A reviewer that remembers every bug your codebase ever had
Most AI code review tools read the diff in front of them and forget it exists the moment the review posts. Bug Genealogy is bugweg's answer to a different question: has this exact pattern shown up before, and did the last fix actually stick?
How it actually works
Every issue bugweg flags is embedded with Voyage's voyage-code-3 model — tuned for code retrieval, 1024 dimensions — and checked against bug_fingerprints: every pattern this specific repository has ever had flagged, going back to the first PR bugweg reviewed on it. Above a 0.82 cosine-similarity threshold, the PR comment tells you this exact pattern has appeared before, exactly how many times, and links every prior fix — not a vague "this looks similar," a concrete count and a trail.
This is real and live on every PR that reaches bugweg's LLM review stage — not a roadmap item, not a simplified version of a bigger planned feature. See the full, unvarnished status of every bugweg feature (including which ones are still simplified or planned) on Features.
Why this matters more than another lint pass
A team that keeps re-fixing the same class of bug isn't missing a linter rule — it's missing institutional memory. Reviewers rotate, PRs get approved on a tight deadline, and the same null-check gets skipped a second time six months after the first incident. Bug Genealogy is the mechanism that closes that gap automatically, without anyone having to remember to write a custom lint rule for last quarter's incident.
Paired with the Recurrence Firewall, a match above a repo's configured threshold can actually block the merge via bugweg's GitHub Check Run — not just leave a comment that gets scrolled past.
How it's different from a full graph index
Some tools, such as Greptile, build a full graph index of your entire repository. bugweg deliberately doesn't — it only ever holds the changed lines of a PR in memory, embeds the bug-level fingerprint, and discards the raw source. That's a narrower scope by design: see the full trade-off comparison on Compare, including where whole-repo indexing is genuinely the better fit.
Code Risk Prediction applies the same memory at the file level. See how this fits into the full review pipeline on How it works, or install bugweg on a repo to see it flag a real recurrence.