Show HN: VibeDrift – Measure drift in AI-generated codebases

Cool project, but not really. From what I can see Allium is preventive, it gives the AI a spec to code against so intent doesn’t get lost. VibeDrift is diagnostic, it analyzes code that already exists and measures where patterns diverged. They’re actually complementary.

That’s an interesting angle. VibeDrift currently focuses on the artifact i.e. the code that actually lands in your repo rather than the LLM output itself. The reasoning is that regardless of why the drift happened (different sessions, different prompts, different models), the codebase is what you ship and maintain. That said, tracking prompt-to-output consistency is a genuinely different problem and not something I’ve explored yet. Would be curious what patterns you’ve seen there. I’m always open to suggestions and feedbacks. There is always room for improvement

Claude Code memory helps the AI remember context within its own sessions. But it’s still one model’s view of what “should” be consistent. VibeDrift doesn’t rely on any AI’s memory or opinion. It looks at the code that actually exists in your repo and measures what the majority of files do vs which files break from that pattern. It’s also deterministic, meaning same codebase, same score every time, which matters if you want to track drift over time or gate PRs in CI.

CodeRabbit reviews one PR at a time using context from past PRs. VibeDrift scans the entire codebase at once and compares every file against its directory peers. Different question: not “does this PR look good” but “does this file follow the same patterns as the files sitting next to it.” Also runs fully locally, zero data sent. Curious what your PoC does though.

This is for multi file codebases written across multiple sessions/prompts.

Prompts help guide generation, but they don’t guarantee consistency over time.

VibeDrift checks the codebase itself and flags where files contradict each other.

Probably overkill for a weekend project, but shows up fast as things grow.

I’d suggest you give it a try and let me know your feedback.

Qodo reviews individual PRs. VibeDrift compares files against each other across your whole codebase. Qodo won’t tell you that the file you’re adding uses raw SQL while 7 sibling files use a repository pattern, it’s looking at the PR in isolation, not the project. If that makes sense.