← Articles

How to Know If an AI Pull Request Is Safe to Merge

24 August 2026 · Wednesday

Merge when every check passes and every comment has an answer. Then check that the tests prove what a user would see, and that the pull request does one thing. That's the bar a senior engineer holds, and you can check it without reading every line. The free AI starter kit we build our own products with takes a pull request to that bar and keeps it there.

What has to pass

Run every gate your repo has and read the results, not just the green tick: formatting, lint, type check, the tests for this change, coverage against your bar, and a production build whenever dependencies, config or imports changed. A change that works in development can still fail the production build, so run it.

Answer every review comment

Collect every signal on the pull request: review bots, coverage reports, code quality tools and people. Then give each comment one of three answers:

  • Accept and fix it in this pull request. Most CI findings land here.
  • Reject it with a reason. A bot's false alarm still gets a polite, specific reply so the thread closes.
  • Defer it as a gap. If it's real but outside this change, log it as an issue, link it, and say so in the thread.

Push the fixes, let the checks run again, and repeat until nothing blocks.

Ask what a senior engineer would ask

  • Does each piece of logic have one home? Logic belongs in the part of the code that owns it, not in a screen or a button handler.
  • Is any rule written twice? If the server and the app both work out the same thing, they'll disagree one day.
  • Do the tests check the result? A test that only checks a function was called passes even when the app is broken.
  • Does it do one thing? One concern per pull request keeps it easy to check and easy to roll back.

Know what's proven and what isn't

Ask for the status in three words: code, wired or verified. Code means it's written, wired means it's connected to the app, and verified means someone saw it work. Anything not verified gets named in the pull request, so you merge knowing exactly what's left.

Red flags before you merge

What you see What it means What to do
A comment with no reply An open question going into main Answer, fix or log it
Tests that only check a call happened Tests that pass when the app is broken Test the end result
A loosened test or a longer timeout A real failure hidden Fix the cause
--no-verify in the history Hooks skipped Rerun the hooks and fix what fails
Five unrelated changes in one PR Hard to check, hard to undo Split it

Get the review set we use

The starter kit's review set answers every comment, fixes behind your checks and gives you a pass or fail with evidence before you merge. It works with Claude Code and Codex. Get the starter kit free on GitHub.

Questions

Is a green CI run enough to merge an AI pull request?

No. CI only proves what your checks cover. Also make sure every review comment has an answer and the tests check what a user would see, not just that a function was called.

What do I do with review bot comments I disagree with?

Reply with the specific reason and resolve the thread. Every comment gets an answer, so nothing is left open when you merge.

Should one pull request fix several things?

Keep it to one concern. Log anything else you find as a separate issue and link it, so the pull request stays easy to check and easy to roll back.

Let's grow your business

Tell us about your business, and we'll call you if it's a fit.

How do you want to work with us?
Stage of the business