X dropped staged pins without a question, including the selected submodule's own pin, which the loss check didn't count. A staged pin is a decision, and X sits one Shift away from x, so dropped staged pins now count as a loss and make X ask (D31, amends D30). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
22 KiB
Actions (draft)
Status: Restore (key x, D23; hard restore
X, D30),
Switch branch (key b, D24) and
Update pin (key u, D29) are implemented.
Stage pin was dropped (D29).
Principle
Offer as few actions as possible, but enough to work with submodules. submarine shows what differs from what git records. The actions resolve such a difference in one of two directions:
| Action | Direction | Key |
|---|---|---|
| Restore | The worktree follows the recorded state | x, X (hard) |
| Switch branch | Both: the branch tip is recorded, then restored to | b |
| Update pin | Both: a switch to the tip of the pin's ref | u |
There is nothing else for now. Committing and pushing are done outside
submarine. Recording a commit made in a submodule is a switch to its branch
(b); a commit on no branch must get one first.
Restore
Makes the selected repository, and everything below it, exactly what git records. It works on any submodule and on the superproject (header selected). Restoring the superproject gives the exact project that the current root commit specifies, with the pins you have staged.
Target
| Selected | Target commit | Branch |
|---|---|---|
| Superproject | HEAD |
stays on the current branch |
| Submodule | The pin in the parent's index (what the UI shows as Pinned); with a staged pin you choose, see below | see Branches |
| Nested submodule, during the recursion | The gitlink in the restored parent's index | as for a submodule |
Staged pins are kept, at every depth. A staged pin is a decision, not
drift. Each repository's files are reset to its HEAD, but its staged pins
stay in its index, and the submodules below are restored to them. So a
nested submodule with a staged pin and uncommitted changes keeps the pin
and is reset to it; its changes are discarded (and listed in the question).
Choosing the pin. If the selected submodule has a staged pin itself, x
first asks which pin to restore to:
Restore libs/staged to s staged pin · c committed pin (unstages it) any other key cancels
s: the staged pin, which stays staged.c: the committed pin (the gitlink in the parent'sHEAD). The parent's index is set back to it, which is how a pin is unstaged in submarine.
The loss check then runs for the chosen pin (the two can differ: a staged
commit that exists only locally is lost by c, not by s).
Two exceptions, where a staged pin is not kept:
- The repository moves. If a repository is off its pin, restore checks
out another commit; staged pins in its own index belonged to the old one.
They are counted as a loss (
2 staged pins), so the question appears. Its submodules follow the target commit's tree. - An added, not yet committed submodule is no staged pin: its
.gitmodulesentry is reset with the other files. Its directory then counts as an untracked file.
Hard restore
X restores to the committed pins, at every depth: what the commits
record, without any staged pin.
- The selected submodule goes to the committed pin, without the
s/cquestion; a staged pin of it is unstaged (as withc). A submodule that was added but not committed yet has no committed pin and goes to the one in the index. - Every restored repository's index is reset to its
HEAD, the superproject included: its staged pins are dropped, and the submodules below follow the committed tree.
So x and X differ in the staged pins of repositories that the restore
does not move: x (s or c) keeps them, X drops them. This matters
for the superproject, for a selected submodule without a staged pin of its
own, and for nested repositories that are on their pin. A repository that
moves loses its staged pins either way (see above); then x → c and X
end the same, but c asks first.
The dropped staged pins are counted in the loss (2 staged pins), the
selected submodule's own staged pin included, and like any loss they make
restore ask (D31): Restore libs/staged to the committed pins? Discards 1 staged pin. The
commits of a dropped pin stay in the submodule; if HEAD leaves them and no
ref reaches them, they count as commits on no branch, as always.
Postconditions
After a successful restore of repository R with target T:
- R's HEAD is T. The index and the worktree equal T's tree, except for the staged pins that were kept. No modified, added, removed or conflicted files remain.
- No in-progress operation remains in R: merge, rebase, cherry-pick, revert.
- No untracked files or directories remain in R. This includes leftover directories of submodules that T no longer has (untracked repository). Ignored files are kept.
- Every submodule that R's index records is present and initialized, even if
it was never initialized, is inactive (
submodule.<name>.active) or hasupdate = none / merge / rebaseconfigured. Its URL comes from T's.gitmodules(synced into the config). - Each of those submodules is restored the same way, with its gitlink in R's index (T's, or a kept staged pin) as its target, recursively at any depth.
So the subtree shows in sync everywhere, except for states restore does
not change: unpushed, behind/ahead of the ref, and detached HEAD on pin
where no branch is at the pin (see Branches).
The recursion follows T, not the tree loaded in the UI. A submodule that is only in T (for example, after switching branches) is initialized. A submodule that is only on disk is deleted as an untracked directory.
Branches
-
Superproject: stays on its branch. The branch is not moved, since the target is its own
HEAD. -
Submodule on the pin: if HEAD is a branch that points at the target, the branch stays checked out.
-
Submodule off the pin: HEAD is detached at the target. The branch that was checked out is not moved, so its commits stay on it.
-
Submodule detached at the target (after the step above, or because it was already so, also when it was in sync): HEAD is put on a branch whose tip is the target, the first that fits of (D28):
- a local branch tracking the ref the UI shows (or named like it), e.g.
local
maintrackingorigin/main, also when the pin is behind or ahead of the ref; - the ref itself, if its tip is the target and no local branch has its name: the local branch is created, tracking it;
- any other local branch, the first alphabetically;
- any other remote branch whose name no local branch has, created as a tracking branch; the ref's remote first, then alphabetically.
If git refuses one (checked out in another worktree, a name conflict), the next is tried. If none fits, HEAD stays detached.
- a local branch tracking the ref the UI shows (or named like it), e.g.
local
Restore never moves or deletes a branch. A local branch that is behind or ahead of its remote branch is no candidate and is not fast-forwarded. The only branches restore creates are tracking branches for remote branches at the target, in steps 2 and 4.
What can be lost
Restore discards work. The things it can lose:
| Lost | Where it comes from |
|---|---|
| Changed files | modified, added, removed, renamed files (changed) |
| Untracked files | changed (untracked) |
| Conflicts and in-progress operations | conflict; merge, rebase, cherry-pick, revert state |
| Commits | A detached HEAD off the pin whose commits are on no branch, tag or remote ref (only the reflog still has them) |
| Staged pins | Staged pins in a repository that restore moves to another commit (see Target); with X all staged pins |
| Repositories | An untracked directory that is a repository with its own .git directory (a clone that was never added as a submodule). Its commits may exist nowhere else. A removed submodule whose git directory is absorbed into .git/modules loses only its worktree. Its objects stay |
Nothing else is lost. Branches, stashes, remote refs, ignored files and absorbed submodule git directories stay.
Confirmation
Restore asks only if something would be lost. Never asking is unsafe
(git clean -ff deletes repositories that exist nowhere else, and nothing
undoes a restore); always asking puts a question on the harmless, common
cases (off pin, missing, a staged pin) and teaches a reflexive y.
-
Nothing would be lost: restore runs at once.
-
Something would be lost: a line above the footer says what is discarded, summed over the whole subtree on disk:
Restore src/app? Discards 7 changed files in 3 repositories, 2 commits on no branch, 1 untracked repository. y restore · n cancelyruns it, any other key cancels. -
The state can't be read for a repository: it is counted as
n repositories of unknown stateand needs the samey.
The check reads the disk at the moment of the keypress (RestoreLoss), not
the loaded statuses, so loading or stale statuses don't matter. Nothing is
lost that is not in the table above; ignored files are never touched.
Network
A restore may need objects that are not on disk: a submodule that was never
cloned (no git directory in .git/modules), or a target commit that was
never fetched. Restore may clone and fetch, but only what is missing,
non-interactively (GIT_TERMINAL_PROMPT=0, ssh BatchMode) and with a
timeout. If the commit is on no remote, restore fails with commit … is not available on any remote. This is the one place besides f that uses the
network (decision D23).
Running, progress and errors
- A line above the footer shows
checking …and thenrestoring …with a spinner. Only one restore runs at a time; pressingxmeanwhile flasheswait for the restore of src/app to finish. Everything else keeps working. - A running fetch does not block a restore.
- Done: flash
restored src/app; the app reloads the tree and all statuses, because the structure can have changed. - Failure: the restore stops with git's message, shown in red above the
footer (
restore of src/app failed: …) until the next key. Restore is idempotent, so it can be run again after fixing the cause. Repositories that were restored before the failure stay restored.
Implementation note
git.Provider.Restore (internal/submodule/git/restore.go) walks the
repositories itself, top-down; the children of a repository run
concurrently, bounded by the provider's semaphore. For each repository
(T = target) it runs, with -c submodule.recurse=false:
cat-file / fetch --all # only if T is missing locally
rebase|cherry-pick|revert --quit # whichever is in progress; --quit, not --abort: no ref moves
checkout --force --detach T # only if HEAD is not T, so a branch is never moved
reset --hard # also ends a merge
update-index --cacheinfo … # put the staged pins back (only if HEAD did not move)
clean -ffd # -ff: untracked repositories too; ignored files stay
submodule sync
submodule update --init --checkout --force -- <not yet initialized children>
# then every gitlink of the index (ls-files -s) is restored the same way
--checkout overrides submodule.<name>.update, and the explicit paths
override submodule.<name>.active, so no child module is skipped.
For a submodule, after submodule sync, attachBranch puts a detached HEAD
on a branch: symbolic-ref -q HEAD (on a branch: done), for-each-ref --points-at=T refs/heads refs/remotes (no branch at T: done), then the ref
is resolved like Status does, and the first candidate (see
Branches) that switch / switch -c <local> --track <remote>
accepts wins. The branch checked out before the restore is passed to the
ref resolution. A child's .gitmodules branch is read once per parent.
With c and X the parent's index gets the committed pin first
(update-index --cacheinfo in the parent). With X (RestoreHard) the
staged pins are not put back, and the loss check takes the children's
targets from HEAD's tree instead of the index.
RestoreLoss (lossOf) follows the repositories on disk: per repository
status (without submodules), the in-progress state files, and
rev-list --count HEAD --not --branches --tags --remotes if HEAD is not the
target. A directory in the untracked list is searched for .git
directories (a standalone repository); an absorbed submodule leaves only a
.git file and loses just its worktree.
Switch branch
Moves the selected repository to another branch, in one step: the tip of the branch becomes the pin, and the repository is restored to it.
The popup
b opens a popup in the middle of the screen: a search line on top, below it
the branches of the selection, local ones first, then remote ones
(origin/…), each group with the newest tip first. Each row shows the name
(* marks the checked-out branch), the short SHA and the age of the tip.
The list is read from the disk; fetch (f) is what brings new remote
branches.
| Key | Effect |
|---|---|
| text | Filters: every word must be in the name, case-insensitive. The selection stays on its branch if it still matches |
↑ ↓, pgup pgdn |
Select |
enter |
Switch to the selected branch |
esc |
Close |
Mouse: a click on a branch switches to it (like enter), the wheel moves the
selection, a click outside the popup closes it; a click on the title, the
search line or an empty row does nothing. Everything else is ignored while the
popup is open. A header selection switches the superproject.
Effect
For a submodule, with the tip T the list showed:
- The parent's index records
Tas the pin (a staged pin; ifTis the committed pin, the index equalsHEADagain). - The submodule, and everything below it, is restored to
T. - HEAD is put on the branch (the branch name is the local one):
- a local branch: checked out;
- a remote branch
origin/feat: the local branch is the one checked out before if it tracksorigin/feat(e.g.dev), elsefeat. It is created, tracking it, if it does not exist; if it exists atTit is checked out; if it is only behindT(an ancestor of it) it is fast-forwarded toTand checked out (D29); if it has commits of its own, or is checked out in another worktree, it is not moved and HEAD stays detached atT.
For the superproject step 1 is left out: the branch is checked out via the restore, so its submodules follow the pins of the new commit (staged pins of the old state are lost, and counted).
Staging comes first, so a failed restore can be repeated with x. If the
branch moved since the list was read (a fetch), the switch aborts with
origin/feat moved; open the list again; the staged SHA is always the one
shown.
Confirmation, network, errors
As for restore: the loss is read from the disk and the question
(Switch src/app to origin/feat? Discards … y switch · n cancel) appears
only if something would be lost. A submodule that is not initialized can't be
switched (src/app is not initialized; restore it first). Network: only what
is missing below, like restore. Lines above the footer: checking …,
switching src/app to origin/feat…; flash switched src/app to feat (or
… to origin/feat (detached: local feat can't be fast-forwarded)); failure
switch of src/app to origin/feat failed: …, until the next key. One
restore or switch runs at a time.
Implementation note
git.Provider.Switch (internal/submodule/git/switch.go): verify the ref is
still T; update-index --cacheinfo in the parent; restoreRepo (see
restore); then switch / switch -c <local> --track <remote-branch>; a local branch
behind T is first moved with merge-base --is-ancestor <local> T, a look
at worktree list --porcelain and update-ref refs/heads/<local> T <old>
(which fails if the branch moved meanwhile). The restore leaves HEAD of the
switched repository alone (Switch puts it on the chosen branch); the
submodules below it are attached as in a restore.
SwitchLoss is lossOf with T as the target. Branches is one
for-each-ref, without the origin/HEAD symref.
Update pin
Moves the pin of a submodule to the tip of its ref, the branch the row shows
(origin/main ↓2). It is a switch to that branch, with the
tip the row shows: nothing is fetched, so u gives exactly the commits the
row counts. f is what brings newer ones.
On the superproject the pin is HEAD and the ref is the upstream of its
branch (the header's main ↓1): u fast-forwards the branch to it, like
git pull --ff-only, and restores the submodules to the pins of the new
commit.
Available when
All of:
- it is initialized and its state is loaded;
- its ref is a remote branch (not a tag, a SHA or a local branch) and has a tip; for the superproject: its branch has an upstream;
- the pin is behind the ref and not ahead (
Behind > 0,Ahead == 0). A diverged pin (↑n↓m) is left tob: moving it would drop the pin's own commits from the record.
Effect
The same as switching to the ref: the tip is staged
as the pin, the submodule and everything below it are restored to it, and
HEAD is put on the local branch. A local branch that is only behind is
fast-forwarded, so the common case (local main at the old pin) ends on
main. A checked-out branch that tracks the ref under another name (dev
tracking origin/main) is the one used, so the update stays on it. For the
superproject staged pins are lost, as in a switch (and counted).
If the ref no longer points at the tip the row showed (a fetch meanwhile),
nothing happens: can't update src/app: origin/main moved; check the new state. The watcher has updated the row by then; u again takes the new
tip.
Confirmation, lines, errors
As for switch: the question appears only if something would be lost
(Update src/app to origin/main? Discards … y update · n cancel). Lines
above the footer: checking src/app…, updating src/app to origin/main…;
flash updated src/app to main; failure update of src/app to origin/main failed: ….
Implementation note
The UI only: startPinUpdate (internal/tui/pinupdate.go) reads the
branches (Switcher.Branches), takes the one named like Pin.Ref, checks
that its tip still starts with the abbreviated Status.Tip.SHA, and then
runs the switch flow with it (SwitchLoss, Switch).
Keys and palette
All three are submodule keys (keyMap.Submodule), shown in the detail
pane's palette, or in the footer when the pane can't show it. They are
always offered, the palette is static: restore is idempotent, and u says
why it can't update.
| Key | Action | Pressed when not possible (flash) |
|---|---|---|
x restore |
Restore the selection | this data source can't restore (mock); wait for the restore of … to finish |
b switch branch |
Switch branch | this data source can't switch branches; wait for the restore of … to finish; src/app is not initialized; restore it first |
u update pin |
Update pin | this data source can't update pins; wait for the restore of … to finish; the superproject has no upstream branch; src/app is still loading; src/app is not initialized; restore it first; the pin of src/app is not on a branch; the pin of src/app has diverged from origin/main; b switches the branch; the pin of src/app is ahead of origin/main; src/app is up to date with origin/main (the superproject without the pin of); after the check can't update src/app: main is a local branch; b switches to it |
R would sit next to r (reload), a dangerous mix-up for a destructive
action, hence x.
Testing
- Provider (
restore_test.go), on fixtures fromscripts/make-fixture.sh: a restored superproject is exact (every repository clean,submodule status --recursivewithout prefixes,libs/missinginitialized), branches are not moved, ignored files survive, a merge and a rebase in progress end, a standalone clone and a removed submodule's directory are deleted, a missing submodule is cloned again,update = noneand inactive submodules are restored, an unavailable commit is reported. A detached submodule ends on the local branch tracking the ref, else on the ref created as a tracking branch (before other branches at the pin), else on the first local branch; a local branch elsewhere is not moved; a checked-out branch stays; a branch of another worktree is skipped; nested submodules are attached too. The loss is empty for off pin, missing and unpushed submodules, counts files, nested changes, merges, detached commits (not those on a branch) and standalone repositories. - UI (
internal/tui/restore_test.go): restores at once without a loss, asks with one (yrestores, anything else cancels), the mock can't restore. - Switch: provider (
switch_test.go): list withoutorigin/HEAD, local first; a remote pick stages the tip and creates a tracking branch; a local branch that is only behind is fast-forwarded; one with commits of its own, or checked out in another worktree, is not moved; a moved ref aborts; the superproject switch ends exact; the loss is the restore's. Picker (branches_test.go): filter, selection, scrolling. UI (switch_test.go): opens, filters, enter switches, asks with a loss, esc closes, the mock can't. - Update pin: UI (
pinupdate_test.go): switches to the tip of the ref, asks with a loss, aborts if the ref moved, refuses (with its flash) a pin that is up to date, ahead, diverged, on a tag, on a local branch or not initialized; updates the superproject, not one without an upstream. Provider: the superproject's branch is fast-forwarded; a checked-out branch that tracks the remote branch under another name is kept. - By hand in tmux on a fixture, see architecture.md.
Open questions
Decided for restore: ignored files are kept (clean without -x); a
submodule is restored to the index gitlink, so a staged pin is its target and
survives; restore may clone and fetch what is missing; the key is x; it
asks only if something would be lost.
Open (update pin): updating several submodules at once (on the superproject or a collapsed parent), deliberately left out for now.