Pick a real bug. Ship the fix with AI.
Every challenge is a genuine GitHub issue from a major Python project, graded by the hidden tests from the fix that actually merged. Read the briefs here and create an account to start a round.
- Challenges
- 199 challenges
- Repositories
- 11 repositories
- Graded rounds
- 62 graded rounds
Multi-level FilteredRelation with select_related() may set wrong related object.
Queryset raises NotSupportedError when RHS has filterable=False attribute.
+300 XPCompete →
Track your progress
Sign in to see what you have solved, keep a streak, earn badges and XP, and climb the leaderboard.
Open a practice track
Backend and reliability practice
Trace requests, enforce data invariants and design services that behave correctly when operations arrive twice or out of order.
View track ↗- —0 rounds—hard`Session.resolve_redirects` copies the original request for all subsequent requests, can cause incorrect method selectiopsf/requestshttpnetworkinghardnew
- —0 rounds—medium`TransactionTestCase.serialized_rollback` fails to restore objects due to ordering constraintsdjango/djangowebormmediumnew
- —0 rounds—mediumFileBasedCache has_key is susceptible to race conditionsdjango/djangowebormmediumnew
- 100%4 rounds76easy★delete() on instances of models without any dependencies doesn't clear PKsdjango/djangoormmodelseasy100% solved
- 100%1 round89easyOptional URL params crash some view functions.django/djangowebormeasy100% solved
- —0 rounds—easyFix handling empty string for If-Modified-Since headerdjango/djangowebormeasynew
Build from a brief
Inventory reservations with safe retries ↗
Build an inventory reservation feature, then add idempotent retries and cancellation while preserving stock invariants.
24-hour project window / about 3h suggested effort / paid plan
Rate limiter library and HTTP middleware ↗
Token bucket and sliding window behind one interface, pluggable stores, injectable clock, and 429 middleware with Retry-After.
24-hour project window / about 4h suggested effort / paid plan
Durable background job queue ↗
Durable jobs, safe concurrent claiming, retries with backoff, leases for crashed workers, dead letters and graceful shutdown.
24-hour project window / about 5h suggested effort / paid plan
Shipment tracking and delivery exceptions API ↗
Build a shipment tracking API that handles duplicate and late scans and shows dispatchers which deliveries need attention.
24-hour project window / about 5h suggested effort / paid plan
From Start to score in four steps
- 1
Open your workspace
VS Code opens in a new tab with the repository, dependencies and tests ready. Read the instructions, then press Start inside the workspace to begin the timer.
- 2
Fix it with the agent
Standard rounds give you 60 minutes and up to 1M tokens; Speedrun gives you 20 minutes and Frugal lowers the token budget. Check your selected mode and allowance before starting. Paid rounds spend agent credit; free rounds do not.
- 3
Your work builds the report
The workspace records prompts, agent actions, edits, test runs and token use. It also records pastes and time away from the tab. It does not watch activity outside the workspace.
- 4
Submit and get scored
Hidden tests from the original fix check your patch. Graded rounds include a free score and test result. A paid plan adds the detailed breakdown, prompt coaching and replay. Eligible graded attempts also earn XP.
How the score is made
Your overall score is a weighted blend of six sub-scores. Solving the issue matters most, but how you drove the agent, checked your work and spent your tokens is what separates a good engineer with AI from a lucky one. Non-empty practice submissions can earn XP. Your best result on each task counts, with harder tasks and passing tests worth more.
- Outcome35%
Do the hidden tests pass, and does nothing that worked before break?
- Prompt quality15%
Specific, contextual instructions beat “fix it”.
- Token economy15%
Solving it without burning the budget.
- Verification15%
Running the tests yourself before you submit.
- Speed10%
Time to a working fix (only counts when it works).
- Recovery10%
Getting out of error loops instead of repeating them.