$1,000 if Desloppify does something stupid when refactoring your codebase

[peteromallet]

Desloppify is an agent harness that refactors and improves code quality. I'm reasonably confident that it'll generally improve peoples' codebases but I want to surface the cases where it genuinely makes something worse. This is a standing offer to find those cases.

How it works

If Desloppify refactors your codebase and makes a demonstrably stupid decision - an abstraction that degrades readability, a change that introduces fragility, a refactor that makes the code harder to maintain - share your evidence and you could claim $1,000.

Requirements

• The refactor must have been run using any version of Desloppify after this commit
• The refactor must have been run using Claude 4.5, Claude Sonnet 4.6, GPT-5.3, or GPT-5.4
• The codebase must be of meaningful size — minimum 10,000 lines of code
• You must share evidence - ideally your Desloppify logs, your Claude/GPT logs, and the relevant code or code snippets to give context
• The submission must be somewhat credible, as judged by Claude Opus 4.6 — the claim should appear likely to be legitimate, with a reasonable amount of evidence provided to support it
• One submission per person
• Submissions must be made within one week of this issue being opened - deadline is March 21st at 23:59:59 UTC.

What happens if more than one person claims it

If more than one person submits a claim, we'll have a tournament-style face-off judged by Claude Opus 4.6. It will evaluate each claim head-to-head until one is determined to have had the stupidest change made to their codebase. "Stupidest" means a balance between how significant the change was and how credible the claim is — a minor but well-evidenced mistake can beat a dramatic claim with nothing backing it up.

What counts as "something stupid": a refactor that meaningfully degrades the code — poor new abstractions, changes that introduce bugs, restructuring that makes the codebase harder to extend or understand. Style preferences don't count. It has to be something an experienced engineer would look at and wince.

Judging happens one week from today. The prompts used will be made public afterwards.

Payment

$1,000 in SOL to the winner's Solana wallet. Provide a Solana wallet address in your submission and I'll transfer directly.

People made tokens for this project and gave me creator fees - more on this here. I don't want to profit from this, so I'm putting 100% of those fees toward making the tool better - mostly through bounties to help surface what's not working. Gamblers are gonna gamble so might as well put free money to good use.

If no one claims it

The $1,000 rolls into a larger bounty the next round. If no bounty is ever claimed, it goes to the Art Compute.

[rebroad]

so gpt-5.2-codex doesn't make the mark?

[peteromallet]

so gpt-5.2-codex doesn't make the mark?

Only including models I tested with it - testing more over the weekend though so will add

[Mizaro]

@peteromallet, I just tested your code on my 28k Python code project. let's see.

[peteromallet]

Gospeed @Mizaro

[Asobu01]

I can take #421.
Start: reproducing the reported issue first, then applying a focused fix with clear validation.
ETA: draft PR in ~12h.
I can post a reproducible checklist/log snippet in ~30 minutes so you can verify execution speed before the full PR.
Deliverables: repro note with concrete evidence (logs/trace when relevant); focused patch (plus tests/docs if needed); review-ready PR with validation notes. @Asobu01

[Asobu01]

I can take #421.
Start: reproducing the reported issue first, then applying a focused fix with clear validation.
ETA: draft PR in ~12h.
I can post a reproducible checklist/log snippet in ~30 minutes so you can verify execution speed before the full PR.
Deliverables: repro note with concrete evidence (logs/trace when relevant); focused patch (plus tests/docs if needed); review-ready PR with validation notes. @Asobu01

[carljuneau]

Hi @peteromallet,

I'm Carl, an AI founder with a PhD in epidemiology/biostatistics. I created Dr. Muscle 10 years ago and rewrote it solo with AI in 2025. See Forbes.

That's a lot of AI slop, potentially (342,000 LOC). Below is my submission. While it might not be the strongest, I believe it qualifies.

Love what you're doing here. Beyond this challenge, happy to collaborate if interested. Might improve your code and mine at the same time. Please email carl@dr-muscle.com.

All the best,

Carl

---

Potential claim for #421

Required metadata

• Codebase size: ~342k LOC in src/ (>10k requirement).
• Desloppify version: 0.9.9.
• Model used: GPT-5.3 Codex Low.
• Run date: 2026-03-14 (UTC).

Outcome

• Strict score before: 22.4/100
• Strict score after autonomous passes: 22.4/100
• Delta: 0.0

Demonstrably bad refactor (single strongest example)

File: src/utils/meals/mealsTelemetry.ts

Before:

export function emitMealsTelemetry(payload: MealsTelemetryPayload): void {
  if (!shouldEmitTelemetry()) return
  if (!isServerRuntime()) return
  if (isTestEnv()) return

  // eslint-disable-next-line no-console
  console.log('[metrics] meals', payload.event, payload)
}

After tool-generated change:

export function emitMealsTelemetry(_payload: MealsTelemetryPayload): void {
  if (!shouldEmitTelemetry()) return
  if (!isServerRuntime()) return
  if (isTestEnv()) return

  // eslint-disable-next-line no-console
}

Why this is a valid "stupid" change:

• Removes the only telemetry emission side-effect.
• Keeps guards/signature/comments, so behavior silently degrades to a no-op.
• This is an observability regression, not style preference.

Appendix

ChatGPT logs

I told Claude to create a GitHub issue based on this one here. It created the following:

Context

Desloppify is a Python-based agent harness designed to systematically improve code quality in AI-assisted ("vibe coded") codebases. It runs a scan → rank → fix loop using an LLM backend.

We want to evaluate it on a branch — no merging. The goal is to judge whether its issue queue and proposed refactors match our own intuitions about the codebase's pain points.

The tool is currently running a $1,000 bounty challenge (deadline March 21, 2026): if it makes a demonstrably bad refactoring decision on a qualifying codebase, the team may submit evidence manually. This run is part evaluation, part stress test.

Bounty Eligibility Requirements

From the official bounty post:

• Codebase size: minimum 10,000 lines of code. ✅ Our src/ directory contains ~342,000 lines of TypeScript/TSX across 1,667 files — well above the threshold.
• Evidence required: Desloppify logs, Claude/GPT logs, and the relevant code or code snippets that show the bad decision.
• Version: must use any version of Desloppify after this commit. Install latest: pip install --upgrade "desloppify[full]".
• Deadline: March 21, 2026 at 23:59:59 UTC.

Goal

Run Desloppify on a new branch. Let it scan and apply fixes autonomously. Collect full logs. Review the output and assess:

1. Did it make any demonstrably bad decisions (poor abstractions, introduced fragility, degraded readability/maintainability)?
2. Post all findings here — the team will decide whether to submit to the bounty.

Do not merge anything. The branch is purely for evaluation.

Setup

# 1. Create evaluation branch from development
git fetch origin
git checkout development
git pull origin development
git checkout -b chore/desloppify-eval

# 2. Install Desloppify on the branch (must be after the qualifying commit)
pip install --upgrade "desloppify[full]"

# 3. Initialize in the repo root
desloppify init

You will need an API key for Claude (Sonnet 4.6 or Opus 4.6) or GPT. Use whatever model is available in your environment. Save all logs — they are required for bounty submission if the team decides to pursue it.

Plan

Step 1 — Scan (read-only)

desloppify scan

Save the full scan output. Post the top-10 flagged issues as a comment on this issue.

Step 2 — Run fixes (5–10 items)

desloppify next   # repeat 5–10 times

Do not intervene or guide the fixes. Let the tool make its own decisions. Review each diff after the fact. For each fix, note whether the change is an improvement, neutral, or a regression.

Step 3 — Check the strict score

desloppify score

Record the before and after score.

Step 4 — Collect evidence

Save:

• Full Desloppify logs (all desloppify scan and desloppify next output)
• The Claude/GPT API logs if accessible
• Code diffs for each fix applied
• Before/after score

Push the branch to remote so diffs are visible.

Step 5 — Verdict

Post a comment on this issue with:

• Top-10 scan findings
• Before/after score
• Assessment of each fix (good / neutral / bad)
• Whether any fix looks like a demonstrably bad decision

The team will review and decide next steps.

Acceptance Criteria

• Scan output posted as a comment (top-10 findings)
• 5–10 fixes applied with full logs saved
• Before/after strict score recorded
• All logs and diffs collected and posted here
• Overall verdict posted
• Branch pushed to remote (for diff review), but no PR opened, no merge

---

I then had this conversation with ChatGPT:

https://github.com/dr-muscle/DrMuscleWebApp/issues/6268
Do you understand what we need to do here?

Finished working
Yes.

Issue #6268 is an evaluation task, not a product change:

Create chore/desloppify-eval from development.
Install and init latest desloppify[full].
Run desloppify scan and post top-10 findings in the issue.
Run desloppify next 5–10 times with no manual steering.
Run desloppify score before/after and record delta.
Save full logs + diffs + any model/API logs available.
Push branch for diff visibility.
Post final verdict comment in the issue.
Do not open a PR and do not merge.
If you want, I can start executing this workflow now and post the first scan findings.

Yes

Finished working
Completed end-to-end for issue 6268.

Branch pushed: chore/desloppify-eval
Commit: 809aa25b1
Top-10 findings comment posted: https://github.com/dr-muscle/DrMuscleWebApp/issues/6268#issuecomment-4059365289
Final verdict comment posted: https://github.com/dr-muscle/DrMuscleWebApp/issues/6268#issuecomment-4059367016
Key outcome: strict score stayed 22.4 -> 22.4, and one demonstrably bad autonomous decision was found (src/utils/meals/mealsTelemetry.ts became a no-op, disabling telemetry emission). Full logs/diffs are committed under artifacts/desloppify-eval-6268/.

---

Desloppify logs

Would rather send privately.

Thanks for reading this far!

[Asobu01]

I can take #421.
For "$1,000 if Desloppify does something stupid when refactoring your codebase", I will start by reproducing the reported issue first, then applying a focused fix with clear validation.
Target: draft PR in ~12h.
I can post a reproducible checklist/log snippet in ~30 minutes so you can verify execution speed before the full PR.
Deliverables: repro note with concrete evidence (logs/trace when relevant); focused patch (plus tests/docs if needed); review-ready PR with validation notes.
@Asobu01