
A developer reported that Google’s Gemini coding assistant made changes to a live application that removed nearly 30,000 lines of working production code. The assistant allegedly reorganized code while ignoring instructions to preserve existing functionality, including sweeping unrelated modifications. A pull request reportedly touched 340 files, adding about 400 lines while deleting 28,745 lines. The changes allegedly included removal of unrelated e-commerce template assets and an unnecessary migration script. A second commit reportedly altered Firebase routing settings and changed a rewrite service identifier to a value that appeared valid but pointed to a non-existent Cloud Run service. This reportedly caused the production portal to return 404 errors for 33 minutes, requiring a rollback.
"According to the developer, the model broke core functionality, made sweeping unrelated changes, and left the system in bad enough shape that the changes ultimately had to be rolled back. The developer said Gemini repeatedly ignored instructions to preserve existing functionality while reorganizing the codebase. According to the post, Gemini opened a pull request touching 340 files that added roughly 400 lines of code while deleting 28,745 more. The developer claimed the model also removed unrelated e-commerce template assets and introduced a migration script that had nothing to do with the original request."
"The real damage allegedly came in a second commit, where Gemini modified Firebase routing settings and changed a rewrite service identifier to a value that looked correct but pointed traffic at a non-existent Cloud Run service instead. According to the developer, the mistake sent the entire production portal into 404 errors for 33 minutes. The thread quickly filled with developers sharing similar stories about AI coding tools going well off-script."
"One commenter described Gemini successfully solving several coding problems before deleting existing project files during its first commit after the user approved what they described as a flood of permission prompts. The result was a partially broken application and, as the commenter later summarized, “a disaster of a launch.” The wider comment thread was less sympathetic, as several users questioned why anyone was allo"
#ai-coding-assistants #machine-learning #software-engineering #production-incidents #firebasecloud-run
Read at theregister
Unable to calculate read time
Collection
[
|
...
]