Show HN: CodeDiff – Fast (<100ms), robust (99.95%) syntax-aware code diff

Developer ivankovic released CodeDiff v0.1.1, an open-source syntax-aware code diff tool written in Rust that parses 24 languages with TreeSitter and falls back to a modified Myers line diffing algorithm for unparseable files. Tested on the last 50 commits of roughly 7,400 Gentoo package-list repositories, CodeDiff diffed 99.95% of file pairs in 120 seconds using 6GB of RAM on an 11-year-old Intel E3-1275 v5, with a median speed of 7ms and 93% of files diffed in under 100ms. On a hand-solved dataset of 1,400 diffs, CodeDiff was perfect on 70% and near-perfect on 90%, and perfect 89% of the time when measuring only the AST diff; the tool is licensed AGPL3.0 and targets p99 under 400ms and 99% perfect diffs across 30 languages for v1.0.

Hi folks Because of AI, there is more code to review. I don't know about you, but I get annoyed when the default git diff doesn't match something "obvious". I built CodeDiff to help. It's a syntax-aware, careful implementation in Rust of all current research in code diffing. It parses 24 languages with TreeSitter and falls back to a modified Myers line diffing algorithm for any file it cannot parse. I wouldn't necessarily call it a novel algorithm, but it isn't a re-implementation of existing tools either. I tested it on the last 50 commits of all 7400-or-so repositories that are in the Gentoo package list. CodeDiff can diff 99.95% of those file pairs in 120s and 6GB of RAM on a rusty 11 year old Intel E3-1275 v5. The median speed is 7ms and 93% of files are diffed in under 100ms. I also did a bit of hand solving and have a test dataset of 1400 hand solved diffs. CodeDiff is perfect on 70% and near-perfect on 90% of them. But accuracy is actually a complex thing to measure, because if you are only after the AST diff, then it is perfect 89% of the time. Deciding what to paint however is not an objective thing so if you do try it out, check out the render options press M and make it fit your style. If you just want to see how it looks like: https://ivankovic.github.io/codediff/showcase/?case=python-r... https://ivankovic.github.io/codediff/showcase/?case=python-refactoring&view=codediff If you want to install it: https://github.com/ivankovic/codediff https://github.com/ivankovic/codediff Do feel free to file issues. It's v0.1.1 at the moment, but I have been using it for months as my git difftool and I plan to keep that stable. Using it as a library directly from Rust however, might change until v1.0. The final goal for v1.0 is p99 < 400ms and 99% perfect diff, including multi-file diff support in 30 languages. Credit where credit is due: There is a bunch of people researching diffs in the past 20 or so years. Without those papers, CodeDiff would be a lot harder to build. Especially without APTED and Myers algorithms. License: AGPL3.0, I kind of like copy-left. If your company prevents use of AGPL software tell them to throw a few bucks my way for my token usage and we can talk. AI disclosure: I use AI assisted coding. Claude and Mistral were used, heavily. Comments URL: https://news.ycombinator.com/item?id=49891989 https://news.ycombinator.com/item?id=49891989 Points: 1 Comments: 0