Nice work! I watched the demo and can see how it will generate fixes for you, which you then copy and paste into the editor. Perhaps you could consider automating this process like Aider[1] does, whereby you force the LLM to generate a git diff for the fix and automatically commit it.

1. https://github.com/paul-gauthier/aider

Haha thanks! Yeah, I think that's definitely a logical next step. We do something similar for the larger bug resolution platform we've been working on, so it shouldn't be too hard to port over!

I've come across Aider before and was trying to remember the name just the other day - thanks!!

I thought of something similar these days but with a different approach - rather than settrace, it would use a subclass of bdb.Bdb (the standard library base debugger, on top of which Pdb is built) to actually have the LLM run a real debugging session. It'd place breakpoints (or postmortem sessions after an uncaught exception) to drop into a repl which allows going up/down the frame stack at a given execution point, listing local state for frames, running code on the repl to try out hypotheses or understand the cause of an exception, look at methods available for the objects in scope, etc. This is similar to what you'd get by running the `%debug` magic on IPython after an uncaught exception in a cell (try it out).

The quick LLM input/repl output look is more suitable for local models though, where you can control hidden state cache, have lower latency, and enforce a grammar to ensure it doesn't go off the rails/commands implemented for interacting with the debugger, which afaik you can't do with services like OpenAI's. This is something I'd like to see more of - having low level control of a model gives qualitatively different ways of using it which I haven't seen people explore that much.

So interestingly enough, we first tried letting GPT interact with pdb, through just a set of directed prompts, but we found that it kept hallucinating commands, not responding with the correct syntax and really struggling with line numbers. That's why we pivoted to just getting all the relevant data upfront GPT could need and letting GPT synthesize that data into a singular root cause.

I think we're going to explore the local model approach though - you raise some really great points about having more granular control over the state of the model.

Interesting! Did you try the function calling API? I feel you with the line number troubles, it's hard to get something consistent there. Using diffs with GPT-4 isn't much better in my experience; I didn't extensively test that, but from what I did it rarely produced synctatically valid diffs that could just be sent to `patch`. One approach I started playing with was using tree-sitter to add markers to code and let the LLM specify marker ranges for deletion/insertion/replacement, but alas, I got distracted before fully going through with it.

In any case, I'll keep an eye on the project, good luck! Let me know if you ever need an extra set of hands, I find this stuff pretty interesting to think about :)

I actually coded something very close to this and it worked surprisingly well: https://github.com/janpf/debuggAIr

Ooh, interesting - starred and going to dig into this later today!

I've done a manual version of this with chatgpt.

I had ipdb, told it to request any variables that I should look at, suggest what to do next, what it would expect - it was quite good, but took a lot of persuading, just having an LLM that was more tuned to this would be better.

On the reddit discussion, one user pointed out that email adresses were incidently collected. https://www.reddit.com/r/programming/s/lBfxL7f2KM

Was removed here: https://github.com/leapingio/leaping/commit/e42d5198abe48875...

They should probably remove that Posthog API key as well...

Why would they do that??? Are we really going into this era where we have to reverse engineer every binary we touch?

This sounds great, but breaks pytest on the projects I've tried it on - both Django projects, both in different ways. I'm interested enough that I'll be keeping an eye on progress though!

Oof, I'm sorry to hear that - I don't think we had any Django projects in the set of projects we were testing this out on. I just filed an issue here and hopefully fix it asap - https://github.com/leapingio/leaping/issues/2

No worries! Good luck with the project.

> To achieve this, we first instrument the test using sys.settrace (or, on versions of python >3.12, the far better sys.monitoring!) to keep a history of all the functions that were called, along with the calling line numbers. We then re-run the test and use AST parsing to find all the variable assignments and keep track of those changes over time. We also use AST parsing to obtain the source code for these functions.

I don't want to be negative on someone's Show HN post, but it seems like getting all of this and showing it to the user would be way more helpful than showing it to the LLM?

My standard sometimes when I'm thinking about this kind of stuff is "would I want this if the LLM was swapped out for an actual human?" So would I want a service that gets all this useful information, then hands it off to a Python coder (even a very good Python coder) with no other real context about the overall project, and then I had to ask them why my test broke instead of being able to look at the info myself? I don't think I'd want that. I've worked with co-workers who I really respect; I still don't want to do remote debugging with them over Slack, I want to be able to see the data myself.

Going through a middleperson to find out which code paths my code has hit will nearly always be slower than just showing me the full list of every code path my code just hit. Of course I want filtering and search and all of that, but I want those as ways of filtering the data, not ways of controlling access to the data.

It feels like you've made something really useful -- an omniscient debugger that tracks state changes over time -- and then you've hooked it up to something that would make it considerably less useful. I've done debugging with state libraries like Redux where I can track changes to data over time, it makes debugging way easier. It's great, it changes my relationship to how I think about code. So it's genuinely super cool to be able to use something like that in other situations. But at no point have I ever thought while using a state tracking tool, "I wish I had to have a conversation with this thing in order to get access to the timeline."

Again, I don't want to be too negative. AI is all the hotness so I guess if you can pump all of that data into an LLM there's no reason not to since it'll generate more attention for the project. But it might not be a bad idea to also allow straight querying of the data passed to the LLM and data export that could be used to build more visual, user-controlled tools.

Just opinion me, feel free to disregard.

Not wanting to be negative too, but I’ve used debuggers like gdb, the ones in JetBrains’s IDEs, XCode’s, and every time it’s not lack of information that’s stopping me from solving the issue. Coding common lisp with sly and emacs or smalltalk with pharo is much entertaining than chatting with an LLM. Coding with a good debugger is very close to that (even the one inside the browser for JavaScript). I think we can design better tools than hook everything to a LLM that requires 128gb of ram to run locally.

Interesting! I think you’re right in saying that the middleman you’re talking about has to be really good for something like this to actually be useful, especially for people very comfortable with debugging tools + their codebase. If I understand correctly, you’re saying that the most productive tool for you would be one that can present you with more relevant data, in a structured way (Redux, etc). At first, we actually did think of making a nice IDE with all that data, but found out it kind of already exists - https://pytrace.com/, and we found it to be more cumbersome to use than anything! Our belief is that these tools can help, but there’s an irreducible amount of reasoning that needs to happen, which takes time and effort, and we think a tool like this might be able to reduce that by offloading the reasoning to LLMs. I guess what I’m saying is that there’s a cap on how much useful information a tool can give a developer to enable them to reason better, and I’m really interested in seeing if it’s possible to reduce how much reasoning is needed in the first place. Curious to hear what you think - thanks for the thoughtful comment!

:shrug: I'm one person on the Internet, so if using the LLM makes it work better for more of your users, go with that.

I do think data filtering and visualization is an important value add, Pytrace looks cool, but it doesn't look much different from what I get if I debug Javascript in a web browser, so I think there's a ton of room to improve. Visual representations of data transformations and code paths are a relatively unexplored area across the entire software industry imo.

If an LLM could flat-out reduce my need to reason and fix the bug for me, great -- but I've worked with coders that I respect a lot, and remote debugging with them has always been a pain and made narrowing down issues harder. I've never enjoyed debugging something remotely where I was working through a middleperson and couldn't look at what was going on; it's helpful to have multiple eyes on the code, but not if I have to use someone else's eyes to look at what's happening.

So for me, in order to successfully reduce the amount of reasoning I need to do and overcome the downside of me not being able to visualize the timeline/data, the LLM would need to be better at fixing these bugs than professional developers in industry: developers who are already intimately familiar with the codebases I'm debugging because they wrote a significant portion of the code. It would need to be better at coding than professional humans. And I just don't think there's anyone who would say that GPT-4 is close to that level yet.

What I could see is, maybe -- if I have access to that data, and the LLM is just kind of on-the-side, maybe at that point it can offer helpful advice and there wouldn't be a downside because I would still be able to debug as fast as I can using all of the available data, and if the LLM can occasionally find something I missed, then great. Peer-debugging sessions with multiple coders are great, so at least in theory I could see some value from an LLM on that stuff, even if I'm a little skeptical about potential performance. And if the LLM wasn't in front of the entire data, if it didn't work then no worries, the data is still there.

But again, if people like it, then it doesn't matter what I think. Why I wouldn't use the tool is less important than why someone would use the tool, and if integration with the LLM makes people want to use the tool, then... I mean, not everyone has identical work styles. Different things might work for different people.