Codex subscription authentication with gpt-5.6-sol, manual review returned four checks and one finding:

Check	Actual result
The Specialist executed the inline CSV review	PASS; the review cited the child handoff and arithmetic results.
Custom MCP results and failures were reported accurately	PASS; valid metrics and the connector error matched execution output.
The molecule call produced the stated artifact/descriptors	PASS; the returned artifact version and values were identified.
The model inspected the saved molecule preview as requested	FAIL; its catalog lookup did not read the structure content.

The corrections encountered managed-input access problems and reached fix limit reached / Issues found. Opening the structure manually in the application proved that the viewer worked, but did not retroactively change the model's recorded inspection or review result. This is a real unresolved review outcome, not a successful review badge.

Auto-review controls

Open Agent controls → Auto-review to configure review after future responses. This is a conversation preference; it is distinct from Ask for approval and Delegation. The built-in Reviewer row in Settings has no ordinary edit/delete/enable controls and is excluded from the normal Specialist picker.

UI state or control	Meaning
Request review unavailable	Check for an active response/review, a missing eligible completed response, or unavailable model setup.
Reviewing…	Review is still running; do not treat it as complete.
Reviewer · n findings · n checks	Opens the checks and their evidence. A zero-finding result is still bounded by what was checked.
Corrections requested	Main Agent may run a follow-up correction cycle. Inspect new operations and their results.
Issues found / fix limit reached	The review did not resolve every finding. Read the latest explanation before starting a new attempt.
Go to transcript	Opens the dedicated Session Reviewer page.
Expand / Collapse Reviewer log	Reveals or hides its operation log; a truncated log is not complete evidence.
Re-run review	Requests another review; it is not an “accept all findings” button.

Run Auto-review with a separate model

1. Under Settings → Model → Reviewer, choose an available fixed model. The exercised setup used gpt-5.6-sol for Main and gpt-5.6-luna for Reviewer.
2. Open Agent controls → Auto-review in the target conversation, confirm On, then send the next request.
3. After the response, expand the automatically created Reviewer card. Check its model, criterion, evidence and result.
4. When Corrections requested appears, inspect Main's correction and the subsequent review before deciding whether the finding is resolved.

What “resolved” establishes

A check can be resolved because a requested attempt was made and its permission failure was accurately reported. That does not establish that the file was readable or its calculations passed. Read the criterion, tool result and remaining findings together. If access remains blocked, follow the file-handoff known issue before starting another review.

Implementation reference: SessionReviewerPanel.tsx, ComposerAgentControlsMenu.tsx.