> Copilot keeps hallucinating on our legacy codebase and its driving me crazy. Honestly, if safety is your main concern, Cursor AI Code Editor is probably the most reliable pick right now. It indexes your entire local project much better than the standard extensions, which helps it understand how your legacy functions actually interact across different files. Its been way more stable for me when doing heavy refactoring under pressure. You might also want to look into Sourcegraph Cody AI Pro for its context-awareness features. It specifically searches your whole repo before suggesting changes so it doesnt just make things up based on generic training data.

• index your codebase fully before starting any work
• feed it specific relevant files to limit the search space
• run unit tests after every single change to verify logic It wont solve everything instantly but it definitely cuts down on those wild hallucinations when youre dealing with messy codebases.

Saw this post while grabbing coffee earlier but yeah, those hallucinations are a nightmare on tight deadlines. If youre dealing with complex junk like you said, you might want to look into Amazon Q Developer Professional Tier. Its basically designed for these enterprise messes and it has a feature specifically for code transformation which automates upgrading legacy code. This helps with dependency mapping way better than chat prompts imo. Total lifesaver. Another one people sleep on is Tabnine AI Code Assistant Pro Edition. Training this on your local repo usually cuts down the nonsense hallucinations because it learns your specific patterns. Try running their local-only model if youre in a pinch with that 2-week sprint. It keeps the context window tight on your actual code rather than guessing based on random open-source stuff. Both are solid choices for getting through that Seattle crunch.