portable: set next version

[got-portable.git] / TODO

got:
- Teach 'got merge' to merge changes into an arbitrary subdirectory
  of the work tree. This would be nice for merging vendor branches.
  Say you have a branch 'llvm-12' which intially contains a 12.0
  release tree as published by the LLVM project, added to the repository
  with a command such as 'got import -b llvm-12'. On the main branch we
  would want to merge files from the llvm-12 branch into /usr/src/gnu/llvm
  instead of the root directory checked out at /usr/src.
  The next LLVM release 12.1 would later be committed onto the llvm-12
  branch and then merged into main at /usr/src/gnu/llvm in the same way.
- Teach 'got merge' to forward a branch reference if possible, instead of
  creating a merge commit. Forwarding should only be done if linear
  history exists from the tip of the branch being merged to the tip of
  the work tree's branch, and if the tip of the work tree's branch is
  itself not a merge commit (this makes "stacked" merges possible
  by default, and prevents a 'main' branch reference from being forwarded
  to a vendor branch in case no new commits were added to 'main' since
  the previous vendor merge). Provide an option (-M) which forces creation
  of a merge commit, for cases where users deem forwarding undesirable.
- When a clone fails the HEAD symref will always point to "refs/heads/main"
  (ie. the internal default HEAD symref of Got). Resuming a failed clone with
  'got fetch' is supposed to work. To make this easier, if the HEAD symref
  points to a non-existent reference it should be updated by 'got fetch'
  to match the HEAD symref sent by the server.
- If invoked in a work tree, got fetch could default to fetching the work
  tree's current branch, instead of fetching the remote repository's HEAD.
- 'got patch' should be able to detect an already applied patch.
- 'got patch' should ideally do more passes if a patch doesn't apply and
  try fancy things (like ignoring context and whitespaces) only in later
  passes.
- investigate whether it's worth for 'got patch' to memory-map the files to
  edit.  (c.f. Plan A / Plan B in Larry' patch.)
- when fetching pack files got should verify that the requested branch tips
  are present in the pack file sent by the server, before making this pack
  file visible to readers of the repository

network protocol:
- add http(s) transport with libtls, speaking the two Git HTTP protocols
  (both "dumb" and "smart" need to work) via got-fetch-pack, or a new helper
  like got-fetch-http; it is fine if HTTP remains a fetch-only protocol, and
  works only against servers which don't require authentication for fetches;
  anything beyond this would require a full-featured HTTP client (Git uses
  libcurl, which we cannot use as it is not in the OpenBSD base system)

tog:
- make 'tog log' respond to key presses while 'loading...' history; loading
  can be slow for paths in a deep history if the path has not been changed
  very often, and 'tog log' blocks far too long in this case
- make it possible to view the contents of tag objects
- verify signed tag objects
- make it possible to toggle the parent to diff against in merge commits

gotwebd:
- reply with a non-200 reply code when an error occurs
- fix COMMITS page for paths that were deleted and/or re-added to the
  repository. One way would be not to let the commit graph filter paths.
  As an additional optimization we could keep a tailq or the object-id
  set for everything traversed in the repo to have fast reverse-lookups.
  (has the additional requirement to invalidate it when the reference
  timestamp changes)
- run-time changes of addresses on interfaces are being ignored by gotwebd
- perhaps "bind interface" should be removed in favour of using only IP
  addresses? This makes the address family selection explicit and avoids
  having to monitor interfaces for dynamic address changes.
- add breadcrumbs to navigate in the directory tree
- add more action links, like "raw", "blame" and "commits" in the blob
  page.
- add a "rawdiff" page to serve a diff as text/plain (or text/x-patch),
  or maybe "rawcommit" that includes the commentary too.
- consider changing the URL scheme to avoid so many query parameters

gotd:
- ensure all error messages are propagated to clients before disconnecting,
  there are probably still some cases where such error reporting is skipped
- client connection timeout handling needs to be checked by regress tests,
  and is likely in need of improvement
- implement stress-tests to observe and fix misbehaviour under load
- listener's fd-reserve limit needs to be reviewed and perhaps adjusted 
- implement pre-commit checks (in lieu of hook scripts):
  1. deny branch history rewriting ('got send -f') via gotd.conf
  2. allow/deny creation/deletion of references via gotd.conf
  3. deny modifications within a given reference namespace via gotd.conf
  4. entirely hide a given reference namespace from clients via gotd.conf
  5. allow/deny addition of binary files to a repo via gotd.conf
  6. enforce a particular blob size limit via gotd.conf
  7. optionally reject merge commits via gotd.conf
- implement post-commit-event libexec handlers (in lieu of hook scripts):
  1. commit email notification, plaintext smtp to localhost port 25
  2. general-purpose HTTP(s) GET/POST request as commit notification,
     should use a format string to encode commit-info in the request URL
  3. perform the equivalent of 'got send' to another repository
- keep track of available repository disk space and fail gracefully when
  uploaded pack files would fill up the disk too much, keeping a reserve
- reuse packed non-delta objects directly (without re-deltification) for speed
- create packfiles with filemode 444, like git does

gotadmin:
- teach 'gotadmin cleanup' to remove redundant pack files
- add support for generating git-fast-export streams from a repository
- add support for importing git-fast-export streams into a repository
- create packfiles with filemode 444, like git does
- speed up 'gotadmin pack -a' is too slow on repositories with many pack files