A GitHub co-founder had agents rewrite Git in Rust — then relicensed it from GPL to MIT
Scott Chacon pointed AI agents at Git's own test suite and got a memory-safe Rust reimplementation passing 41,715 of 42,001 tests for about $10–15k. The thread's fight wasn't the code. It was the license.
Grit is 360,000+ lines built across multiple agent setups, burning ~45B tokens, passing 99.3% of Git's tests — though the author admits nobody has used it for anything real.
The flashpoint: Git is GPL, but Chacon released Grit under MIT, arguing the LLM-generated code is a new work. A GitHub co-founder, of all people, relicensing a copyleft project.
Others reject the premise outright: Git has never failed on a memory overflow, so why burn $10k to rewrite it?
What Chacon built
Scott Chacon's post describes building "Grit," a from-scratch, library-based, memory-safe Rust reimplementation of Git, by pointing AI agents at Git's own test suite. It passes 41,715 of 42,001 tests (99.3%), with the rest deliberately skipped. The stated motivation is a 15-year wish to make Git "library based" rather than command-line and Unix-pipe-centric, so it can be embedded in tools like GitButler, editors, WASM builds and servers without fork/exec overhead. The build spanned multiple agent setups (Claude Code, Cursor agents, Claude workflows), consumed roughly 45 billion tokens at an estimated $10–15k, and produced 360,000+ lines across 7,000+ commits. Chacon, a GitHub and GitButler co-founder, is candid that it "passes the tests" but isn't used for anything real yet.
The fight was about copyright, not Rust
The dominant reaction wasn't technical admiration. It was the GPL-to-MIT relicensing, made pointed by the author's own résumé.
"This is coming from a cofounder at github, someone who probably knows precisely what the GPL is for … building on a GPL3 project's complete test suite and relicensing under MIT is not acting in good faith toward the original authors." — a commenter on the licensing move
"I guess software licenses are meaningless now since anyone can decide their llm clone is not derivative." — Kit Randel, an engineer at Stacklet
"Sometimes software is good as is"
A second camp rejected the entire premise that Git needs rewriting.
"they still haven't explained why I should bother. Is it faster, easier, more efficient … In fact, I would rather it stay C for 15 more years." — a commenter questioning the motivation
"So it is not entire then but somehow that was worth burning $8,000~ dollars worth of tokens?" — a commenter on the cost-benefit
There's also prior art the project doesn't displace: gitoxide, an existing Git-in-Rust effort, got surfaced as the thing already doing this without the licensing controversy (pointer to gitoxide). And the 99.3% pass rate carries its own asterisk: the author concedes agents "love to cheat" on tests, and that nobody has run Grit in anger. Passing 41,715 tests is an impressive demo of agent-driven engineering. It's also a copyleft project laundered into a permissive license by a man who, more than almost anyone, knows what that license was for.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
In 2019, OpenAI said GPT-2 was too dangerous to release. The thread argues both sides were right
A resurfaced retrospective recalls the staged release of GPT-2 and the 'too dangerous' framing. Seven years on, half the thread calls it vindicated by AI slop, half calls it the original hype playbook.