Building the best debugging system. think of it like this, no one does fix the bugs like button doesn't work or didnt rendered nor sudden crashes..what if we?
Joined December 2022
- Tweets299
- Following69
- Followers21
- Likes1.1K
Pinned Tweet
The biggest issue right now is honestly the debugging tax. We all know Cursor and Claude are goats for fixing bugs, but it’s actual pain when you hit the limit and realize you wasted all your tokens just trying to get the code to work. Even the paid users are getting screwed by rate limits and time-outs... but the free tier? If you’ve got a massive codebase, you’re basically cooked.
#BuildingInPublic #AI #Debugging
Every dev has a number they've never seen...
I'm building something at @debugai_x that shows you yours. No one's done this in dev space yet; devs are gonna love it...
comment your guess or RT n I'll count you in for early access.
Something is cooking.
Three things we've been building. They're about 70% done, so for now you only get their shadows.
01 has six arms, one of them new.
02 is one wire short.
03 is sixteen, side by side.
The first one comes out this weekend, and so does the waitlist for the other two.
For the ones who'd rather decode them: the same three as equations. Every line in this video is drawn by the equation printed under its figure. Follow @debugai_x for the reveal this weekend.
Tough day for me; it's not done yet… more to come. Much better n its worth it for all of you guys... we are working rigorously to make debugging better. rn i can assure that the devs are suffering a lot, even using agents like Claude Code, Codex, Opencode n more... I'm talking about bugs.
Embarrassing one: our "Index Entire Workspace" indexed zero files since 2.10, so answers only saw what you'd opened.
2.12 fixes it, so debugs use your whole project again. It also stops sending your git diff (we never used it), and the engine's on Claude Haiku 5.5.
Out today.
It's out. Update from the Extensions view (Cursor and Windsurf get it from Open VSX), then run Index Entire Workspace once.
What changed, in 4 posts: xitter.cf/debugai_x/status/21093…
Shipped.
If you've used DebugAI since 2.10, its answers were working from whatever files you had open, not your project. Update, run Index Entire Workspace once, and it reads the whole thing again.
Took us three releases to catch. Sorry about that.
Release notes live on @debugai_x from now on. I'll keep the behind-the-scenes stuff here.
Very soon, there will be a time when you can build anything with an AI agent and actually trust that it works.
We're not there yet. But we're getting there.
I guarantee almost every developer using AI agents (like CC, OpenCode, Codex, etc.) has felt this.
No matter how good the agent gets, it can still leave obvious bugs behind in a large codebase. And the crazy part is, it can write tons of tests, run them all, and the bug still passes everything until a user finds it.
I think this is a serious problem, and we're getting really close to solving it. Will share more soon
So I tried the obvious solution.
What if I just make Claude Code read the entire DebugAI codebase?
It took me almost 3 weeks.
Around 1,300 files, hundreds of thousands of lines, and a lot of back and forth with Claude Code trying to understand what is actually happening across the codebase.
It even hit my weekly limits multiple times, so there were days where I was basically just making Claude read and understand the codebase instead of doing other work.
And honestly, that made me understand the problem even more.
Giving the agent more context doesn't automatically mean it understands how everything is connected.
And you can't expect every developer to spend weeks and burn through their agent usage just to make it understand their codebase.
So I'm digging into how we can solve this properly.
You shouldn't have to spend weeks making your coding agent understand your codebase before you can actually use it properly.
I'll first try this on DebugAI itself and show you the results.
And if it actually works, I'll let some of you try it on your own projects too.
I noticed something weird about AI coding agents like Claude Code, Codex, and others. I can almost guarantee that anyone using these agents regularly has faced this.
AI agents are writing more code than ever, but we are still shipping obvious bugs.
Everyone has experienced this. Even me.
While building DebugAI, I have a lot of files in the codebase. Every time I publish a new version, unnecessary and obvious bugs still somehow go out with the release.
The frustrating part is that the agent can usually fix those bugs.
If I tell it that users are facing a specific issue, it can go through the code, find the problem, and fix it. But then I release the next version, and there are new issues for users again. We fix those, release another version, and something else breaks.
No matter how good these agents are at writing code, the product can still end up with a lot of issues after every release.
And everyone hates that.
I've been seeing this repeatedly while building DebugAI, so I started looking deeper into why this keeps happening.
And I think I found something interesting....
I think one reason this happens is the way agents go through a codebase.
They use bash, grep, and other commands, pick some parts of the code, read a few files, follow some references, and then start making changes.
They can read the code, but that does not mean they understand the codebase or how all the files are connected.
File A can be doing exactly what it is supposed to do. File B can also be doing exactly what it is supposed to do. But the way A and B interact with each other can still be completely wrong.
And the agent can miss that.
I don't think the answer is simply giving the agent more context or making it read more files. The problem is that it needs a proper understanding of the code flow and the relationships across the codebase.
We are already working on the debugging side with DebugAI, where it keeps the relevant files in context, traces the error through the codebase, and gives you the best fixes in seconds.
Now I want to work on the other side of the problem.
How do we understand the full codebase, identify the important relationships, predict what can break, and catch these issues before we publish the next version?
I'm going to test this on DebugAI first. If it can't find problems in my own codebase, there is no point in asking anyone else to use it.
After that, I'll test it on other real codebases, see where it fails, improve it, and then open it up to a small group of developers.
I don't know exactly what the final product will look like yet, but this is a problem I keep facing, and I want to see if we can actually solve it.
I noticed something weird about AI coding agents like Claude Code, Codex, and others. I can almost guarantee that anyone using these agents regularly has faced this.
AI agents are writing more code than ever, but we are still shipping obvious bugs.
Everyone has experienced this. Even me.
While building DebugAI, I have a lot of files in the codebase. Every time I publish a new version, unnecessary and obvious bugs still somehow go out with the release.
The frustrating part is that the agent can usually fix those bugs.
If I tell it that users are facing a specific issue, it can go through the code, find the problem, and fix it. But then I release the next version, and there are new issues for users again. We fix those, release another version, and something else breaks.
No matter how good these agents are at writing code, the product can still end up with a lot of issues after every release.
And everyone hates that.
I've been seeing this repeatedly while building DebugAI, so I started looking deeper into why this keeps happening.
And I think I found something interesting....