Is there any public repo to reproduce this?

Working good locally so far. The only way to get the deleted type would be to use head_commit as first param and prev_commit as second.

Can you try to set the HEAD to the sha where the file was first created and run a diff like this:
repo.diff('HEAD^', 'HEAD')

That would be bulletproof. Your Git version would be useful too.

Does a command line git diff prev_commit head_commit works as expected ?

Sorry for not including steps to reproduce, but here you go:

$ git clone https://github.com/isaacs/readable-stream
$ cd readable-stream
$ irb

irb> require 'git'
irb> repo = Git.open('.')
irb> diff = repo.diff('5aec72b01618ddfc46ad09cdbc66dc8d26211e52', 'c49f6387e8de23dd8da96b513037ca99800014c3')
irb> diff.map(&:type)
[ ..., "+deleted", ... ]

Funny (?) cuz yours is just the tip of the iceberg.

Any file line, within any kind of file, containing \bfilemode\b.{6} will modify the previously value of the type and mode.

Also, any line containing index\b.{7}\.\..{7} will also modify src, dst and even the mode in some cases.

It's weird that an small validation (that does not prevent mode from been overwritten) has been added for the later scenario, but just for mode:

final[current_file][:mode] = m[3].strip if m[3]

I'll create some tests. I think we can easily prevent problems
* Improving the regexps just a little bit
* Protecting those values if they are already set

Ideally the parsing logic should be improved to detect context and forget about specifics within those.

Working on it....

Damn, you can even "create files" that do not even exist by adding/having lines following diff --git a\/(.*?) b\/(.*?)

Basically .patch files, logs, or things like that can create tons of invalid diff entires if they are included. in the diff output.

Opening a new issue for a general refactor of the thing.

Anyways, all the previously listed scenarios should be fixed.

Tests e55f3bd

Thanks @robertodecurnex. Verified it works on my end.