AI keeps changing code without fixing the bug: how do you establish a minimal failing path?

Reduce the user action to a repeatable failing input

This applies when one edit temporarily removes an error but another input fails or an agent experiments across unrelated files. Preserve working-tree state and separate existing user changes from current attempts. Record the stack trace, request path and actual input, retaining the structure that triggers the same failure after removing private information.

Confirm the failure location and the version actually running

Follow the actual execution chain to confirm entry point, dependency versions and failure location, including stale builds or programs running from another directory. Reproduce with a small sample and state expected versus actual behavior. For external-service failures, preserve a safe response sample or representative substitute; a stub that always succeeds does not locate the original problem.

Fix the identified path and inspect the diff

Make the smallest change supported by the identified cause and review changed files and effects after each attempt. Avoid mixing unrelated refactoring and dependency upgrades. Add an assertion for the original error where useful, with diagnostic input in its failure message. Use reversible experiments for uncertain causes and retain evidence rather than accumulating ineffective experiments in the final patch.

Show the original sample failing before repair and passing afterward

Confirm the reproduction fails before the fix and meets expectations afterward, then test adjacent boundaries and established behavior. Interface and API bugs need their actual path checked rather than just an isolated function. Report executed commands and outcomes. If an external environment is unavailable, describe the remaining scope instead of claiming all tests passed.

This article did not run your project or claim a successful repair; it describes evidence-based debugging. Minimization must preserve the conditions that actually cause the bug, and concurrency failures may require timing traces. Passing tests cannot prove every input correct, but reproducible evidence supports a fix. Do not delete failing assertions to obtain green results.