How to commit without an index file?

In my program, I have the file paths, permissions and Git object IDs of everything that I want to appear in the next commit. Hence, I want to create a commit without creating an index file.

Essentially it's like to combining `--cacheinfo` from `update-index` with `commit` like:

	git commit \
		-m "Commit message" \
		--branch my-branch \
		--cacheinfo "100644,0123456789abcdef0123456789abcdef01234567,path/to/first/file” \
		--cacheinfo "100755,123456789abcdef0123456789abcdef012345670,path/to/second/file” \
		--cacheinfo "100644,23456789abcdef0123456789abcdef0123456701,path/to/third/file"

In this example, the command would not read the index file, but would get the information necessary to create the commit tree as arguments (or from STDIN). The command would handle recursively creating trees for the intermediate directories. The object ID for the final tree would be used for the `tree` field of the commit.

I tried tricks for simulating standard streams as an index file to Git, but they didn't work out:

	$ GIT_INDEX_FILE=/dev/stdout git read-tree HEAD
	fatal: Unable to create '/dev/stdout.lock': No such file or directory

The best I could come up with is creating a no checkout worktree for `my-branch` and using a FIFO for the index file, instead of a conventional file. The shell script equivalent of it would be as follows (I would be happy if you can point out any errors in this script as well):

	set -e
	
	git_dir=$(git rev-parse --git-dir)
	worktree_name=my-program
	worktree_location=$git_dir/$worktree_name
	
	# Ran only once
	function init_worktree {
		branch_name=my-branch
		worktree_gitdir=$git_dir/worktrees/$worktree_name
		
		git worktree add "$worktree_location" "$branch_name" --no-checkout
		rm -f "$worktree_gitdir/index”
		mkfifo "$worktree_gitdir/index”
		
		echo "Initialized worktree at $worktree_location”
	}

	# Ran on every commit
	function commit_worktree {
		cd "$worktree_location”
		git read-tree —empty
		git update-index --add \
		--cacheinfo "100644,0123456789abcdef0123456789abcdef01234567,path/to/first/file” \
		--cacheinfo "100755,123456789abcdef0123456789abcdef012345670,path/to/second/file” \
		--cacheinfo "100644,23456789abcdef0123456789abcdef0123456701,path/to/third/file”
		# etc. Will be received as arguments or from STDIN.
		git commit -m "Commit message”
	}

So I guess this is the best it gets? If I wanna make it better, I guess I will need to use libgit2?

Thank you

On 2021-01-18 at 16:58:56, Utku wrote:

You could use git mktree, git commit-tree, and git update-ref for this. If you have the object IDs for all the existing blobs and trees, then you can create any necessary trees, pass the root tree to git commit-tree, and then use git update-ref to update the branch from the old value to the new value.

Note that if you're doing this in an otherwise busy repository, you should definitely specify the old value of the ref to avoid a race condition where someone else updates it at the same time.

If this seems like a lot of inconvenient work, then you'll see why we have the index.

If you just want to create a commit without modifying the _main_ index file, then you can use the GIT_INDEX_FILE environment variable to use a temporary file as the index and then read data into there and commit using git read-tree, git update-index, git commit-tree, and git update-ref. We do that elsewhere in Git, and this will definitely be easier than writing your own trees using git mktree.

This may be slightly easier with libgit2, depending on your needs.