Claude Code and Codex on GitHub Pull Requests: Review, Fix Checks, and Address Comments (2026)
Use Claude Code and Codex side by side on GitHub pull requests: one writes the change, the other reviews it, and either one fixes failing checks or review comments. Paradev, GitHub Actions, Copilot, and Codex cloud compared.
For macOS on Apple Silicon and Intel. Linux AppImage is experimental.
Paradev was called Superset until October 2026.
To use Claude Code and Codex together on GitHub pull requests, check out each pull request in its own workspace, run one agent to write or fix the change and the other to review it, and read both results next to the PR's checks before you merge. In Paradev, press ⌘N, choose Pull request, and pick the PR. Paradev checks out its branch in a new workspace where you can start Claude Code, Codex, or both. The Pull requests view shows the same PR's description, checks, review threads, and diff, and you can select lines in the diff and send them to an agent as a comment.
For work that should start without you, an Automation (Pro) runs an agent when GitHub reports an event: a review request, a failed check, or a new comment. The alternatives are Anthropic's Claude Code action or OpenAI's Codex action in GitHub Actions, GitHub Copilot's cloud agent and code review, and Codex code review. They run on GitHub runners or in the vendor's cloud and post back to the PR. Paradev runs the agent CLIs on your own machines or your own Linux server, with your local environment, or in cloud workspaces.
This guide covers a small team that already works in GitHub pull requests. It is published by Paradev. Details about other tools come from their own docs as of October 2026.
Why run two different agents on one pull request?
Claude Code and Codex use different models and make different mistakes. Having the second one review the first one's diff is a cheap way to catch what the author missed. Teams use three patterns:
- Write and review. Claude Code writes the change. Codex reviews it with
/review, or the other way around. - Race. Both agents get the same task in separate workspaces, and you keep the better diff. See how to compare coding agents side by side.
- Divide the follow-up. One agent fixes failing checks while the other answers review comments, each on its own copy of the branch.
The third pattern needs care. Two agents that push to the same PR branch at the same time can overwrite each other's commits. Let one agent push at a time, or have the second one work on a branch of its own and open a follow-up PR.
How do I run Claude Code and Codex on the same pull request in Paradev?
-
Connect GitHub for the project. The Pull requests view then lists PRs from your projects' repositories, with a compact check summary (passed out of total) on each row.
-
Select the PR and click Start Workspace, or press ⌘N and choose Pull request. From the CLI:
-
In the same workspace, start a second agent session with Codex and type
/reviewwhen Claude Code is done. Both agents work in the same worktree, so run them one after the other. -
Read the result in the Changes pane, then go back to the PR's Summary tab. The Checks list shows each check's status with a link to its run, and the header says "All 12 passed" or how many failed. Cancelled checks count as failures.
To hand a specific line to an agent, open the PR's Code tab, select the lines, and send a comment. The agent receives the file, the line range, and the side of the diff. If a workspace is already linked to the PR, the comment goes to one of its running agents or a new session. If not, Paradev creates a PR-checkout workspace and starts the agent with your comment as the first prompt. You can also post the comment on GitHub at the same time.
On the iPhone and iPad app, an open PR shows Fix Checks with Agent when checks fail or the branch has conflicts, and Address Comments with Agent when reviewers left comments.
How do I make an agent respond to pull request events automatically?
Create an automation with a GitHub trigger. Each one reads as a sentence you fill in. Useful ones for this job:
| Trigger | What it starts |
|---|---|
| Review requested from Me | A first review pass before you read the PR yourself |
| Workflow failed | An agent that reads the failing logs and proposes a fix |
| Changes requested | An agent that works through the requested changes |
| Pull request opened | A review of every new PR in the repository |
Use Workflow failed rather than Checks completed for fixes: Checks completed also fires when every check passes.
For pull request events such as Review requested, Pull request opened, and Changes requested, when the project is linked to that repository, the run's workspace checks out the PR's branch and reuses the PR's workspace if one exists. Runs started by comments, checks, or workflow runs can start on a fresh branch, so tell those agents to run gh pr checkout <number> first. A prompt for the review case:
Pick Claude Code for one automation and Codex for another, and both review the same PR. Things to know:
- Forks are only partly skipped. Pull request events from forks (opened, pushed, reviews, review comments, threads) do not start runs, and a run never checks out a fork's branch. A top-level comment on a fork's pull request can still fire Comment added, and checks and workflow runs from fork pull requests still fire. Filter comment triggers by people you trust.
- Watch for loops. An agent's own comment can match a Comment added trigger. Filter that trigger by people.
- The device must be online. If the target machine is asleep, the run is recorded as missed and is not retried. Use Run now later.
The review-requested recipe and the PR feedback recipe have full prompts.
What do other tools do for the same job?
| Option | Where it runs | Agents | Strength | Tradeoff |
|---|---|---|---|---|
| Paradev | Your machines, a shared host, your own Linux server (BYOC), or cloud workspaces | Claude Code, Codex, and other CLI agents | Local environment; review the diff before anything is pushed; both agents in one app | Desktop app on macOS (Linux AppImage experimental); a device must be online for automations |
| Claude Code action or Codex action in GitHub Actions | Your GitHub runners | Claude Code or Codex | Runs from a workflow, for example a review when a PR opens, with no machine to keep on | On GitHub-hosted runners, each run starts fresh; secrets and services must be set up in CI |
| GitHub Copilot cloud agent and code review | GitHub (Actions-based or cloud environment) | Copilot | Built into GitHub: start an agent session from an issue or a failing Actions run, and request a Copilot code review on a PR | Copilot only |
| Codex code review | OpenAI's cloud | Codex | Comment @codex review for a review on the PR, or @codex with other instructions to start a cloud chat that can push a fix; optional automatic reviews | Codex only; the environment is OpenAI's |
| CodeRabbit and similar review bots | Vendor service | Their own reviewer | Automated reviews on pull requests, set up once for the repository | Their own reviewer, not Claude Code or Codex |
If your checks need services that only run on a developer machine or a shared box, such as a seeded database, a VPN, or a large test fixture, Paradev's local runs are the better fit. If every check already runs in CI and you want reviews on every PR with no machine to maintain, a GitHub Action or a review bot is simpler. Many teams use a review bot on every PR and keep agent fix-ups in Paradev, where a person looks at the diff before it is pushed.
Frequently Asked Questions
Can Claude Code fix failing CI checks on a pull request?
Yes. Check out the PR in a workspace and ask Claude Code to read the failing logs with gh run view --log-failed and fix them. In Paradev, a Workflow failed trigger can start this for you.
Can Codex review a pull request that Claude Code wrote?
Yes. Run Codex in the same workspace after Claude Code finishes and type /review, or set a Codex automation on Review requested or Pull request opened. On GitHub, a comment with @codex review asks Codex cloud for a review.
Will the agent merge the pull request?
Only if your prompt and permissions allow it. The review prompt in this guide tells the agent to comment and not to approve, push, or merge. The CI-fix example tells the agent to push its fix, so remove "and push" if you want to read the diff first. Branch protection rules on GitHub still apply.
Does the agent post comments as me?
When it uses gh in a Paradev workspace, yes. gh is signed in with your GitHub account, so the review shows up as yours.
Which Paradev plan do I need?
Paradev is open source under the MIT License. Automations, remote hosts, the iPhone and iPad app, and the Linear and Slack integrations are on Pro. Bring Your Own Cloud needs a paid plan, and Team and Ultra include cloud workspaces. See pricing for what each plan includes.