# OmniDiff 0.3.0: Rust Diff Tool That Understands Your Code > Source: > Published: 2026-11-05 01:07:47+00:00 Your AI coding assistant just rewrote three functions, added a formatting pass, and touched 15 files. `git diff` shows 400 changed lines. None of them tell you what actually changed. **OmniDiff 0.3.0**, released this week, is a Rust-based `git difftool` replacement that diffs code by its syntax tree rather than line count — and version 0.3.0 adds the feature developers have been asking for: embedded language support. ## How It Works OmniDiff parses both versions of a file with [tree-sitter](https://tree-sitter.github.io/tree-sitter/) (25 languages supported) and runs the APTED tree-diff algorithm on the resulting ASTs. The result: a reformatted function shows as “moved,” not deleted and re-added. A variable rename shows as an update, not 40 changed lines. When a file has an unknown extension, OmniDiff falls back to plain line diffs — so it never breaks your workflow. The accuracy difference versus legacy tools is not marginal. Against a labeled corpus of real-world diffs, OmniDiff gets 90% of changes completely correct. [Difftastic](https://difftastic.wilfred.me.uk/), the previous best-in-class structural diff tool, lands at 64% on the same benchmark. OmniDiff is also faster at worst-case: its slowest file clocks 803ms vs Difftastic’s 2000ms+. Median diff time is 4.2ms. ## What’s New in 0.3.0: Embedded Language Diffing The 0.3.0 release ships one major new capability: OmniDiff now detects when code of a different language is embedded inside a string and diffs each layer separately. Concretely: - Shell inside YAML `run:` and`script:` blocks (GitHub Actions, GitLab CI, Ansible) is diffed as shell - CSS and HTML inside template literals (styled-components, css`` `...` `` , html`` `...` `` ) are diffed as CSS/HTML - TSX inside JSON `"content"` fields is diffed as TSX - Tailwind/clsx class lists in JSX `className` ,`cn()` ,`cva()` ,`twMerge()` — diffed class by class, not as a changed string blob If you work with GitHub Actions, Ansible playbooks, or any styled-components codebase, this is a direct quality-of-life upgrade. Version 0.3.0 also ships a 22% performance improvement on average and 22–43% improvement on large files. ## Binary Files, Finally Standard `git diff` on a binary file prints “Binary files differ” and stops. OmniDiff does not. It diffs: - **Archives** (zip, jar, tar, gz, xz, bz2) — member by member, with Java .class files decompiled to listings - **Fonts** (TTF, OTF, WOFF, WOFF2, EOT) — glyph by glyph - **PDFs** — page by page, with changed region descriptions - **Images** — pixel-aware diff with a change description Design tokens, bundled fonts, exported PDFs — these live in git repositories. Knowing what actually changed in a font file or a bundled archive matters, and OmniDiff is currently the only `git difftool` that provides it. ## Installing and Configuring OmniDiff installs from crates.io: ``` cargo install --locked omnidiff ``` Then run the interactive setup wizard, which asks whether to configure globally or for the current repo only: ``` omnidiff git configure ``` To use OmniDiff with `git diff` and `git log -p` (not just `git difftool`): ``` GIT_EXTERNAL_DIFF=omnidiff git diff GIT_EXTERNAL_DIFF=omnidiff git log -p ``` A [VS Code plugin](https://marketplace.visualstudio.com/items?itemName=ivankovic.omnidiff) is also available for in-editor diffs. ## Who Should Switch If you are reviewing pull requests that contain AI-generated code, the answer is: everyone. AI coding assistants produce reformats, moves, and rewrites that trip line-based diffs into showing false positives. OmniDiff’s structural approach cuts that noise. It is also the right tool for repos with binary assets — font files, bundled JARs, exported PDFs — where “Binary files differ” tells you nothing. The project is [open source on GitHub](https://github.com/ivankovic/omnidiff) under AGPLv3, with a showcase at [ivankovic.github.io/omnidiff/showcase](https://ivankovic.github.io/omnidiff/showcase/). The [Hacker News thread](https://news.ycombinator.com/item?id=50325253) from the Show HN post has 316 points and is worth reading for edge cases and comparisons.