GitCraft

Objects make the repository

Loose objects are encoded with Git’s type-and-length header, addressed by SHA-1, compressed with zlib, and stored under their derived path. Recursive tree construction turns filesystem entries into an object graph; commit creation then points to a tree and its parent history.

The implemented local commands are init, cat-file, hash-object, ls-tree, write-tree, and commit-tree. That boundary is visible because compatibility is more useful when its limits are explicit.

Clone is a binary data pipeline

The clone path discovers references and requests objects through Git’s smart-HTTP upload-pack flow. The response is parsed as a packfile rather than treated as an archive. Base objects are decoded first, then offset and reference deltas are applied to reconstruct their final bytes.

Binary lengths, variable-width encodings, and delta instructions demand defensive cursor management. Rust’s result types and bounded parsing help keep malformed input from becoming unchecked indexing or partial output.

Materialization comes last

Reconstructed objects are written into the local object database before the selected tree is walked into the working directory. Keeping network exchange, object reconstruction, storage, and checkout as distinct stages made the end-to-end flow easier to reason about.

Clippy passes with warnings denied, but the current test command reports zero tests. That is a real limitation: the next useful work is a compatibility matrix and deterministic fixtures for hashes, packfiles, deltas, and checkout—not a broader claim about replacing Git.