AI-Assisted Debugging OA: Format, Strategy and Reported Repos

Since January 2026 many Amazon SDE online assessments end the coding section with a task in a code repository, done with a built-in AI assistant. 29 candidate reports from January to October 2026 are collected here, and 15 more describe the two repos with their own pages. This page covers the format, what the assistant will and will not do, every repo candidates have reported, and how they say they worked. Two repos come up often enough to have their own practice pages, each with a runnable reconstruction: the movie database app and the Jira-like issue tracker.

Format

• The coding section is 100 minutes: a 40-minute algorithm question, then the 60-minute repository task. The OA email describes it as two coding problems, "one will be a traditional code-writing question, and in the other you will have access to an AI coding assistant in a code repository environment." Each question's timer starts when you open it, and one candidate recalls that time left over from the first does not carry over to the second.
• It runs on HackerRank, in an IDE that candidates describe as VS Code-like. Reports say there is no camera, but you need a single screen in full screen, and pop-ups warn you not to switch windows.
• You choose your stack before starting, and one candidate says it cannot be changed once you are in. Reports name Django (Python), Spring Boot (Java), Node.js and C++, and one reply says you may pick a different language for the algorithm question.
• The usual setup is a repo, a task description in a README or markdown file that names two or three broken features, and failing tests, most often six. Running the tests can turn up failures the description never mentioned. One candidate describes the files as "very much like a real project", with only a few that need changes.
• Some versions ask you to build rather than fix: one candidate added rate limiting and keyword blocking with a user blacklist to a full-stack project, and another built a package-tracking aggregator.
• One candidate (C++ version) advises leaving time at the end to push your code.
• Candidates say you have about a week to take the OA once it arrives; one recruiter allowed a month.
• Not every 2026 OA has the task. Some candidates still got two algorithm questions in 90 minutes, and one SDE II candidate received three OA emails for the same job, one with the AI task and two without; a commenter who did the AI one found the others then showed as completed.
• After the coding come the Work Simulation and Work Style surveys, which have their own page.

The AI assistant

• It has guardrails. It will tell you which file, function or range of lines to look at, but it refuses to write the fix and refuses to answer once your question gets close to the answer. It cannot edit files; one candidate says it can generate code that you paste in yourself.
• Opinions of it vary: "fairly reliable" (on the feature-building version), "average", "basically useless, it just gives hints", "of little use". One candidate on the movie app found it could see only the file you have selected.
• In one report the chat box was broken: the candidate could not type into it, could not reach support, and did the last half hour by hand. They still got the onsite.
• What candidates say worked: paste each failing test and ask which function or lines are wrong; ask it to point you at the API file, then debug in the order of the API calls; copy the error from the terminal and ask which file it comes from.
• One candidate recommends doing the official practice question first, to learn what the assistant will and will not answer. A post describing a Django blog task (create and publish a post through POST /api/posts/, two named tests) was identified in a reply as HackerRank's demo rather than the OA, so expect that one as practice.

Reported repos

• Movie database app (the most reported; see its page): watch lists with six failing tests, a search that finds nothing, recommendations, actor-follow notifications, Django filters.
• Jira-like issue tracker (see its page): comments that do not show after create or update, and issues that cannot be created.
• Loan app. The task named three issues: a user cannot create a loan, cannot view loans already created, and can fund a loan with an insufficient balance. With those fixed, 5 of 6 tests passed; the sixth checked that a user cannot repay a loan with an insufficient balance, which the task never mentioned. The candidate passed the OA.
• Password reset. The verification code was never generated, and a code had to expire 30 seconds after it was generated (the fix: store a generated-at time and check it on verify). After those fixes every test failed, because a test URL file had many variables not found; fixing that file got the tests running. One more failure was not in the task: the new password must be saved to the user's profile. A reply hit tests that could not run because pytest was not installed, and the poster believes debugging the test setup is part of the task.
• Return decisions. One version is an Amazon return system: each customer gets a risk score from how many items they have returned, the items' value, how long the account has been active and so on, and a high score makes an item not returnable. Some items were returnable when they should not have been, and the task was to find six problems. Two C++ versions follow the same theme: a multi-file return routing engine with bugs mainly in parser.cpp and engine.cpp (the candidate fixed the parser and kept hitting syntax errors in the engine, and still moved forward), and a return policy engine where the provided JSON parser was buggy too (the candidate found four or five bugs without passing every test, and passed).
• Review moderation. When a user adds or edits a review, call the existing content-check function; content with bad words gets 403 instead of 201, increments the user's violation count (and saves the user), flags the user after more than three violations, and keeps a flagged user blocked even for clean content; the violated words are stored too. Tests check the status, the count, the flag and the words. The candidate passed 4 of 6 tests and still heard from HR.
• Scheduled payments (Django): six failing tests; the bugs were things like == that should be !=, < 0 that should be <= 0, or a variable that was never updated.
• Bank app: six failing tests (the report with the broken assistant).
• Car rental (Node.js backend): three independent bugs: My Cars shows other users' cars, a newly added car does not appear, and edits are lost on refresh. The poster's suggested approach: check each function against the API documentation, looking at the filter conditions, the insert logic and whether updates persist.
• File and CSV reading (Node.js, from a commenter who passed every test of the repo task): read the file names in a folder with fs and parse a CSV; an unused import in the controller is the one you need, and on the login page you must choose admin login first.
• Other reports: a C++ project run through test.sh, with logic errors, empty unfinished functions and syntax errors; a Django + React project; a Node.js project using Mongoose; a repo whose six tests covered controller validation, persistence and status codes; an object-oriented C++ repo whose six tests each had several problems. The first reporter's advice for Java: look first at the service and controller files, where sorting or status errors sat, and expect a SQL query that needs an extra condition.
• Build-a-feature versions: rate limiting plus keyword blocking and a user blacklist; and a package-tracking aggregator that takes a JSON list of shipments with delivery rules and deadlines and a JSONL stream of carrier scan events, works out each package's current location, decides ON_TIME, LATE or PENDING, and outputs the ledger sorted by tracking id.