Debugging and codebase interviews: practise on real bugs.

The setup is simple: a realistic codebase, a symptom and a clock. Stripe’s Bug Bash is the best known example. Candidates describe an unfamiliar repository with a failing test or an open issue, worked in their own language and editor for 45 to 60 minutes. Other companies run similar codebase rounds under other names, and AI-enabled rounds like Meta’s often open with the same kind of bug.

• A method: reproduce, narrow down, form a hypothesis, test it
• Reading skill: following data through files you did not write
• Root cause over symptom: fixing why it broke instead of hiding the failure
• Clear narration: a well explained diagnosis can count for more than a rushed fix

1. 01

   Reproduce it

   Run the failing test or the steps in the ticket first. Read the actual error, not your guess about it.

2. 02

   Find the path

   Trace from the entry point to where the value goes wrong. Search for names, follow calls and read the tests that cover the area.

3. 03

   Test a small hypothesis

   Change one thing that would prove or disprove your idea. Print or assert before you rewrite.

4. 04

   Fix the cause

   Make the smallest correct change that fits the code around it.

5. 05

   Prove it

   Rerun the failing test, then the whole suite, so you know nothing else broke.

Toy examples hide the hard part. Real projects have layers, conventions and tests that interact, and the bug is rarely where the stack trace first points. The only way to get faster at orienting yourself is to do it on many codebases you have not seen.

Every StackYard task is a bug from a real open source project or a realistic multi-module service, written up as a ticket with steps to reproduce and the expected and actual behaviour.

• Real tickets from real open source repositories and realistic services, in Python and TypeScript
• A browser IDE with a terminal and the project’s own test suite, nothing to install
• Pair, an AI assistant beside the editor that reads the code you are working on
• Example cases to run while you work, and hidden checks when you submit, so a pass means the fix holds
• A time budget on every task, from 10 to 45 minutes