git v2.47.0 is a large C codebase with roughly 390,000 lines, 124 builtin commands, and behavior spanning object storage, refs, index state, patch formats, transports, lockfiles, and packfiles.
The agent has read-only access to the full C source tree as reference, but must write its implementation entirely in Zig.
We asked agents to rebuild git v2.47.0 in Zig.
The agent is given Git's source code with tests stripped. It can read the full C implementation to understand repository behavior, but cannot see the verifier tests and must infer correct behavior from the source and documentation alone.
The verifier builds a native binary from the Zig code and runs Git's own integration test suite against it. The test suite contains 29,075 assertions across 1,007 scripts, covering checkout, diff, merge, rebase, submodules, and more. If the agents code doesn't compile, we give it a score of 0.
There is no reward for architectural progress unless it shows up as real passing Git assertions.
The task runs in a Modal container with 4 CPUs, 16 GB RAM, and no internet access. The container image stages /app/zig-port,/app/git-src, and a reference git binary for local experiments.
The environment intentionally omits gcc, so the agent can compare behavior locally but cannot just compile the C source or wrap it into Zig.
/app/git-src/ before zig build, so the submission cannot compile or link the bundled C source at build time.git before tests and uses strace plus ELF checks to reject wrapper binaries that exec the real git.