cd /news/ai-tools/ai-wrote-most-of-my-merged-prs-last-… · home › topics › ai-tools › article
[ARTICLE · art-173434] src=dev.to ↗ pub= topic=ai-tools verified=true sentiment=↑ positive

AI wrote most of my merged PRs. Last night I typed four lines myself.

A developer known as connabyss landed three AI-assisted pull requests in external open-source projects this week — docling-pipelines #686, django-modeltranslation #831 and #832 — while also shipping Bite, a tool that reverts a "fixed" commit and reruns tests to check whether any test actually guards the fix. The developer also hand-typed a four-line fix to their Past Forever language runner (PR #27) so it sets a non-zero exit code when a program crashes, instead of reporting success to CI. "A program has two ways of answering. What it prints is for people. The exit code is for other programs," the developer wrote.

recommendation

After an AI agent claims a bug is fixed, revert that commit and rerun your tests; if they still pass, no test guards the fix and you need one.

more in playbook: Working with AI coding agents
read3 min views8 publishedNov 20, 2026

This week, maintainers merged three of my pull requests into projects I don't own. Then, a little before midnight, I typed my first fix by hand.

It was four lines. I was nervous anyway.

I'm new. Not "new to Rust" new. New new. Most of the code under my name was typed by Claude Code. My job so far has been the part after the AI says "fixed": running it, checking it, and asking whether anything would notice if it was wrong.

That habit turned into a tool, Bite. It takes a "fixed" commit back out of your code and reruns your tests. If they still pass, no test guards that fix.

This week I decided "checking the AI" isn't the same as knowing how to code. So I started learning.

My language project, Past Forever, has a command-line runner. When a program crashed, the runner printed the error and then told the computer everything went fine. Exit code 0. Any script or CI job running it would shrug and carry on.

Here's the whole fix (PR #27):

-  console.log(F.text(src, { why }));
+  const result = F.text(src, { why });
+  console.log(result);
+  if (result.includes('stopped: ')) process.exitCode = 1;

Claude walked me through it step by step. I typed every character in the GitHub web editor, ran the tests, and then had to explain it back in my own words before merging. My explanation was one sentence. It counted.

What I learned: a program has two ways of answering. What it prints is for people. The exit code is for other programs. Mine was lying to the programs.

docling-pipelines #686 (merged). The reviewer asked for five changes. One was that when you escape text, "the backslash must be escaped first." I pushed one commit that covered every point, and it was approved the same day.

django-modeltranslation #831 (merged). Bite found this one: two settings helpers reset a value to nothing when they finished, instead of putting back what was there before. I opened it on October 27. It sat for three weeks. I sent one polite nudge. It got merged this week.

django-modeltranslation #832 (merged). The maintainer requested a change: don't patch Django when the file loads, do it in the app's ready() step, with a guard so it only happens once. I made the change, added a test, and it was merged the same day. My PR description said up front that I wrote it with help from Claude Code.

wardenIQ is a test-case generator that can run on a local AI model. I test it on a CPU-only setup, which turns out to be useful: I'm the slow machine.

Outside contributors have been helping on my repos for weeks. This week two more showed up.

imshah-id opened Trawl PR #28. I reviewed it and asked for changes. The day before, I was the one getting change requests.

jeelthacker1 asked to take Bite #251. I said it's theirs.

AI did most of the typing in everything above, and helped draft this post. Every link and number in it was checked against the source. The week's real change is that I'm starting to read what it writes, and sometimes write it myself.

If you're also learning with an AI next to you: make it explain, then explain it back. And check its "fixed" claims. Some of them don't hold.

I credit everyone who breaks something. There's a Hall of Fame with a Breakers list.

pipx install bite-check, then bite --free. Small issues, each written to be a good first one:

I'm connabyss on GitHub. I post here when something breaks or ships, which this week was both.

── more in #ai-tools 4 stories · sorted by recency
── more on @claude code 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/ai-wrote-most-of-my-…] indexed:0 read:3min 2026-11-20 · —