Okay, I have to get this off my chest. I've been a daily Codeium user for months, especially in my CI/CD and Go/Python workflows. Since the last major model update a couple weeks ago, I've noticed a significant drop in quality.
It's not just about getting a "dumber" answer. The issues are more specific:
* **Context handling seems broken.** It'll reference a function I defined 10 lines earlier, but then hallucinate its parameters or return type. This is brutal for refactoring.
* **Increased "confidently wrong" answers.** It suggests a non-existent GitHub Actions property or a deprecated Kubernetes `kubectl` flag, and it writes it like it's gospel.
* **Autocomplete has gotten *weird*.** It'll aggressively suggest completions that are syntactically correct but completely illogical for the context, like inserting a `log.Fatal` right in the middle of a loop for no reason.
I went from relying on it for boilerplate and spotting my typos to having to double-check every single suggestion. The latency is better, but the accuracy is way down.
Has anyone else hit this, or is it just my setup/languages? I'm trying to pinpoint if it's a broader regression or something specific to the stack I'm using. Really hoping this is a temporary blip, because it was a fantastic tool before this.
Ship fast, measure faster.