Last updated · Santosh Kumar Radha

An AI agent that reviews pull requests and fixes the comments

CodeAF's review specialist reads a GitHub pull request and reports findings; senior-dev can then fix them. What it checks, what it posts and its limits.

CodeAF, the open-source (Apache 2.0) coding agent by AgentField, reviews a GitHub pull request with its built-in /review specialist and can then fix the findings with senior-dev, its coding agent, on your yes. Review reads code and reports; it runs no tests and posts nothing until you ask, and the fix is a separate task on its own branch.

At a glance

CodeAF /review
Start it/review owner/repo#123 focus on the retry logic, or codeaf review run owner/repo#123 at a shell
What it readsthe pull request's commit on GitHub, checked out in a temporary folder; not uncommitted or unpushed edits
How it reviewsmaps the change, plans, sends up to eight reviewers at parts of it at once, then filters, checks evidence and challenges each finding
Runs testsno; reviewers have read-only file tools
What you getfindings with file and line, severity, blocking status and a suggested fix; review-report.md and review-report.json
Posting to GitHuba separate step you ask for, after you have read the report
Default ceiling$5 and 2 hours; at 80% of the dollar limit it starts no new reviewers
HostsGitHub pull requests only

Step 1: run the review

Push your branch and open the pull request first, because review reads the pull request as GitHub has it. Then, in a conversation in the repository:

text
/review #123 focus on cancellation and error handling

/review alone finds the current branch's open pull request. A public repository needs nothing else; for a private one it uses GH_TOKEN, GITHUB_TOKEN or the token from gh auth login, and only for the one clone (review manual). A review of a mid-sized pull request commonly takes half an hour to an hour. From a script:

sh
codeaf review run --max-cost 5 --max-hours 2 owner/repo#123 focus on retries

Step 2: read the report before anything is posted

The chat gets a short account whose first line counts the findings, such as 5 findings, 2 blocking the merge (1 critical, 1 important, 3 suggestion)., then one line per finding, and it ends Nothing was posted to GitHub. The full findings and evidence are in review-report.md; the proposed GitHub review and the commit reviewed are in review-report.json.

Check coverage. A report whose first line says Partial review covers only part of the change, because a session failed, a dimension degraded or a limit stopped work. A partial review cannot be posted as an approval. No findings does not guarantee the change has no bugs.

Step 3: post it, if you want it on GitHub

Ask the chat to post the saved report, or run /review post <report-folder>. It posts the review as it stands, against the commit it reviewed. If the pull request has moved on, run the review again for the newer commit.

Step 4: fix the findings with senior-dev

If findings block the merge and the conversation's folder is the pull request's checkout, the chat can offer a senior-dev task to fix the findings you choose. It waits for your yes. senior-dev works on its own branch, runs your build and tests on what it hands in, and stops at $10 and 3 hours by default. Neither the review nor the fix merges the pull request.

Comments from human reviewers work the same way: the conversation can read the pull request with ordinary gh commands, and you hand the comments you agree with to senior-dev as its brief. The CodeAF docs describe no command that turns a pull request's review comments into fixes automatically. Pushing the fix is yours or the conversation's to do, never the task's; see an AI agent that fixes GitHub issues.

Two playbooks for a review gate

  • Review every change. A manager's @dev member hands each change to senior-dev; an @review member reads the branch's diff against main and runs the tests; the manager merges only on a pass. In the documented run the second change took 2 minutes and 45 seconds from message to merge, and the team spent $0.87 across both changes. This @review is a normal team member that runs tests, not the GitHub /review specialist.
  • Gate every merge with a review panel. Three reviewers read the same diff at once, one each for security, tests and API shape, and the manager merges only on three passes. The documented change went from message to merge in under three minutes for about $0.50.

For who may approve what along the way, see audit and approvals. To run reviews from CI, see a coding agent in CI.

About the benchmark

/review is based on PR-AF, which publishes its own Martian Code-Review-Bench results. Those describe PR-AF's published runs; they do not establish the same score for CodeAF's imported specialist, your models or a particular review.

When something else fits better

  • GitHub Copilot code review lives in the pull request page. It "reviews your pull requests, identifies issues, and suggests fixes you can apply in a couple of clicks", can run automatically when an owner turns that on, and can pass suggestions to the Copilot coding agent, which opens a pull request with the fixes (public preview). With Copilot approvals enabled, its approving review can satisfy a required-approval rule.
  • CodeRabbit reviews each pull request on GitHub, GitLab, Azure DevOps and Bitbucket, and its coding agent can turn a review finding into a commit or a stacked pull request for you to approve. If you are not on GitHub, it covers what CodeAF's /review does not.
  • Claude Code offers a review workflow in GitHub Actions that posts inline comments with --comment, and a separate Code Review product for automatic review on every pull request.

CodeAF's difference is that the review runs on your machine with the models you choose, and nothing reaches GitHub until you read it.

Questions people ask

Is there an AI agent that reviews pull requests and fixes the comments itself?

CodeAF does both in two steps. /review reads a GitHub pull request and reports findings with file, line and a suggested fix; the chat can then offer a senior-dev task to fix the findings you pick, on its own branch, and it waits for your yes. Neither step merges the pull request.

Does CodeAF's review run my tests?

No. CodeAF's /review reviewers have read-only file tools and do not run your tests, build or app. Use your CI alongside it, or the review-every-change playbook, where a team member reads the branch and runs the tests before anything merges.

Will CodeAF post comments on my pull request without asking?

No. The analysis posts nothing and ends Nothing was posted to GitHub. You ask the chat to post the saved report, or run /review post <report-folder>, after reading it. A partial review cannot be posted as an approval.

Does CodeAF review GitLab or Bitbucket merge requests?

Not today. CodeAF's /review supports GitHub pull requests only; a local branch without a pull request and other hosting services are outside its scope. For a local branch, a CodeAF team member can still read the diff and run tests, as in the review-every-change playbook.

How much does an AI pull request review cost with CodeAF?

CodeAF is free; you pay your model provider. A /review run defaults to a ceiling of $5 and two hours, and at 80% of the dollar limit it stops starting new reviewers and finishes with what it has. --max-cost and --max-hours set your own limits.

Sources

Last updated 2026-11-21. Checked against the CodeAF v0.8 docs and manual and the GitHub Copilot, CodeRabbit and Claude Code documentation on that date.