Блог пользователя AmShZ

Автор AmShZ, 4 дня назад, По-английски

Hello everyone!

Have you ever wondered what a competitive programming contest for real-world software engineering would look like?

Instead of writing code from scratch, imagine being given an existing project full of bugs, with the clock ticking and a live leaderboard showing everyone's progress.

That's what we're building at Repovive!

We're excited to announce Repovive SWE Round 1: Vive Market, a live debugging contest where your goal is to find and fix bugs in Python projects.

You'll also have an AI agent working alongside you. The interesting part is that the agent knows what the software is supposed to do, but hasn't seen the actual implementation. More on that below!

Prizes

We have a $750 prize pool for this round!

  • 1st place: $150
  • 2nd place: $100
  • 3rd place: $50
  • 4th and 5th place: $25 each

There are also additional prizes:

  • Random Top 50: 10 randomly selected participants from the top 50 will receive $20 each.
  • Random Participants: 10 randomly selected participants from all participants will receive $20 each.

The problems

Each problem is a small Python project containing several bugs. Most bugs can be fixed by changing just one or two lines of code.

You can approach the problems in any order. Every project contains some easier bugs, so we recommend exploring all four projects before spending too much time on a single one.

The goal is not to rewrite everything. It's to understand the existing code, identify what's wrong, and make the smallest changes needed to fix it.

Scoring

  • Each bug you fix earns you points. When you submit, every bug is checked using its own public and hidden tests.
  • If your changes break something that previously worked, that submission receives 0 points.
  • You can submit as many times as you want. Only your best score on each problem counts, so a bad submission never lowers your existing score.
  • Your final score is the sum of your best scores across all four problems.
  • Submissions made after the contest ends do not count.

Tiebreakers

If two participants have the same total score:

  1. The participant with fewer submissions ranks higher.
  2. If they are still tied, the participant who reached that score earlier ranks higher.

So make sure to test your changes before submitting!

Your AI teammate

Every workspace includes a built-in AI agent to help you understand the project.

The agent has access to:

  • The rule book.
  • The project's data files.
  • An outline of the code, including function names, signatures, and docstrings.

However, the agent has not seen the actual function bodies. It doesn't know where the bugs are, cannot execute code, and cannot modify your files.

Think of it as a teammate who understands exactly what the software is supposed to do, but has never seen how it was implemented.

You investigate the code, and the agent helps you understand the intended behavior.

Using any AI tool other than the built-in agent is strictly prohibited. This includes ChatGPT, Copilot, Claude, and any other AI assistant or code completion tool.

We're excited to explore this new contest format, and we'd love to have you join us!

Whether you're an experienced competitive programmer or simply enjoy debugging and understanding real software, we hope you'll find something interesting.

Since we're still experimenting with this format, your feedback will be especially valuable to us.

See you on the leaderboard!

  • Проголосовать: нравится
  • +56
  • Проголосовать: не нравится

»
4 дня назад, скрыть # |
← Rev. 6  
Проголосовать: нравится +5 Проголосовать: не нравится

Wow nice idea for contest

  • »
    »
    4 дня назад, скрыть # ^ |
     
    Проголосовать: нравится +25 Проголосовать: не нравится

    if by "nice" you mean "terrible and not fun", then yes

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +20 Проголосовать: не нравится

As a tester, I had a great time testing these problems :>

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +26 Проголосовать: не нравится

As a tester, use your AI teammate.

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +24 Проголосовать: не нравится

As a Problem setter, I highly recommend attempting all the questions :)

»
4 дня назад, скрыть # |
← Rev. 2  
Проголосовать: нравится +16 Проголосовать: не нравится

clashes with upcoming codeforces round:(

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +10 Проголосовать: не нравится

Can't I just copy paste the code to the agent and ask it what the bugs are?

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +15 Проголосовать: не нравится

as a tester, GLHF

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +23 Проголосовать: не нравится

As a tester, I really like this idea. It is a great simulation of how software engineering interviews will look in the future.

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +15 Проголосовать: не нравится

As a Problem setter, I enjoyed setting the problems

»
4 дня назад, скрыть # |
← Rev. 2  
Проголосовать: нравится +13 Проголосовать: не нравится

Time is updated to Saturday

Start Time: Saturday, 14 November 2026, 14:35 (UTC)

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +8 Проголосовать: не нравится

I'd call it bug hunt rather than software engineering when it's small toy projects with clear spec that'll never feel the touch of a real user.

  • »
    »
    4 дня назад, скрыть # ^ |
     
    Проголосовать: нравится +3 Проголосовать: не нравится

    We kept the projects small on purpose so the format fits in 3 hours. Bigger, messier repos are coming as we learn what works.

    Would love your feedback if you give it a try <3

    • »
      »
      »
      2 дня назад, скрыть # ^ |
       
      Проголосовать: нравится 0 Проголосовать: не нравится

      As a former maintainer of a repo with multiple 10k+ LOC files, the best SWE improvement of a large project is often... splitting it into multiple smaller focused ones. As stuff gets larger, complexity in detail takes a back seat to complexity in high level structure, so if you just increased size, you'd just end up with a hard "dig through this Epic of Codegamesh" followed by comparatively trivial bug hunt. There's definitely a SWE problem in redesigning code (repo/directory/library/package structure, files, functions...) but it's not a problem with an objective solution.

      In fact I'd argue that any true SWE problem doesn't permit an objective grading system by definition. Real life problems aren't constrained like that. Figuring out what sort of problem you should even be solving is the hardest part, designing how solutions could possibly look is hard, implementation is comparatively trivial. There's an extremely important user factor (who's going to use this, in what ways) and an important future factor (what happens if something needs to change; this includes QA).

      The name "bug hunt" describes the problems this contest describes much more accurately.

      • »
        »
        »
        »
        2 дня назад, скрыть # ^ |
         
        Проголосовать: нравится 0 Проголосовать: не нравится

        Yeah that's fair

        The hard part of real SWE is figuring out what to build and designing it well and you can't really auto grade that We're only testing a small piece of it here: jumping into code you didn't write and fixing it without breaking anything

        Thanks for the feedback, appreciate it <3

»
4 дня назад, скрыть # |
 
Проголосовать: нравится +5 Проголосовать: не нравится

As a tester, I had a really fun time trying out these problems. The idea of debugging a real project with an AI teammate is super cool and different from standard CP. The progression from easy to medium was smooth, and it's awesome how the team actually listens to feedback and updates things to make it clearer. Highly recommend giving it a shot!!!

»
3 дня назад, скрыть # |
 
Проголосовать: нравится +11 Проголосовать: не нравится

finally a contest where the bugs are someone else’s fault

»
3 дня назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

this gives me ptsd from the amazon oa please no

»
3 дня назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

why no c++

»
3 дня назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

can i verbally abuse my ai agent, i mean thats what i do with my cp team mates

»
2 дня назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

ONE HOUR TO GO!

»
2 дня назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

when will the leaderboard be public? or it won't?

»
2 дня назад, скрыть # |
 
Проголосовать: нравится +27 Проголосовать: не нравится

cussing the agent out was surprisingly effective

»
2 дня назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

Interesting contest, it's good for me to know what small Python projects looks like, and also let me learn to read and understand code that I seldom use under AI assist.

»
2 дня назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

Need to ban cry he is just too good. Share your secrets man how do you do it so fast :)

»
46 часов назад, скрыть # |
 
Проголосовать: нравится +26 Проголосовать: не нравится

Some observations:

  • Forbidding the agent from reading implementation was only a mild inconvenience — it explicitly asked me to copypaste the function bodies.
  • It doesn't know initially where the bugs are and won't tell me directly how to fix them, but given enough input-output from code it tells me to run, it'll figure out what the bugs are and where, at least in principle, and then will start hinting me that info. For example when explaining several steps a function takes, it can make the step with a bug bolded.
  • Nonetheless there's no chance to figure out how to test what incident, read a pile of rules, work out expected outputs by hand and what parts of implementation intend to handle those rules, on time without this language processing tool.
  • The agent defaults to asking me to work out those expected outputs by hand. When I point out it's difficult for a human, it quickly relents and is willing to narrow down scope or check stuff for me within the limits of being unable to run code.
  • My approach quickly became: fix the public tests, which takes care of some bugs, then ask the agent for commands that showcase remaining bug reports and let me test correctness, run them and keep dumping the outputs and asking the agent to narrow down the bug's location; then asking it to state what rule a line should follow but doesn't; then guessing fixes until the agent approves.
  • There are two types of bugs: obvious nonsense like wrong sign of inequality, array indexing, random extra crap that has nothing to do with the function at hand etc; and then logic errors in what's intended to happen. The former are very easy to fix — in fact so easy that when I submitted D with 5 minutes left and saw there are 2 more easier bugs I didn't know about, I was able to guess the fixes in time with only info from the agent about what functions are involved in the logic paths of those bugs' reports. The latter require a lot more testing and comparing (also the agent's work) output against rules.
  • Difficulty of subtasks (individual bugs) is reasonably scaled. The problems don't appear to be in any order of difficulty.
  • Sometimes I found rewriting a function into much simpler form easier than trying to figure out what it's doing.
  • »
    »
    45 часов назад, скрыть # ^ |
     
    Проголосовать: нравится 0 Проголосовать: не нравится

    If you were to make the agent, how would you make it behave?

    Or, what do you think the perfect experience for that agent would be for those types of tasks?

    • »
      »
      »
      40 часов назад, скрыть # ^ |
       
      Проголосовать: нравится +8 Проголосовать: не нравится

      You need an umpire that understands the goals of the contest and can make own judgements on what actions would follow those goals the best, independent of the rules or the contestants' input. Basically: the perfect experience is AGI.

      It's a fundamental problem that LLMs are definitionally incapable of handling. An LLM doesn't understand subtext, doesn't have a spirit of the rules to follow and can be coerced (either by you or the contestants) towards actions that fail a sanity check. There are no helpfulness sliders that can be tuned so the clanker is both sufficiently helpful to let contestants overcome the problem of scale and not so helpful that it wouldn't guide towards a solution because the two are deeply tied together. Choosing and evaluating a useful test, for example, is both. Deciding what sort of test/result would be too useful and de facto doing the contestant's work? That's difficult without true, deep understanding of the problem.

      The projects aren't actually what I'd call "small", which makes the problem of scale the most difficult subproblem that needs to be handled to solve anything. It's one humans are bad at in contest environment, both due to a steep learning curve and limited short-term memory/context with relatively little to draw on — an algorithmic contest uses broad long-term memory of algorithms and a short statement while understanding a set of rules isn't about the language those rules are written in. If you let the agent be nothing but a summary/search tool for what's directly written in a project, you end up with 90% reading and understanding the rules, 9% understanding the implementation, 1% fixes.

      Another awkward thing is that the optimal strategy is understanding as little of a project as needed to fix only what needs fixing. That'll always be true. Headspace is limited, unnecessary context is distracting so pruning it should lead to faster performance. As it stands, the agent massively helps with that, which is by definition helping solve the problem. Even in any ideal situation, good performance would necessarily equal having little clue about most of the problem's project/repo. That works just fine for a contest problem... yet it's the opposite of good software engineering. Competitive programming just doesn't translate well into the real world, no matter what programming.

      • »
        »
        »
        »
        26 часов назад, скрыть # ^ |
         
        Проголосовать: нравится +3 Проголосовать: не нравится

        This is really useful thanks for taking the time <3

        The agent asking you to paste the code and hinting the bug location are both things we'll fix and you're right that the "obvious" bugs are too cheap next to the logic ones

        Taking all of this into the next round :)

»
30 часов назад, скрыть # |
 
Проголосовать: нравится 0 Проголосовать: не нравится

When will the prizes be distributed? AmShZ GreatShinobiOwl