tl;dr: I am tired of having to painfully put a shit ton of logging or breakpoints into AI code, which I haven't written, in order to debug it. And more generally, I'm tired of existing debugging techniques.
That's why for the last 4 months I built this tool called Ariana to help visually understand in your IDE what the code does at runtime, with minimal effort and prior knowledge about the code: https://x.com/AniC_dev/status/1934549383690752363
You can get started with it on VSCode/Cursor for free https://marketplace.visualstudio.com/items?itemName=dedale-d..., but it's weird and experimental so I advise you to read below first.
# Problem
AI code is weird to work with because it is often so subtly wrong and at the same time completely foreign to us, maintainers, now that we don't even write the code. How ironic I have to admit.
# Existing solutions
Traditional debuggers expect you (or the AI if it uses them) to know pretty well where to put breakpoints, and then you have to just step and step and step (but often can't even step back). It's not trivial for everybody (folks coding with AI aren't likely to be in the mood for traditional debuggers). Also you have to reproduce your bug until you reach interesting breakpoints which is annoying enough as of itself even pre-AI imo. Oh and forget about using debuggers on production code lol.
And logging, well at least AI is half decent at doing it, but it's often not already logging the exact right thing at the exact right place the first time you encounter a bug, especially in prod. So you also need to reproduce. Furthermore, it's annoying to maintain enormous amounts of logging, and to read walls of logs. Not even mentioning learning how to best handle observability paradigms, platforms, dashboards... as a small to medium-sized project's maintainer that's not the mood.
# Solution
With Ariana you have basically 0 things to change with your code, just run your JS/TS/Python code normally from the terminal but add the `ariana` command before (for example: ariana python <insert your script>). Behind the hood, it transforms your code like a transpiler, adds custom observability, it runs and watches that instrumented code, and reconstructs logic-flow & data-flow visualizations.
(Note that it will transform the files on my server, I'll publish a self hosted version soon. For now I did it this way so I could later add LLMs into the mix, so that observability will be made custom to your program's semantics.)
Then I made a VSCode extension to consume the debugging data like you've seen in the vid: https://marketplace.visualstudio.com/items?itemName=dedale-d... It installs the CLI for you btw.
You can also copy all the data and give to your agent to debug but for now they are too verbose (working on improving that too).
The whole philosophy of the tool is to always run your code with it, know all that the code did the first time a bug appears, so you never have to reproduce bugs again. At least that's the vision. Hope one day some projects will even use it in prod and not just to debug locally (I've heard everything is IO bound on mutexes out there so could be fine).
Thanks for your attention :-)