OmniDiff is a command-line diff tool that does what Unix diff has refused to do for fifty years: look inside the file before deciding how to compare it. For code, it parses 25 languages with tree-sitter grammars and diffs structurally—matching moved functions and renamed variables rather than treating every line as an opaque string. For binaries, it identifies the format from the bytes (not the extension) and diffs by content: pixel-by-pixel for images, file-by-file for archives, glyph-by-glyph for fonts. The accuracy numbers, self-reported against a judged test set, are striking. Code diffs hit 98% accuracy, pictures 96%, cursors 100%, fonts 95%, archives 88%. The comparison columns matter more: Unix diff scores 70% on code, 69% on pictures, 82% on fonts. The gap is real and consistent across content types, with the notable exception of PDFs, where OmniDiff's accuracy drops to 57% (though with fewer than ten test cases, the author flags this honestly). Speed is the other headline. All benchmarks are under 100ms on a mid-range ThinkPad T14s with an i7-1370P. This is cargo-install-from-source software written in Rust with fat LTO, and the first build takes minutes because it compiles every tree-sitter grammar from C. After that, it's fast enough to be invisible in a git difftool workflow. The integration story is already mature for a tool at this stage. Git and Jujutsu get first-class setup wizards. VS Code and Neovim have extensions (with the VS Code extension waiting on an upstream API to replace the built-in diff view rather than bolting on a workaround). Installation covers Homebrew, apt with signed repos, Nix flakes, Arch PKGBUILD, and Gentoo ebuilds alongside prebuilt binaries for Linux, macOS (both architectures), and Windows. The design philosophy is visible in the small decisions. Headless mode activates automatically when stdout isn't a terminal. Context lines default to 3 with function-header annotations, matching the diff -u convention developers already expect. Exit codes follow the diff(1) convention with an opt-in --exit-code flag, and the README explains exactly why the default is 0 even when files differ (git difftool and jj both misinterpret non-zero exits). JSON output has a published schema. What's missing is equally telling. The codereview browser viewer is at v0.0.0 with an explicit use-at-your-own-risk warning. The VS Code extension can't replace the built-in diff view yet. Six languages are recognized by extension but diffed as plain text because no grammar is compiled in. These are honest gaps, not hidden ones—the kind of scoping that suggests the developer knows what 1.0 means and isn't there yet. This is a solo developer's tool built with the kind of engineering discipline you usually see from small teams that have shipped before. The README alone—its precision about terminal color detection, SSH pseudo-terminals, git's GITEXTERNALDIFF 7-argument form—reads like someone who has debugged every edge case they document. If you review code at a terminal, this is worth fifteen minutes of your time to install and try.
