CodeDiff

Fast, robust, syntax-aware code diffing.

Try it in the browser: twenty real changes in codediff's own viewer, each shown as Unix diff marks it and as codediff maps it, one click apart.

Installation

cargo install codediff

This command builds CodeDiff from source. You need a C compiler on PATH and a Rust toolchain, edition 2024 or later (rustc 1.88 or later). The build compiles every tree-sitter grammar from C. The first cargo install takes a few minutes, because of this and the lto = "fat" release profile.

Pre-built binaries for Linux, macOS (Intel and Apple Silicon), and Windows are attached to every GitHub release, along with an unofficial .deb for Debian and Ubuntu, a codediff-completions-and-man.tar.gz holding the man page and shell completions, and a SHA256SUMS.txt covering all of them.

On NixOS, or anywhere with Nix installed, no installation step is needed at all:

nix run github:ivankovic/codediff

Recipes for Arch (AUR), Gentoo, Debian and Nix live in packaging/. None of them are submitted to their respective repositories yet - unlike the VS Code extension, which is published; see Editor integration.

Shell completions and the man page

codediff util completions bash   # also: zsh, fish, powershell, elvish
codediff util man

Both print to stdout, generated from the same definition as --help, so they cannot drift from the real flag list. Every package under packaging/ installs them for you; if you installed a pre-built binary by hand, redirect them wherever your shell expects — or source <(codediff util completions bash) for the current session.

The git-history analysis tools in src/bin/ sit behind an off-by-default stats feature. cargo install does not install these tools. They matter only if you build from a checkout. The stats feature exists because these tools need git2 and its own OpenSSL/libssh2 build dependencies. The diffing tool itself does not need these dependencies. Build these tools with cargo build --features stats.

Editor integration

• VS Code - codediff-vscode, v0.0.1. Search for CodeDiff in the Extensions view, or code --install-extension ivankovic.codediff. Also on Open VSX for VSCodium, Cursor and Windsurf. It paints codediff's verdict onto real editors as decorations, and adds five commands: diff two files, diff against HEAD, against any revision, against the last save, and clear the highlights. The binary ships with the extension, per platform.
• Neovim - codediff.nvim.

Both drive codediff --mode json BEFORE AFTER, which prints one JSON object describing each side's changed ranges, their operation (insert/delete/update/move), a move's real counterpart range in the other file, and the nearest enclosing declaration - enough to place highlights on buffers the editor already has open, with no ANSI parsing. The schema is documented at the top of src/tui/json_output.rs. Its columns are byte offsets, which is what tree-sitter reports and what Neovim consumes directly; an editor using UTF-16 (VS Code) or character offsets must convert per line, which is what the VS Code extension does and tests.

Using CodeDiff

The interactive TUI

codediff with no arguments opens a terminal UI with two panels, "Before" and "After". Terminals 220 columns or wider show both panels side by side. Narrower terminals show one panel at a time. codediff BEFORE AFTER, with two file paths, opens directly into their diff, not an empty viewer. If you build from a checkout instead of cargo install, use cargo run -- in place of codediff in every example below.

• Tab — switch the active panel. v cycles the layout between auto, forced dual, and forced single, and codediff remembers the choice across runs.
• o — open a file selector for the active panel. Type to filter the listing; Backspace widens the filter, and only with nothing left to widen does it go to the parent directory. Ctrl-h toggles dotfiles. Once both panels have a file, codediff computes and draws the diff between them automatically, color-coded by change type (inserted, deleted, updated, moved) using the current overlay theme - see c below. On the empty start screen, the digit keys 1-9 reopen a recently diffed pair.
• r — reload both files from disk and re-diff, keeping the cursor where it is. e opens the focused panel's file in $VISUAL/$EDITOR at the cursor line and re-diffs on return - together they close the read-diff/fix-code loop without leaving the session. While a diff is computing, Esc cancels it and keeps the previous result.
• c — open the color theme picker. Built-in themes: Dark (default), Solarized Dark, Solarized Light, Dracula, Nord, Gruvbox Dark, Monokai, One Dark. Moving the selection previews the theme live; Enter keeps it, and codediff remembers your choice across runs.
• ? — show every keybinding plus a color legend rendered in the active theme's actual colors, and copyright/license/repository info. j/k scrolls it. ? or Esc closes it.
• Arrow keys or h/j/k/l — move the cursor, one line or column at a time, same as a text editor. The range under the cursor, and the matching range on the other panel, highlight whenever it's part of a real change; unchanged (identical) content is never highlighted. The other panel's cursor always follows the matched node. Enter jumps to the matched counterpart on the other panel (and back). Both panels carry a line-number gutter, long lines scroll horizontally with the cursor (… marks a cut edge), and a one-column strip at each panel's right edge maps where the changes are in the whole file.
• n/p — jump straight to the next or previous change. This skips unchanged lines entirely. It wraps around at the start or end of the file. g jumps to a line number.
• / — search the focused panel for text (smart-case: all-lowercase matches insensitively, any capital matches exactly). The match count updates live while typing, and matches highlight in the theme's search color. Enter jumps to the nearest match; a bare Enter repeats the last search. Esc cancels; an empty query clears the current search's highlights. >/< step to the next/previous match once a search is active.
• Ctrl-d/Ctrl-u — move the cursor half a page. Ctrl-e/Ctrl-y — scroll the view one line without moving the cursor. Page Up/Page Down/Home/End — scroll.
• S — toggle syntax highlighting.
• Mouse — the wheel scrolls the panel under the pointer; a left click focuses that panel and places the cursor on the clicked character. Terminal-native text selection stays available via your terminal's usual modifier (typically Shift-drag).
• q or Esc — quit. If a dialog is open, Esc cancels the dialog instead; while a diff is computing, Esc cancels the computation.

The web UI

codediff-web is the same viewer served to a browser tab. It is a second binary behind the off-by-default web feature:

cargo install codediff --features web
codediff-web BEFORE AFTER

It starts a local server, prints the URL, and opens your browser (--no-open to skip that, --port and --host to pick where it listens - the default is a random port on 127.0.0.1). With no arguments it starts empty, like the TUI, and it accepts git's GIT_EXTERNAL_DIFF argument list too. Every key from the list above works in the page, ? included, and the two front ends share one config file: a theme, layout or render option chosen in one is what the other starts with.

Three things differ because a browser is not a terminal. e opens $VISUAL/$EDITOR in the terminal codediff-web was started from (a GUI editor opens wherever it opens), and the page re-diffs when it exits. q stops the server, the way it quits the TUI; closing the tab does not, so a reload does not end the session. Ctrl-Z has no counterpart. Only the machine running codediff-web can use the page: it binds a loopback address, and every request has to carry a token the page got when it loaded, so no other site open in the same browser can read a diff or drive the session. Listening on anything but a loopback address exposes your files to whoever can reach it.

Headless / batch mode

codediff --headless BEFORE AFTER, or its synonym --batch, prints the diff as plain text, with optional color, instead of opening the TUI. Use this for scripts, CI, or any case where stdout is not a real terminal. Headless mode also starts automatically whenever stdout is not a terminal, for example when piped into less or redirected to a file. Because of this, codediff BEFORE AFTER | less works without the flag.

Every printed line is prefixed with its line number, so the moved-chunk headers' "Moved to lines 40-60" cross-references can actually be followed. codediff collapses long runs of unchanged lines. It keeps 3 lines of context on each side of a change (override with --context N), the same convention as diff -u. codediff also prefixes each hunk with the nearest enclosing function, class, or struct line, when that line is not otherwise visible. This shows the location of a change deep inside a large file. The reader does not need to see the whole file around it.

Colors are on by default (git's pager renders them); pass --color never, or set NO_COLOR=1 (see https://no-color.org), to disable ANSI colors, for example when you redirect output to a file - --color always forces them even under NO_COLOR.

codediff exits 0 on success and 2 on error. For scripting, pass --exit-code to additionally get 1 when the files differ, the diff(1) convention. That is opt-in rather than the default for the same reason git diff exits 0 even when files differ: codediff's usual non-interactive callers are version control systems driving it as a display tool, and they read a non-zero exit as "the tool failed" - jj warns on every file, and git difftool with difftool.trustExitCode=true aborts the whole diff. (The 7-argument GIT_EXTERNAL_DIFF form stays at 0 even with --exit-code, since git treats a non-zero exit there as fatal.)

Git integration

codediff doubles as a git difftool backend. Run the interactive setup wizard, which asks whether to configure it globally or for the current repository only:

codediff git configure

Or configure it by hand:

git config difftool.codediff.cmd 'codediff "$LOCAL" "$REMOTE"'
git difftool --tool=codediff

Run git config diff.tool codediff to make plain git difftool use codediff by default, without needing --tool. If you do not want git to ask "view diff ... [Y/n]?" before every file, run git config difftool.prompt false.

git difftool opens the interactive TUI. git diff and git log -p never do. git diff pipes its output through git's pager, and a full-screen TUI cannot draw onto a pipe, so codediff always falls back to plain text there regardless of terminal or GIT_EXTERNAL_DIFF config (see "Headless / batch mode" above). If you want the interactive viewer from git, use git difftool, not git diff. If git difftool still doesn't open interactively over SSH, reconnect with ssh -t — the session needs an allocated pseudo-terminal; tmux panes always have one, so tmux itself is never the cause.

codediff also works directly with git diff and git log -p, through GIT_EXTERNAL_DIFF. This path needs no difftool config:

GIT_EXTERNAL_DIFF=codediff git diff

Files with no tree-sitter grammar (an unrecognized extension, or none at all - Makefile, Dockerfile, ...) fall back to a plain line-level diff instead of the syntax-aware one, so a change touching one of them never blocks git diff from showing the rest.

Binary files - anything codediff cannot read as text, a PDF or an image - get a one-line Binary file <path> differs notice instead of a diff, the same stand-in git and diff(1) print for them. They likewise never block the rest of a git diff: an external diff that exits non-zero makes git abandon the entire run, so codediff reports an unshowable file as a successful diff of nothing rather than as a failure.

Jujutsu (jj) integration

jj does not read git's difftool/diff.external settings, even in a colocated repo, so it needs its own configuration. Run the setup wizard:

codediff jj configure

Or configure it by hand:

jj config set --user merge-tools.codediff.program codediff
jj config set --user merge-tools.codediff.diff-args '["$left","$right"]'
jj config set --user merge-tools.codediff.diff-invocation-mode file-by-file

That registers jj diff --tool codediff. To make it the default for plain jj diff as well:

jj config set --user ui.diff-formatter codediff

Use --repo in place of --user to configure the current repository only.

diff-invocation-mode = "file-by-file" is required. jj's default hands a diff tool two directory trees; codediff diffs two files, so without this setting every invocation fails. With it, jj passes one changed file pair at a time, keeping each file's real path and extension, so language detection works exactly as it does under git.

jj diff runs its formatter under a pager, so codediff renders in its non-interactive text mode there - the same output git diff gets. jj has no equivalent of git difftool's interactive per-file viewer (its terminal-attached hook, ui.diff-editor, is for jj diffedit/jj split, which edit the right-hand side and read it back - not something a read-only viewer should claim to do), so for the full-screen TUI on a jj repo, run codediff BEFORE AFTER directly.

Guiding principles

Robust

CodeDiff must process 100% of all commits in the full test dataset.

The full test dataset holds the git commit history of about 7,400 open-source git repositories, as available on the main branch. This list of repositories comes from the Gentoo Linux distribution. Find it in list_of_repositories.csv.

Measured once in full, on 2026-09-19: every modified code file in the most recent 50 commits of each of those repositories, 442,530 readable before/after pairs in 25 languages, diffed with no size cap under a 120-second budget and a 6 GB memory cap per process. 442,322 completed (99.95%), with no panics. The largest completed pair holds 7.1 million AST nodes on each side; the corpus's largest file has 23.6 million, and the files of that order the run met - all generated tables - needed more memory than the cap. So the goal is not met at the very top, and the paper says so. 96 pairs ran past the budget and 112 past the memory cap; the latter are twenty generated or embedded files - tree-sitter parser tables, codegen, minified bundles, a PNG as a C array - plus one commit of a 40,000-line single-header C++ library. Given 24 GB, eleven of those files complete. The run, its harness and every pair that did not complete are documented in research/data/performance/PROVENANCE.md.

A smaller list of 100 repositories, the "small" dataset, is available in the same directory. Use it for faster iteration when you debug.

Fast

• A median diff in 100ms or less.
• A 99th-percentile diff in 1000ms or less.

Both are met. Over the 2,001 fixtures in src/test/data/diffs/, measured 2026-09-18 on a release build: p50 7.6ms, p90 78.7ms, p99 347ms, slowest 1,355ms. 100ms is the 92.7th percentile — 146 of 2,001 fixtures take longer than that, and 2 take longer than a second.

Two things that number is not. It is not the full dataset the Robust goal above names: a diff of every commit in 7,400 repositories is not something this project runs per change, and the fixture corpus is the proxy it uses instead. And the proxy is a pessimistic one, because it is grown deliberately toward hard cases. The one full-corpus run above says how pessimistic: over 442,322 real file modifications the median is 12.2ms and the 99th percentile 2.1s, so the first goal holds on real commits with room to spare and the second does not, on the roughly 1% of files that are generated tables and bundles rather than code anyone edits by hand.

In code, I accept less readable, more complex code, if that code is faster.

Benchmarks make sure that performance does not regress. make benchmark-quality prints the distribution above; make check-quality compares it against the committed baseline on every push.

Accurate

CodeDiff must match a human's own reading of a change, measured against the hand-authored ground-truth mappings in src/test/data/diffs/:

• 90% of test cases with zero mismatched visible nodes.
• 99% of test cases with at most 1% of visible nodes mismatched.

Visible is the load-bearing word. Most AST nodes are structure the reader never sees on their own - a block, an argument_list, a declaration_list, whose every readable byte belongs to some descendant. Getting one of those wrong does not put anything wrong in front of the reader, so counting it alongside a wrongly-matched identifier overstates how wrong the diff is.

A node is visible when it carries text of its own: a leaf, or an interior node with non-whitespace content its children don't cover (a comment whose // marker is a separate child, say). This is a property of the syntax tree and the source bytes and of nothing else - codediff::diff::nodes::is_structurally_visible is the definition. In particular it does not depend on the diff: a measurement whose own denominator moves when the algorithm changes cannot be used to judge the algorithm. Corpus-wide about 68% of nodes are visible.

make benchmark-quality reports both the raw and the visible mismatch count per fixture.

License

Copyright (C) 2026 Marko Ivankovic

This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation.

See the LICENSE file for the full text of the License.

Cannot use AGPL software?

Contact me for options.

AI policy

This project uses substantial AI assistance, currently Claude Code. Most commits disclose this with a Co-Authored-By trailer and a link to the session that produced them. This project does not hide that fact. This project does not treat AI assistance as a lesser way to write software.

AI-assisted contributions are welcome. Use whatever tools help you do good work. Disclose your use of these tools the same way. You are still responsible for understanding and standing behind whatever you submit.

For Developers, human or otherwise

See CONTRIBUTING.md for the technology overview, code-quality and testing expectations, project structure, and what CI checks on every push and PR. AGENTS.md has additional AI-agent-specific conventions.