Codex Code Review Lands in ChatGPT Desktop With Background Scans

OpenAI gave code review its own surface at DevDay 2026: the ChatGPT desktop app now includes a dedicated review view where developers see changes across multiple projects, read generated summaries, walk through diffs, and question Codex about potential issues before sending feedback to GitHub or GitLab. The feature's sharpest edge is automatic reviews — Codex can scan changes in the background while the developer is offline, so review sessions start with findings already prepared. It is available on all plans.

This piece covers how the review view fits into existing pull-request workflows, what the background-scanning model changes about review economics, and what teams should check before treating agent findings as review signal. It is written for engineers and team leads who own code quality in their organizations.

The review view: what OpenAI shipped

The desktop app now treats review as a first-class mode rather than a chat pattern. The view aggregates changes across multiple projects into one place — relevant for anyone juggling several repositories — and layers three interactions on top of each change set: a summary, the diff itself, and a conversational thread where you can push Codex on specific concerns before anything leaves your machine for GitHub or GitLab.

Making it available on every plan, not as an enterprise upsell, is a deliberate breadth play. Review is the choke point of most teams' delivery pipeline, and OpenAI is positioning Codex inside that choke point for every subscriber tier at once, the same day it expanded Codex to phones and the cloud and open-sourced its harness.

Background reviews: the actual innovation

Interactive AI review has existed in various tools for two years; the shift here is temporal. Automatic reviews run while the developer is away, which inverts the usual sequence — instead of a human opening a diff cold and summoning help, the agent has already read everything, and the human session starts at the triage stage. Overnight, the queue of unreviewed changes becomes a queue of pre-analyzed changes.

The output boundary stays conservative: findings inform feedback that a person sends to GitHub or GitLab. Codex is not auto-approving or auto-commenting on pull requests; the accountable reviewer remains human and the platforms of record remain the existing ones, which makes the feature adoptable without changing team policy.

Why review is the right place for an agent

Review work splits into two unequal parts: comprehension — reading the change, tracing its implications — and judgment — deciding whether it should ship. Comprehension consumes most of the clock time and is exactly what a model with the full diff and repository context accelerates; judgment is where human reviewers earn their role. A pre-briefed review that compresses comprehension can shrink time-to-first-comment, the metric most correlated with how long pull requests sit open.

The risk is the familiar one from every generation of analysis tooling: signal quality. Agent findings that are subtly wrong or confidently irrelevant cost reviewer trust quickly, and a reviewer who starts rubber-stamping agent summaries instead of reading diffs has made review worse, not better. The conversational layer — being able to interrogate Codex about why something is flagged — is the mitigation built into the design.

Who benefits first

Multi-repository maintainers get the clearest immediate value: the aggregated view plus background scanning effectively staffs a first-pass reviewer across every project simultaneously. Solo developers and small teams gain a second pair of eyes they did not have; larger teams gain review coverage on the long tail of changes that today get waved through because senior reviewer time is scarce.

Teams with strict review SLAs should note the desktop-app dependency: this lives in ChatGPT's desktop client, not in the GitHub or GitLab UI, so it augments rather than replaces the platform-native review flow. Where it fits best is the stage before formal review — the author's own pre-flight, or a designated reviewer's preparation.

Host support and PR integration: what to watch

Unpublished at launch: which repository hosts beyond GitHub and GitLab are supported, how background scans are scheduled and scoped, and whether findings can flow directly into PR comments rather than through a human relay. Also worth watching is how this interacts with Codex Security Cloud, announced the same day — correctness review and security scanning are converging on the same always-on, background-agent pattern.

A sensible pilot: pick one active repository, let background reviews run for two weeks, and measure finding precision against what human reviewers independently caught. If the agent's findings overlap heavily with human ones, it is compressing comprehension time; if they diverge wildly in either direction, calibrate before rolling it into the team's review path.