git http-fetch

Object-by-object download over plain HTTP.

🎯 What & why

git http-fetch walks a remote repository over plain HTTP, downloading objects one by one starting from a given commit. It exists for the ancient dumb-HTTP transport, where the server is just a static file host with no Git CGI.

🧠 Mental model

Smart HTTP negotiates a single packfile for the whole transfer; dumb HTTP cannot. http-fetch instead does what an old web spider would: GET an object, parse it, GET its referenced objects, repeat. Many round trips, no negotiation.

🛠️ Synopsis

git http-fetch [-c] [-t] [-a] [-v] [-w <filename>]
               [--recover] [--stdin] <commit> <url>

Pulls the commit's reachable history from <url> into the current repo's object store. With `--stdin`, reads `<sha> <ref>` lines instead.

🎚️ Switches & options

Flag	What it does
`<commit> <url>`	Starting commit SHA and the remote base URL (no trailing `info/refs`).
`-c`	Also fetch commit history (default for normal use; explicit here).
`-t`	Also fetch tags reachable from the walked commits.
`-a`	Equivalent to `-c -t`; fetch commits and tags.
`-v`	Verbose -- print each object SHA as it lands on disk.
`-w <filename>`	Write the resolved tip SHA into `<filename>` after success.
`--recover`	Resume an interrupted fetch by filling missing objects.

💡 Use cases

Pulling from a legacy intranet mirror that only exposes static files.
Bootstrapping a clone from a USB stick or CD-ROM served over a tiny HTTP daemon.
Salvaging objects from a partially-mirrored repo with --recover.
Diagnosing dumb-HTTP transport problems in isolation from porcelain.

🧪 Examples

Fetch one commit's history

git http-fetch -a abcd1234 http://127.0.0.1/repo.git/

Verbose with tip recorded

git http-fetch -v -w FETCH_HEAD abcd1234 http://127.0.0.1/repo.git/

Resume an interrupted pull

git http-fetch --recover abcd1234 http://127.0.0.1/repo.git/

Read SHA/ref pairs from stdin

echo 'abcd1234 refs/heads/main' | git http-fetch --stdin http://127.0.0.1/repo.git/

🎓 Recommendations

Avoid this in modern setups -- use smart HTTP (git fetch against a http-backend CGI) instead.
If forced into dumb HTTP, ensure the server has run update-server-info recently.
Pass -v while diagnosing; the per-object output makes round-trip pain visible.

🪤 Common pitfalls

Object-by-object download is dramatically slower than smart-protocol pack negotiation.
Stale info/refs on the remote leads to 'object not found' errors mid-walk.
Without -a, tags reachable from the commit may be skipped, leaving annotated tags missing.

🔗 Related modules

📝 Quiz

Hit each option, then Check answers. Score is recorded; Next is always open.