like gotd, assume an implicit ".git" if the repo doesn't exist. Merge gotweb_init_repo_dir() and gotweb_load_got_path() since the former only sets repo_dir->name, and we don't realistically know the name until we've opened the directory. While here, unbreak the git worktree handling too. Realistically, moving forward, we should issue a redirect to the canonical URL or at least set the correct metadata in the document to avoid having multiple URLs for the same content. ok stsp; tiny change to repo_dir->name handling done after the ok.

My previous fix for ~user in SSH URLs broke ~user in HTTP URLs, as used on sourcehut. Sanitize the amount of leading and embedded slashes in URLs to make things work in all cases. The root cause of the test problem with naddy's previous diff was likely in the http-server script itself, where a request like GET //repo//info/refs?service=git-upload-pack somehow ended up as GET $testroot///info/refs. We work around this here by avoiding the double slashes.

It broke the clone regression test.

Add a leading '/' and do not produce a double "//" after the path. ok op

The stage command is usually used in a recursive manner, like 'got commit'. Forcing users to specify -R all the time is deemed too inconvenient in practice. discussed on IRC with Lorenz (xha), Omar, and Lucas

This leak resulted in the gotd notify process spinning on CPU in an imsg_read() -> EAGAIN loop once it ran out of file descriptors.

Document http and https protocols in got.conf(5). Copied from got(1) except reworded the part that referred to got.conf(5). Change "send directive" to "send block" in got(1), since got.conf(5) says there's "one configuration directive per line". Adjust table widths: longest protocol name is now "https".

We try to handle early replies from the server, so we call tls_read() at each "tick" of the event loop. For the TLS case, however, bufio_read() will set bio->wantev, and if we rely only on it we can deadlock trying to read data from the server without having sent all the request. Found out the hard way while trying to send several notifications in one go. ok stsp

pointed out by op@

Ephermeral processes on the other end of a pipe might decide to exit which results in EPIPE when writing. This is not a fatal error but is somewhat expected during normal operation (at least until we improve the inter-process communication about notifications).

We would end up raising a spurious error when the server had no more data to send and would abort the clone operation, while everything was fine in fact. Problem reported by vazub via IRC, thanks!

Use the more predictable %F, aka %Y-%m-%d, instead of %G-%m-%d. %G follows the definition of ISO-8601 week-based year, which is weird. In particular, 2024 is one of such years with weird behaviour: $ date -jf %Y-%m-%d +"%F %G-%m-%d" 2024-12-30 2024-12-30 2025-12-30 Diff from Lucas Gabriel Vuotto (thanks!); stsp agrees

we don't need getopt.h; getopt() is in unistd.h. also sort the headers while here. ok stsp@

'got clone' already happens to support such URLs via scp-style syntax: user@127.0.0.1:~/test.git While the clone operation succeded, subsequent fetch operations would raise an error: '/~/test.git' does not appear to be a git repository Make such URLs work when they are read back from got.conf, and also on the clone command line in canonical form: ssh://user@127.0.0.1/~/test.git Not adding a test case because I don't see an easy way to add test coverage for this without spilling test data into home directories or modifying ~/.ssh/ssh_config. Problem found by Jesper Wallin.

Check what happens when patches attempt to edit or delete empty files.

if we've reached EOF without finding a match is fine if and only if the target file is empty, so add this exception. reported by naddy, ok stsp