refreshStashEntries now loads the stash entries on the worker and writes
Model.StashEntries in an onUIThreadUnlessRepoChanged bounce.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
refreshTags now captures the repo generation, loads the tags on the
worker, and writes Model.Tags in an onUIThreadUnlessRepoChanged bounce
rather than directly from the worker goroutine.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
refreshCommitFilesContext now enqueues the Model.CommitFiles write and
CommitFileTreeViewModel.SetTree() call via OnUIThread, instead of running
them directly on the worker goroutine that drives async refreshes. This is
what makes moving SwitchToDiffFilesController's post-refresh work into Then
(previous commit) actually necessary, rather than just future-proofing.
Same repo-switch hazard as the FILES bounce, closed the same way: it
captures the repo generation before the git work and bounces through
onUIThreadUnlessRepoChanged, so the write is dropped if the user switched
repos while it was in flight.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
FileTreeViewModel.RWMutex is removed along with the
withFileTreeViewModelMutex wrapper in FilesController that RLocked it:
every writer (the bounce closure, previous commit) and every reader (key
handlers, disabled-reason callbacks) now runs on the UI thread, so the
mutex is redundant.
RefreshingFilesMutex is removed entirely, including its last use in
repos_helper's DispatchSwitchTo. That use predates the bounce and was
never about FilesController's optimistic-rendering concern; it serialized
a repo switch's onNewRepo() against an in-flight FILES refresh for the
repo being switched away from, so that a slow refresh from the old repo
couldn't write into the freshly-reset model for the new one. Bouncing the
write already broke that guarantee on its own terms — the mutex's critical
section never covered the bounced closure's actual execution, only the
(now-removed) code that enqueued it — so by this point it was only still
locked here without protecting anything real; the previous commit's
repo-generation guard is what now actually closes that race, making this
lock fully redundant rather than just relocated.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
refreshStateFiles now does its git work on the worker and enqueues a
single OnUIThread closure that writes Model.Submodules, Model.Files, and
the FileTreeViewModel state together, instead of writing them directly
from the worker goroutine. refreshStateSubmoduleConfigs becomes a pure
getter (returns the configs; no model write) so the result can be
threaded into that same bounce.
The STAGING handler wraps RefreshStagingPanel in OnUIThread after
fileWg.Wait() so it sees the post-bounce file model rather than the stale
pre-refresh one — without this it would race the files bounce queued just
above it.
Bouncing the write opens a hazard the old synchronous write didn't have:
if the user switches repos while this refresh is in flight, the queued
closure would fire after resetState has replaced the model with a fresh
one for the new repo, silently overwriting it with the previous repo's
files. Guard against this with a repo generation: resetState bumps a
counter on every switch, refreshStateFiles captures it before its git
work, and onUIThreadUnlessRepoChanged drops the bounce if the generation
has moved on. This one helper is the general mechanism the remaining
scopes' bounces will use too; the same guard covers the rebase-continue
prompt, which reads Model.Files right after.
A generation counter, not a comparison of the *Model pointer: switching
away from and back to a repo reuses that repo's cached state (the same
Model pointer), which a pointer comparison would wrongly accept even
though the in-flight data is stale.
PromptToContinueRebase's Then callback (previous commit) now gets an
explanatory comment, since this is the commit that makes it necessary.
The explicit locking around these writes (RefreshingFilesMutex in
refreshFilesAndSubmodules, FileTreeViewModel.RWMutex around the write in
refreshStateFiles) is left in place for now even though it's becoming
redundant, to keep this commit focused on the bounce itself; it's removed
next.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
PromptToContinueRebase and WithEnsureCommittableFiles both read
Model.Files right after a SYNC FILES refresh. This works today because
the model write currently happens synchronously in the worker before
Refresh's wg.Wait() returns, but an upcoming commit will bounce that
write onto the UI thread instead, at which point wg.Wait() no longer
guarantees it's been applied.
Move both reads into Then ahead of that change. Then is already queued
via OnUIThread (previous commit), so this is a behavior-preserving
refactor on its own: the model is fully written by the time Then runs
either way, whether that write is still synchronous or gets bounced
later.
As part of restructuring WithEnsureCommittableFiles, prepareFilesForCommit
and syncRefresh are inlined into their single call sites.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This is preparation for upcoming commits that will bounce refresh-scope
model updates (e.g. Model.Files) onto the UI thread by enqueuing the
write via OnUIThread instead of applying it directly on the worker
goroutine. Once that lands, a Then callback that reads the model must
run after that queued write has been processed, not synchronously at
wg.Wait() time — at that point the workers have returned, but a bounce
they queued may not have been processed yet.
Queuing Then via OnUIThread here, ahead of that change, guarantees the
right ordering once it lands: a bounce queued earlier in the same
refresh is already sitting in the channel by the time wg.Wait()
returns, so Then enqueued after it will always be processed after, and
see the post-refresh model.
The signature change to func() error lets Then propagate errors
through gocui's normal error handler (the same path key-handler errors
take).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
GetIsRefreshingFiles() is never called anywhere in the codebase, so the
flag serves no purpose. Remove it from Gui, StateAccessor, and
IStateAccessor, and drop the two SetIsRefreshingFiles calls in
refreshFilesAndSubmodules that maintained it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Operations that show an inline status ("Pushing", "Fast-forwarding",
"Fetching", …) removed it by relying on the async refresh they trigger
to redraw the view after the item operation had been cleared. That
ordering was never guaranteed: the item operation is cleared on the
worker once the operation's function returns, while the refresh redraws
the item from the UI thread whenever its (asynchronous) git work
happens to finish. If the refresh redrew before the clear, the status
was left on screen with no later redraw to remove it, so the branch (or
tag/remote) stayed stuck showing e.g. "Pushing" indefinitely even though
the operation had completed. This is timing-dependent, which is why it
surfaced as rare, hard-to-reproduce reports and as flaky CI failures.
Fix it by re-rendering in stop() right after clearing the operation,
and by making these refreshes synchronous rather than async. Because a
synchronous refresh has already updated the model and queued its own
redraw by the time stop() runs, and UI-thread callbacks run in order,
the redraw we queue here runs last and draws the up-to-date model with
the status removed. An async refresh couldn't give that guarantee: its
model update might not have landed yet, so the redraw could briefly
flash the pre-operation status.
Pull refreshes through the shared CheckMergeOrRebaseAndSelectHeadCommit,
so that helper becomes synchronous too; its only other caller,
RegularMerge, thereby also refreshes synchronously, which is fine: a
synchronous on-worker refresh is what we want anyway.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Resetting to a commit/branch/tag from the reset menu ran inline on the UI
thread with no spinner; a hard reset to a distant commit can take a while
and blocks the UI meanwhile. Run it on a worker with a waiting status. The
undo/redo callers of ResetToRef already wrap it this way.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The interactive-rebase item in the rebase-onto-ref menu ran inline on the UI
thread with no spinner, unlike its two siblings in the same menu (simple
rebase and rebase onto base branch), which already run on a worker with a
waiting status. Make it consistent.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The regular and squash merges from the merge menu ran inline on the UI
thread, freezing it with no spinner while git worked. Run them on a worker
with a waiting status, like the rebase entry points already do.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Continuing, skipping, or aborting a merge/rebase from the options menu ran
the git command inline on the UI thread, freezing the UI with no spinner
while it worked (a continue can replay many commits). Run the
non-subprocess path on a worker with a waiting status instead, matching how
the other merge/rebase entry points already behave.
The auto-skip recursion in CheckMergeOrRebaseWithRefreshOptions must not
start its own worker: it already runs on the caller's thread (the worker of
the enclosing waiting status, or the UI thread for the synchronous callers).
Route it through genericMergeCommandImpl with the waiting status suppressed
so its behavior is unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When loading the files of a commit we passed --no-renames, so a rename
showed up as a separate delete and add rather than a single R entry.
That made it impossible to work with a rename that also modifies the
file: the modifications were spread across a full deletion and a full
addition instead of appearing as the handful of lines that actually
changed. The staging view already shows renames and lets you stage
their hunks, so there was no good reason for the patch builder to
differ; the flag was only there because the commit-file parser couldn't
cope with the rename record format.
Switch the commit-file loader and the per-file diff to --find-renames,
teach the parser about the rename record (a status followed by two
paths), and carry the previous path through the patch builder so the
diff for a rename is loaded with both paths, which is what makes git
emit the rename in the first place.
A whole-file selection keeps the rename in the header, so the rename
moves or is discarded together with the file's contents. A partial
selection instead strips the rename metadata and points the header at
the new path, so applying the patch only changes the contents and
leaves the rename in place; the blob index line is kept so that a 3-way
apply can still fall back to a blob merge.
Discarding a renamed file from a commit now discards both the new and
the old path, so the new file is removed and the old one is restored.
Changing the rename similarity threshold refreshes the commit files
panel too, not just the files panel, so that a rename can turn into a
delete and add or back. It is disabled while building a patch, however,
because the patch builder caches each file's diff by path and would
desync if a rename changed into a delete and add underneath it.
Finally, copying a file's diff from the commit files panel now passes
both paths for a rename, so the copied diff shows the rename instead of
a new-file add.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Accordion mode expands the focused side panel, but when that panel has
little content (an empty Files panel, a Branches panel with only master)
it just fills the extra height with blank space. The same waste happens
for any panel that gets more height than it has content to show.
When this option is enabled, each side panel is sized to its own content
(plus a blank line, so it's clear there's nothing more below) rather than
to an equal share of the height. The height a small panel gives up flows
to the panels that have more content than fits; those grow up to their
content and then scroll, weighted toward the focused panel in accordion
mode so the two features compose. Only when every panel fits with room to
spare is the leftover shared out equally, regardless of focus: enlarging
the focused panel there would reveal no more content and would only make
the panels jump around as the focus moves.
The option is independent of expandFocusedSidePanel and off by default.
The status panel, and the stash panel when unfocused, keep their fixed
one-line height as before.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The height thresholds that decide between the proportional layout and
the squashed layout (and, within the squashed layout, between 3-row and
1-row unfocused panels) were hard-coded constants tuned for the fixed
set of five side panels. Now that the panels are configurable, a layout
with fewer panels has less to fit, yet was still forced into the
squashed layout at the same height as five panels would be.
Scale the thresholds down in proportion to the panel count so a smaller
layout keeps using the proportional layout at smaller heights. Only ever
scale down: raising the thresholds for more panels would make them
squash sooner, which works against the reason someone adds panels in the
first place (they want to see them).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pressing `d` on a worktree only ever removed the worktree, leaving its branch
behind even though deleting it too is often what you want. Turn the confirmation
into a menu: "Remove worktree", "Remove worktree and delete branch", and "Remove
worktree and delete local and remote branch". The branch-deleting items come
after the plain removal (they do more harm if picked by accident); both are
greyed out for a detached-HEAD worktree, and the local-and-remote one is also
greyed when the branch has no upstream. The plain menu pick is the confirmation,
so the standalone "remove worktree?" prompt is gone (and its now-dead
translation string with it); the dirty-worktree force prompt and the
unmerged-branch warning still appear when relevant.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Picking "Delete local and remote branch" for a single branch that's checked
out in another worktree used to fail with "Some of the selected branches are
checked out by other worktrees. Select them one by one to delete them." That
message only makes sense for a multi-selection; for a single branch there's no
reason we can't remove the worktree and delete both the local and remote branch
in one go. Route that case through the same worktree menu as the local-only
delete, with labels that spell out that the remote goes too. The multi-select
error stays for actual multi-selections.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When you delete a local branch that's checked out in another worktree, the
menu offered to remove or detach the worktree but then stopped there, leaving
the branch you asked to delete still around. Now both actions delete the branch
afterwards, and the labels say so ("Remove worktree and delete branch" /
"Detach worktree and delete branch") to avoid surprises.
Also drop the "Switch to worktree" item: switching abandons the delete the user
asked for, and it's already reachable by checking out the branch or via the
worktrees panel. And drop the now-redundant "remove worktree?" confirmation:
the explicit menu pick is the confirmation (the dirty-worktree force prompt and
the unmerged-branch warning still appear when relevant).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Split the actual worktree removal out of the confirmation in Remove into a
non-confirming helper, and give both Remove and Detach an optional `then`
continuation that runs after a successful removal in place of the default
refresh. Upcoming flows need to delete the worktree's branch once the worktree
is out of the way; threading a continuation through (rather than the caller
firing branch deletion independently) keeps it ordered after the git command
that actually frees the branch. No behavior change yet.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pull the merged-check-and-force-warning step and the actual git deletion
out of ConfirmLocalDelete and ConfirmLocalAndRemoteDelete into helpers, so
that the upcoming worktree-aware delete flows can reuse them instead of
duplicating the logic. No behavior change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lazygit runs git directly rather than through a shell, so a literal "~"
reaches `git worktree add` unexpanded and git creates a directory named
"~" instead of using the home directory.
Expand the tilde ourselves, both for paths typed into the "Other"
location prompt and for the worktree.defaultPath config value.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The old 'n' flow opened a "normal vs detached" menu (the same meaningless
gate the 'w' flow used to have), then asked for a base ref, a path typed from
scratch, and a branch name in three separate prompts.
Replace it with a single picker prompt titled "New worktree for branch",
suggesting local branches not already checked out anywhere, plus remote
branches that don't yet have a local branch of the same name. The entered
value is classified on confirm: an existing local branch checks out into a
new worktree, a remote branch creates a new local tracking branch, and
anything else creates a new branch off the current ref. All three then feed
the same location menu the 'w' flow uses, so paths are chosen from candidates
rather than typed blind. Picking a remote or new branch needs no separate
name prompt — the picker value already is the name. Checked-out branches are
filtered from the suggestions, and a verbatim type-in of one is rejected with
an error.
createWorktree now takes the context to switch focus to once the worktree is
created, so 'n' lands back in the worktrees panel while 'w' still lands in
the branches panel.
This deletes the old NewWorktree / NewWorktreeCheckout core and the now-
orphaned i18n (CreateWorktreeFrom, CreateWorktreeFromDetached, NewWorktreeBase,
NewBranchNameLeaveBlank), completing the migration started for 'w'.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The old flow forced an up-front "normal vs detached" menu (meaningless for
commits, tags and stashes), then asked the user to type a worktree path from
scratch — easy to get wrong, and ambiguous about what relative paths resolve
against. It also offered the same two actions everywhere regardless of what
was selected.
Replace it with per-context "Worktree" menus whose items imply the intent
(new branch + worktree, worktree for an existing branch, detached worktree),
each feeding a shared name -> location -> create pipeline. The location menu
offers candidate parent directories as absolute paths instead of a blank
field, and "Worktree for a branch" is disabled (with a reason) when that
branch is already checked out somewhere, rather than failing after the fact.
Each ref/commit panel binds 'w' in its own controller and calls the matching
typed entry point on the worktree helper, so which menu opens is decided
statically by the call site rather than by dispatching on a ref's dynamic
type. The three commit panels share one menu through BasicCommitsController;
there is no longer a shared worktree-options controller.
The worktrees-panel 'n' flow and its old core (NewWorktree /
NewWorktreeCheckout) are left untouched here so nothing is written and then
rewritten; they migrate, and the dead i18n strings get removed, in a
follow-up commit.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This is the core of "never type a path from scratch": from the repo root,
the configured default path, and the parents of existing worktrees, derive
the ordered list of directories under which a new worktree could be placed.
Pure and unit-tested here; wired into the creation flow in a later commit.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The prompt offering to continue a rebase/merge is opened from a refresh
and then left to sit until the user acts on it. But the operation can
change out from under it: a coding agent (or the user in another
terminal) might continue or abort it, or advance it to a commit with new
conflicts. The prompt then becomes stale — pressing continue fails with
"no rebase in progress" or acts on the wrong state.
Track whether the prompt is showing, and on each refresh dismiss it if
the operation is no longer in the "resolved, ready to continue" state
that the prompt is offering to act on. This runs on the same refreshes
that would open it (including the background poll and the refresh on
window focus), so the prompt disappears on its own shortly after the
operation moves on.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When conflicts of an in-progress rebase/merge/cherry-pick/revert are
resolved, lazygit pops up a prompt offering to continue it. This is
helpful when you started the operation in lazygit and resolved the
conflicts in your editor. But it's confusing when the operation was
started outside lazygit — e.g. by a coding agent in another terminal
that resolves the conflicts but hasn't continued yet because it's still
running tests or fixing the build. lazygit would then prompt unbidden.
Track whether the in-progress operation was started from within lazygit,
and only show the prompt in that case. We record this right after running
a merge/rebase step (in CheckMergeOrRebaseWithRefreshOptions, the
subprocess branch of genericMergeCommand, and the custom-command
conflict path), and clear it whenever a refresh observes that no
operation is in progress — which also handles an operation that was
finished or aborted externally.
The conflict-resolution tests start their operation by running git
directly (not through lazygit's UI), so they call the new test helper
Common.PretendMergeOrRebaseStartedInLazygit to have lazygit treat the
operation as its own and still get the prompt.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Replace the hard-coded side panel order, tab groupings, and window
assignments with values resolved from the gui.sidePanels config. The panel
order (SideWindows and the layout boxes), the tab strips (viewTabMap), the
per-context window names, each window's default view, and the jump-label
groups all now come from the config rather than from five separate
hard-coded lists.
A panel's window name is the name of its first tab, and panels not listed
in the config get their own window name so their views stay hidden instead
of overlapping a visible panel. Three small lookups translate config names
into views, tab titles, and contexts; a test keeps them in sync with the
set of valid names. The lookups are split this way (rather than one
resolver) because configureViewProperties runs before the context tree
exists, so the title/view lookups must not depend on it.
The config is applied to a repo's contexts via applySidePanelConfig on
every repo entry, including the cached-repo path: a repo's per-repo config
can differ from the previously visited one's, so each repo's contexts must
be (re)assigned from its own config rather than kept from when they were
first built.
With the default config this reproduces today's layout exactly.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The three branches of sidePanelChildren each spelled out the five side
windows by name, so the panel order lived in three places and the
status/stash sizing special-cases were tangled into positional literals.
Map each branch over one `windows` slice instead, and fold the
normal-height special-cases (status's fixed height, stash's
collapse-unless-focused) into a single per-window function. Behavior is
unchanged; this isolates the ordering so it can later come from config.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When restoring the commit selection after a refresh we match by hash and
TODO status. The TODO status is part of the match so that a commit being
reverted or cherry-picked is matched to the real commit rather than to
the rebase TODO entry that shares its hash.
But a selected commit can also change its TODO status across a refresh:
when starting an interactive rebase that stops to edit it, the real
commit becomes a TODO entry. Fall back to matching by hash alone when
there is no exact match, so the selection is still restored in that case.
The next commit relies on this to remove bespoke selection-restoration
code in the local commits controller that matched by hash alone, which
the generic mechanism otherwise wouldn't fully replace.
With the recently added external change detection, it happens more often
now that we refresh the commits list because an agent made a commit in
the background. In this case, if we keep the selection index the same,
it now points at a different commit, making the main view show a
different commit too, which is confusing and annoying. To fix this,
track the selected commit and range anchor by hash before reloading,
then restore those rows if both hashes still exist. This also allows us
to get rid of some bespoke code that did this for the specific cases of
reverting a commit or cherry-picking commits, because those are now
handled by the generic mechanism.
Preparation for the next commit, which selects the newly created commit
after a commit succeeds, while leaving the selection alone on failure.
For now success and failure use the same refresh options, so behavior is
unchanged.
We set GIT_OPTIONAL_LOCKS=0 for every git command we run. That env var
only affects `git status`: it tells git not to take the optional lock it
would otherwise use to write the index back after refreshing the cached
stat information. The intent was to avoid contending for index.lock with
git commands the user runs in a terminal.
The downside is that our `git status` never persists the refreshed
stat-cache. So whenever the working tree's cached stat info goes stale
(e.g. editing files and discarding the changes, or a checkout), every
subsequent status re-hashes the affected files to confirm they're clean,
and stays slow until something else writes the index (such as the user
running `git status` in a terminal).
Fix this by only suppressing optional locks for refreshes that run
unattended in the background; foreground refreshes triggered by a user
action now run a plain `git status` that writes the refreshed index back,
just like the command line does. Background refreshes keep passing
--no-optional-locks so they still can't cause lock contention.
RefreshOptions gains a Background flag that the background routines set,
threaded down to the status command.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the storage and snapshot-update half of the external-change-detection
mechanism. RefreshHelper now keeps a mutex-protected snapshot string and exposes
accessors for it; Refresh captures a fresh snapshot at the start of any refresh
whose scope set includes COMMITS or BRANCHES.
We capture before reading the git state, not after. Capturing after would let an
external change that lands between the git state read and the snapshot (say, the
next step of a rebase running in another terminal) leave the stored snapshot
newer than what we actually rendered; the poller would then see no difference
and never refresh again, stranding the UI on the intermediate state. Capturing
first keeps the snapshot from running ahead of the render, so if disk moves
during the refresh the next poll catches it.
No reader of the snapshot exists yet — the polling goroutine that consumes it
comes in a later commit. Keeping the snapshot hook in its own commit isolates
the invariant that the snapshot stays in sync with what the UI has observed,
which is what makes the poller's change-detection predicate work across in-app
commands and focus-in refreshes.
Several downstream conditions in Refresh() relied on multi-scope predicates to
express "if X is in scope, Y also needs refreshing". This makes it hard to add
new code that needs to ask "does this refresh re-read refs?", because the answer
involves mirroring one of those predicates and keeping them in sync forever.
Expand the co-refreshing relationships once, up front, right after the scope set
is built. The downstream conditions then collapse to single-scope checks against
the (now-expanded) set. Behavior is preserved.
Two of the scattered multi-scope conditions are intentionally left as-is because
they express subsumption rather than co-refresh (one branch already does the
work of another internally — expanding would cause double-refresh), and one
expresses mid-function coupling on a flag set inside the COMMITS/BRANCHES block.
Several commands (rewording or amending an earlier commit, custom patch
operations, etc.) are implemented by starting an interactive rebase that stops
at a commit, amending it, and continuing. When no conflict occurs, the user
isn't meant to notice a rebase happened at all.
But a background file refresh can fire while the rebase is mid-flight and render
a dirty working copy of whatever the behind-the-scenes rebase is doing (e.g.
applying a custom patch).
To fix this, we pause the background routines for the duration of any
waiting-status operation — exactly the window in which lazygit is driving the
git operation itself and will refresh once at the end. The boundary is also
right for the conflict case: when a rebase stops on a conflict the operation
returns, the pause releases, and background refreshes resume for the interactive
resolution that follows.
When a user switches into a repo whose .envrc hasn't been approved with
`direnv allow`, the previous behavior was to drop a "blocked" error
popup and leave the user to fix it externally. That meant opening a
terminal, running `direnv allow`, and then either restarting lazygit or
switching repos and back to refresh the env — easy to get wrong, easy
to forget.
When `direnv export json` exits non-zero, follow up with `direnv status
--json` to ask direnv whether the current directory has a not-yet-
allowed .envrc, and if so, get its path. Then show a confirmation popup
with the .envrc contents inline so the user can read what they're
approving. Confirming runs `direnv allow <path>` and re-runs the load
so the new env reaches subprocesses immediately; cancelling leaves the
env unloaded (the same state as before this commit when direnv refused
to load the .envrc).
Using `direnv status --json` instead of parsing the "is blocked"
stderr line means we rely on direnv's structured output rather than
its human-readable error format, which is more stable across versions
and avoids assumptions about output formatting.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When a user opens a repo from the recent-repos menu or jumps between
worktrees inside lazygit, only the env vars present at process startup
reach subprocesses. That breaks pre-commit hooks and other tools whose
dependencies are pulled in by a per-repo .envrc — users were left with
read-only operations because the env their shell would normally load via
direnv never made it into lazygit's git invocations.
Shell out to `direnv export json` after each chdir and apply the JSON
delta via os.Setenv/Unsetenv. direnv tracks the previous load in its own
DIRENV_DIFF env var, so the delta also unloads vars from the old repo
when entering one without a matching .envrc. If direnv isn't on PATH the
call is a no-op, so users who don't use direnv pay nothing and users who
do need no config to opt in. Any stderr direnv emits (loading messages,
"blocked .envrc" errors, etc.) goes to the command log.
The integration test puts a fake direnv on PATH and asserts that a value
it exports reaches a custom command after switching repos. Wiring this
up needed runner.go to support `{{actualPath}}` placeholders in
ExtraEnvVars, mirroring the existing support for ExtraCmdArgs, so the
test can prepend a fixture-relative directory to PATH.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
If we return the error here, we don't switch repos, but the chdir
happened already, so this would be an inconsistent state (a lot of
lazygit's code assumes that the current directory is always the worktree
root). Only log the error; failing to record the current directory is
not the end of the world.
Also, it is very unlikely to happen; RecordCurrentDirectory only writes
to a small file, and if this fails, then either there is filesystem
corruption of the disk is full, and in both cases the user likely has
much bigger problems.
Commit 4f0393f97b caused a regression: for operations that use
WithWaitingStatusSync (examples are squashing fixups, moving commits up or down,
cherry-picking, creating fixup commits, and more), the waiting status wouldn't
show during the operation; however, it would show after the operation was done,
and then linger forever.
The cause: since 4f0393f97b, layout sizes the bottom line from the actual
content of the AppStatus view rather than from the status manager. The async
render path keeps the view in sync (it sets the buffer on the first tick and
clears it to "" when the status ends), but the sync path used by
WithWaitingStatusSync did not:
- It called ForceLayoutAndRedraw before writing anything to the view, so layout
saw an empty buffer and left no room; the status never appeared during the
operation.
- When the operation finished it just broke out of the loop, leaving the last
spinner frame in the buffer. Every subsequent layout kept reserving room for
that stale content, so the status stuck around forever.
Fix this by writing the status into the view before the initial layout, and
clearing it again when stopping.
Until now every keybinding config field was a plain string. That meant a user
couldn't ask for two keys to invoke a command — the config silently accepted
only one form.
Convert every string-typed field across all 13 KeybindingXxxConfig structs to
Keybinding so the union type extends to every command. Defaults wrap their
single-key value in Keybinding{...} so the generated Config.md still renders one
scalar key per binding.
The alt fields keep their separate Binding registrations for now: this commit
does not yet introduce the merge mechanism that folds them into the main field —
that comes in a follow-up. Consumers previously calling opts.GetKeys on a string
field now call opts.GetKeys on the Keybinding, or take .String() / Keys[0] where
a single value is needed.
Adds a Keybinding.String helper for rendering, schema-generator work that
inlines the Keybinding union into each consuming property, and a unit test
covering the user-facing scalar/sequence YAML forms for quit.
This is a pure refactor in preparation for letting users configure multiple
alternate bindings for a single command. Every Binding still has exactly one
key, so nothing changes visibly: the cheatsheet, the on-screen options bar,
and the keybindings menu all render identically.
When a Binding ends up with multiple keys, the on-screen options bar will
show only the first (to avoid clutter); the cheatsheet will show all of them (in
a later commit). For now both paths take Key[0].
MenuItem.Key is changed in the same way, it also has a slice of keys now.
In this commit we keep the name `Key` in Binding, KeybindingOpts and MenuItem,
instead of renaming them to `Keys` right away, in order to keep the diff a bit
more readable. We'll do the rename separately in the next commit.
Constructing a menu item key from a literal character requires
gocui.NewKeyRune('r'), which is a bit noisy. Add a private menuKey helper in
both the controllers and helpers packages so the common case in either reads as
menuKey('r'). Duplicating the one-liner is cheaper than a cross-package import
dependency and avoids forcing every controller file to qualify the call.
The reason for doing this now is that we are going to change MenuItem.Key to a
slice of keys later in the branch, which means we'd have to add `[]gocui.Key{`
at each call site, making them even more noisy. With the menuKey helper we can
just change its signature and leave all clients unchanged.
AutoForwardBranches relies on the worktree model to skip any branch
that's currently checked out in another worktree (so we don't update
its ref behind the worktree's back). The post-fetch refresh wasn't
including the worktrees scope, so any external change to the worktree
list between lazygit's startup and the fetch — a `git worktree add`,
a `git checkout` in a linked worktree, a branch rename — left the
in-memory model stale and the skip check returned false negatives.
Add WORKTREES to the post-fetch refresh scope when auto-forwarding is
enabled. We gate on the config so users with auto-forward disabled
don't pay for an extra `git worktree list` plus per-worktree rev-parse
on every fetch tick.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The post-fetch logic was duplicated in `backgroundFetch` and the manual
fetch handler: refresh a fixed set of views, then auto-forward branches
if the fetch succeeded. The two had already drifted on the refresh
scope; folding them into a single helper makes the duplication go
away and prevents it from drifting again.
Pass the fetch error through so we preserve the previous behaviour of
refreshing unconditionally but only auto-forwarding on success.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
If `git worktree list` fails, we want the Worktrees model to fall back
to an empty slice so callers iterating over it stay correct. The error
branch was setting it to `[]`, but the line below unconditionally
overwrote it with the nil `worktrees` value from the failed call.
Use an else branch so the empty-slice fallback actually sticks.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The branches-panel PR icons only worked for github.com remotes. There was no
fundamental reason — the auth library we already vendor (cli/go-gh) supports
enterprise tokens out of the box (GH_ENTERPRISE_TOKEN, gh auth's keyring),
and the user-facing 'services' config has long been the documented way to
tell lazygit "this domain is a github service" for the View-PR-URL feature.
The fetcher just hardcoded github.com in three places:
- a substring check on the remote URL to decide we're "in a github repo",
- the GraphQL endpoint (always api.github.com/graphql), and
- the auth lookup (always against the default host).
Plumb the resolved web domain through instead. Detection now goes through
the hosting_service ("is this remote's provider 'github'?"), which means a
user with services: { 'git.acme.com': 'github:git.acme.com' } configured
gets PR icons on their GHE remotes too.
Replacing the substring check with a provider check also tightens a latent
bug in getGithubRemotes: it previously accepted any remote whose URL parsed
with the default regex, including gitlab and bitbucket — masked today only
by the InGithubRepo gate, but exposed once the gate goes away.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
After a successful paste DidPaste is true, hiding the "X commits copied"
indicator but leaving the buffer populated. From the user's perspective
this looks like a clean slate, so a new shift+C should start fresh. It is
important to reset DidPaste first, before populating the buffer with the new
commits, because otherwise each loop iteration would overwrite the previous one
since Add() rebuilds the set via SelectedHashSet() which returns empty while
DidPaste is set.
Now that HandleRender no longer does an implicit Render(), we can use it inside
OnUIThreadContentOnly to save performance, on the assumption that rerendering a
view that contains an inline spinner never changes the layout.
refreshBranches runs on a worker goroutine and re-renders the commits view
directly to refresh the branch-head visualization. As with refreshView, this
currently only flushes because HandleRender ends with self.c.Render(). Wrap the
explicit HandleRender call (along with the LocalCommitsMutex pair around it) in
OnUIThread so it keeps working once Render() is removed.
An async refresh dispatches refreshXyz on a worker goroutine, which then calls
refreshView -> PostRefreshUpdate -> HandleRender. Today the final
self.c.Render() inside ListContextTrait.HandleRender is what triggers a UI flush
from the worker. We're going to remove that Render() call, so prepare by
wrapping refreshView's body in OnUIThread.
This moves the entire rendering of the view (and the ReApplyFilter/ReApplySearch
stuff) to the UI thread, not just the layout. I don't expect this to make a
difference in practice, and it is already one step towards my long-term goal of
moving all view rendering to the UI thread (see
https://github.com/jesseduffield/lazygit/issues/2974#issuecomment-1729154768).
Unrelated to this branch, just because we're touching this code: there's little
reason to set the color on the background thread but the text on the UI thread.
Set them both together on the UI thread. Avoids a data race (unlikely to be a
problem in practice, we're talking about a 64-bit int, but still).
In 0d195077e4 we improved the performance of the status bar spinner by
avoiding a layout. This is fine from one spinner tick to the next, but it's a
problem when spinning starts or ends (or in the hypothetical case that the
status text changes in the middle of the operation, which we never do in
lazygit, but theoretically could). In this case a layout is needed so that the
rest of the status bar gets pushed over appropriately (or moves back to the left
when the spinner ends), and also so that the bottom line is shown or hidden
properly for users who set gui.showBottomLine to false.
To fix this, keep track of the status string width and force a layout whenever
it changes. This includes the beginning and end of an operation when it changes
from empty to non-empty or vice versa.
There is currently no observable misbehavior from this bug, but that's only
because we must have a HandleRender call somewhere that forces a full layout
when an operation starts or ends. We will remove the Render() call from
HandleRender at the end of this branch, at which point the misbehavior would be
visible if we didn't fix it here.
This bundles the keyName and a rune, so that we don't have to pass these around
separately everywhere. This should make it easier to swap out the rune for a
string when we upgrade to tcell v3.
I copied all files except dot files (.github and .gitignore), the _examples
folder, and go.mod/go.sum.
At some point we may want to copy the files back to the gocui repo when other
clients (e.g. lazydocker) want to use the newer versions of them.
If the repo has multiple remotes, but only one of them is on Github (the others
might for example point to a self-hosted Critic server or something like that),
lazygit would still present a menu to choose the remote for pull requests, but
it would contain only that single entry. That's pointless, pick it automatically
without prompting.
We add some tests while we're at it; these wouldn't have caught the problem,
because they only test getGithubBaseRemote which already takes the filtered
github remotes. It's still better than not having any tests; the real issue
could only have been caught with an integration test, which we don't bother
adding.
If we used to have a base remote (e.g. because there was only one), but now we
don't (e.g. because the user has added a second one, and we don't know yet which
one is the base), stop showing the cached PRs. This is especially important in
the case that the user cancels the prompt, in which case we don't have a base
remote for the rest of the session, but would keep showing the (potentially
stale) cached PRs forever.
This is useful when cancelling out of the commit panel mid-sentence (after
having typed the space for the next word); when entering the commit message
panel again, the space was gone and you had to type it again. Small thing, but
it just seems better to resume the panel in exactly the state that you left it
in. (Which we actually don't do; we don't remember the cursor position, or which
of the subject/description panels was active. That would be a separate
improvement.)
The save path and the load path used to be asymmetric. On save, the textarea
getters applied strings.TrimSpace, which stripped any leading blank lines, a
trailing newline after the cursor, or indentation on the very first line of the
description — all of which are legitimate user content. On load,
SplitCommitMessageAndDescription did TrimSpace on the description as well, and
the preserved message was routed through that same git-format split because
HandleCommitPress passed it as OpenCommitMessagePanel's InitialMessage. The
result: every round-trip through "escape and reopen" silently mutated the
message.
The fix is to treat our own preservation file as its own format, distinct from
git's canonical "summary\n\nbody" format:
- The textarea getters return raw content. strings.TrimSpace moves to the one
place that still needs it: the empty-summary check in HandleCommitConfirm (git
itself strips trailing whitespace and blank lines, so no pre-trim is needed
before -m).
- SplitPreservedCommitMessage / SetPreservedMessageInView split on the single
"\n" our Join uses, without any trimming — truly lossless.
- SplitCommitMessageAndDescription keeps its git-format behavior but replaces
TrimSpace with TrimPrefix("\n"), so it strips only the blank-line separator
and leaves body indentation intact.
- HandleCommitPress now mirrors HandleWIPCommitPress: it no longer passes the
preserved message as InitialMessage. OpenCommitMessagePanel resolves the
preserved content itself, uses it for display via the preservation-format
setter, and stores it as the initial message so the close-time "did the user
change anything?" check still correctly detects a cleared panel.
- GetInitialMessage no longer trims. With raw getters on both sides of the
comparison, trimming here caused spurious non-matches (e.g. for preserved
content with trailing whitespace). The original motivation — matching a
"WIP: " prefix with trailing space — works unchanged.
- UpdateCommitPanelView becomes dead code and is removed; its one remaining
caller (history cycling, always git-format) goes directly through
SetMessageAndDescriptionInView.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
We'll want to call just the preservation step when the app is quitting
with the commit message panel open: at that point the view-hiding and
context-popping in CloseCommitMessagePanel are pointless (the app is
about to exit), and some of the focus-lost side effects could be
surprising during teardown.
There's no reason for this; when opening the panel next time, it will be
populated again. Also, it doesn't make sense to do this only when not preserving
the message.
When no base repository is configured using gh, we have several heuristics for
choosing a likely base repo before prompting the user. One of these was that if
the origin remote exists, we'd choose that. This may be the right choice in some
cases, but in many others it's not; for example, in the common scenario of a
forking setup where the upstream repo's remote is called after its owner, and
the fork remote is called origin, we would choose origin as the base repo, which
results in not showing any PRs. Don't do that, and instead prompt the user for
the base repo in this case.
This makes it easier to recover them if they were deleted accidentally.
We only do this for local branches for now, we don't bother for remote branches;
it would be trickier there, because the branch on the server might not actually
point to the same hash as our local remote tracking branch does.
For esthetic reasons, checking out a branch (or other ref) blocks the UI until
the refresh is done, so it's important that the refresh doesn't do unnecessary
work. Refreshing pull requests is unnecessary (but costly, when waiting for it)
when a branch is checked out that already existed locally. However, it is
required when checking out a remote branch for the first time, so that the PR
icon appears immediately when there is one.
When you rebase a branch and there are conflicts, lazygit asks you to continue
the rebase when it detects that all conflicts have been resolved. However, it is
common to make additional changes beyond just fixing the conflicts (e.g. to fix
build failures), and when doing that while the "Continue the rebase?" prompt is
showing, lazygit detects that too and asks you if you want to stage those newly
modified files too. This is all well and good (and can be disabled for those who
don't like it); however, lazygit would treat out-of-date submodules as unstaged
changes and would offer to stage those as well, and this is pretty bad and
almost never what you want. Fix this by excluding submodules from that second
check.
Previously, when showing subcommits of a branch or tag, scrolling down, and then
going out and entering subcommits again, it would select the first commit but
keep the previous scroll position, so that the selection was not visible. Always
scroll to the top to fix this.
I got confused when I investigated the bug that was fixed in the previous
commit, so hopefully these will be helpful for the next person reading this
code.
The symptom of the bug was that confirmation panels with wrapped text were one
character too wide, which could sometimes result in them being one line too
heigh. One concrete example is the "Discard file changes" confirmation that
appears when pressing `d` in the commit files panel.
I got very confused when investigating this bug, and the reason is that
getPopupPanelDimensionsAux took two parameters panelWidth and panelHeight, but
the second was actually contentHeight. Rename it to contentHeight, and change
the first one to also be the content width rather than the panel width. This is
easier for clients, and less confusing.
Without this check, the selection was being reset to 0 whenever
ReApplyFilter was called for the current filter context, even when
the user wasn't actively typing in the search prompt (e.g. when the
model updates in the background). This was causing unexpected cursor
jumps.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When opening a filter prompt, reset PrevSearchIndex to -1 to avoid
stale search state from a previous search/filter session.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This avoids a naming collision with GetFilter from the
IFilterableContext interface, which will be implemented by
FileTreeViewModel in the next commit.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
On very large screens they would get ridiculously wide.
The limits we chose here are a little arbitrary:
- 80 for confirmations and prompts
- 90 for menus
- auto-wrap width plus 25 for the commit message editor (or 100 if auto-wrap is
off)
It is possible to scroll the selection out of view using the mouse wheel; after
doing this, it would sometimes scroll into view by itself again, for example
when a background fetch occurred. In the files panel this would even happen
every 10s with every regular files refresh.
Fix this by adding a scrollIntoView parameter to HandleFocus, which is false by
default, and is only set to true from controllers that change the selection.
We want to do this whenever we switch branches; it wasn't done consistently
though. There are many different ways to switch branches, and only some of these
would reset the selection of all three panels (branches, commits, and reflog).
The confirmation used to make sense back when the Open MergeTool command was its
own top-level command; however, that command was changed in 703f053a7e to open a
menu instead, and Open MergeTool is now just a submenu entry in that menu, so it
no longer needs a confirmation.
We move the code to push the branches context into CheckoutRef, this way it
works consistently no matter where we call it from. Previously, checking out
remote branches or tags would switch to the branches view, but checking out a
commit did not.
Note that it now also takes effect for undoing or redoing a checkout, which may
be a bit questionable; but I still think it makes sense for this, too.
This doesn't really solve a pressing problem, because I guess it's unlikely that
users add spaces at the beginning or end of what they type into a prompt; but it
could happen, and in this case we almost always want to strip it. Just adding
this here for completeness while I was working on this code.
The only exception is the input prompt of custom commands, because who knows
what users want to use that input for in their custom command.
Most of our prompts don't (shouldn't) allow empty input, but most callers didn't
check, and would run into cryptic errors when the user pressed enter at an empty
prompt (e.g. when creating a new branch). Now we simply don't allow hitting
enter in this case, and show an error toast instead.
This behavior is opt-out, because there are a few cases where empty input is
supported (e.g. creating a stash).
Previously it was used both for the Confirm handler and the Cancel handler, as
well as for the Confirm handler of confirmation popups (not prompts). There was
no other way to do it given how wrappedConfirmationFunction was shared between
all these; but now there is. The logic is really only needed for the Confirm
handler of prompts.
This doesn't fix anything, it just makes things clearer.
And call this new helper function from both wrappedConfirmationFunction and
wrappedPromptConfirmationFunction; this gives us more flexibility to do
different things in each of those.
The existing diff parser incorrectly treated subsequent lines beginning
with "---" as filename headers while processing hunks. This caused
corruption when dashed lines appeared within diffs themselves.
Restrict filename detection to only occur between hunks.
This will put whatever git's default merge variant is as the first menu item,
and add a second item which is the opposite (no-ff if the default is ff, and
vice versa).
If users prefer to always have the same option first no matter whether it's
applicable, they can make ff always appear first by setting git's "merge.ff"
config to "true" or "only", or by setting lazygit's "git.merging.args" config to
"--ff" or "--ff-only"; if they want no-ff to appear first, they can do that by
setting git's "merge.ff" config to "false", or by setting lazygit's
"git.merging.args" config to "--no-ff". Which of these they choose depends on
whether they want the config to also apply to other git clients including the
cli, or only to lazygit.
- Squash and FastForwardOnly are mutually exclusive, and instead of asserting
this at runtime, model the API so that they can't be passed together.
- FastForwardOnly is unused, so remove it; however, we are going to need --ff
and --no-ff in the next commit, so add those instead.
- Instead of putting the enum into the MergeOpts struct, replace the struct by
the enum. We can reintroduce the struct when we add more arguments, but for
now it's an unnecessary indirection.
Stashing doesn't affect submodules, so if you have a working copy that has
out-of-date submodules but no other changes, and then you revert or paste a
commit (or invoke one of the many other lazygit commands that auto-stash, e.g.
undo), lazygit would previously try to stash changes (which did nothing, but
also didn't return an error), perform the operation, and then pop the stash
again. If no stashes existed before, then this would only cause a confusing
error popup ("error: refs/stash@{0} is not a valid reference"), but if there
were stashes, this would try to pop the newest one of these, which is very
undesirable and confusing.
Whenever git returns the error "The previous cherry-pick is now empty", we would
previously continue the rebase; this works for rebase because it behaves the
same as "git rebase --skip" in this case. That's not true for cherry-pick
though; if you continue a cherry-pick where the current commit is empty, it will
return the same error again, causing lazygit to be stuck in an endless loop.
Fix this by skipping instead of continuing; this shouldn't make a difference for
rebase, but works for cherry-pick.
Theoretically we could have a similar problem for revert (if you are trying to
revert a commit that has already been undone through some other means); this
should then be fixed in the same way with this change. However, the change is
not relevant for revert because git returns a different error in this case.
When focusing the main view, going into full screen mode by pressing '+' twice,
and then opening the search prompt ('/') or a menu (e.g. '?' or ':'), the full
screen display would switch to the focused side panel.
Fix this by always excluding popups from the window arrangement logic. No popup
should ever have any influence on how the views beneath it are laid out.
This is likely to do bad things; for example, if the prompt is the shell command
prompt, then we would run into what looks like a deadlock bug in tcell. In other
cases, the characters in the following lines might be treated as random commands
after the prompt is confirmed.
Replace merge-tool with merge options menu that allows resolving all
conflicts for selected files as ours, theirs, or union, while still
providing access to the merge tool.
Previously, the feedback you got when pressing "-" was just a "Checking out..."
status in the bottom line. This was both easy to miss if you are used to looking
for an inline status in the branches panel, and it didn't provide information
about which branch was being checked out, which can be annoying in very large
repos where checking out takes a while, and you only see at the end if you are
now on the right branch.
Improve this by trying to figure out which branch was the previously checked out
one, and then checking it out normally so that you get an inline status next to
it (as if you had pressed space on it). There are cases where this fails, e.g.
when the previously checked out ref was a detached head, in which case we fall
back to the previous behavior.
So far, confirmations and prompts were handled by the same view, context, and
controller, with a bunch of conditional code based on whether the view is
editable. This was more or less ok so far, since it does save a little bit of
code duplication; however, now we need separate views, because we don't have
dynamic keybindings, but we want to map "confirm" to different keys in
confirmations (the "universal.confirm" user config) and prompts (hard-coded to
enter, because it doesn't make sense to customize it there).
It also allows us to get rid of the conditional code, which is a nice benefit;
and the code duplication is actually not *that* bad.
In some cases we set a disabled reason but leave the text empty, so that we
don't get an error toast when the item is invoked. In such a case it looks
awkward if there is a tooltip showing "Disabled: " with no following text.
When showing a confirmation whose text ended with a line feed, we would make the
popup panel one line less tall than it needs to be, so it would show a scroll
bar. One example where this occurred is the very first popup that users ever see
(the "seriously you rock" message for new users); that's a pretty bad first
impression.
This happens because our code to suppress trailing newlines in views doesn't
work for styled text, and the text in confirmations is bold. This code checks if
the last character of the text is a line feed, and in this case it isn't; the
escape sequence for turning bold off comes after it.
We should really fix this by improving that mechanism, but this would require
some tricky logic, so as a quick fix, trim trailing (and leading) linefeeds from
the text that we display in a confirmation, before making it bold.
This makes it easier to use the full ref in the git merge-base call, which
avoids ambiguities when there's a tag with the same name as the current branch.
This fixes a hash coloring bug in the local commits panel when there's a tag
with the same name as the checked out branch; in this case all commit hashes
that should be yellow were painted as red.
Also, fix two other commands that stage all files under the hood:
- when continuing a rebase after resolving conflicts, we auto-stage all files,
but in this case we never want to include untracked files, regardless of the
filter
- likewise, pressing ctrl-f to find a base commit for fixup stages all files for
convenience, but again, this should only stage files that are already tracked
When filtering for a file path we use the --follow option for "git log", so it
will follow renames of the file, which is great. However, if you then selected
one of the commits before a rename, you didn't see a diff, because we would pass
the original filter path to the "git show" call.
To fix this, call git log with the --name-status option when filtering by path,
so that each commit reports which file paths are touched in this commit;
remember these in the commit object, so that we can pass them to the "git show"
call later.
Be careful not to store too many such paths unnecessarily. When filtering by
folder rather than file, all these paths will necessarily be inside the original
filter path, so detect this and don't store them in this case.
There is some unfortunate code duplication between loading commits and loading
reflog commits, which I am too lazy to clean up right now.
I don't know why this function argument was added, but I don't like unnecessary
indirections, so I'm removing it as SubCommitsHelper has access to everything it
needs to do it itself.
Recycle reflog commits only for the non-filtered ones.
If we're not filtering, FilteredReflogCommits and ReflogCommits are the same. If
we then enter filtering and get reflog entries again, and pass a
lastReflogCommit, we'd recycle the previous FilteredReflogCommits, which are
unfiltered, so that's no good. Work around this by simply never using the
recycle mechanism when getting the filtered reflog commits.
We could do better by remembering what the last filter path or author was, and
only suppressing the recycling when it changed; but that's more complexity than
I want to add, so hopefully this is good enough.
I can only guess what happened here: in aa750c0819, this code to manually load
the reflog commits was added, to make sorting branches by recency work when the
reflog is filtered by path. At that time we didn't have separate ReflogCommits
and FilteredReflogCommits models yet. Then, FilteredReflogCommits was introduced
(in 8822c409e2), probably for the very purpose of being able to get rid of this
again; but then it was forgotton to actually get rid of it.
The funny thing is that the introduction of FilteredReflogCommits happened in
the very next commit, 15 minutes later.
Since filtering switches to half-screen mode in the local commits panel, most
people probably didn't notice, but we do also filter those other views. So when
leaving half-screen mode (but not filtering), you could switch to sub-commits or
stashes, and those would show the filtered view only after the next refresh
(e.g. after a background fetch). It's worse when leaving filtering, because this
goes back to normal screen mode, and you would often see an empty stashes panel
after that (until the next background fetch), which is quite confusing.
I also find it questionable to always switch focus to the commits panel when
entering filtering. If it is initiated from subcommits, reflog, or stashes,
maybe we want to stay there. I'm not changing this now since I'm unsure how much
people rely on the current behavior.
When exiting filtering mode while the focus is not in the local commits panel,
the call to HandleFocus would render the selection in the commits panel as if
the panel had the focus.
The call to HandleFocus was introduced recently in 4981419ba9 in an attempt to
rerender the main view correctly, but it should only have been called when local
commits is the focused context. But instead, we can use PostRefreshUpdate, which
handles this distinction.
When entering filtering we would only call FocusLine, which takes care of
highlighting the selected line in the commits list, but not of re-rendering the
main view. HandleFocus does that.
When exiting filtering, the HandleFocus call was missing entirely.
The tests needed to be reworked a little bit to make this testable.
At the same time, we change the defaults for both of them to "date" (they were
"recency" and "alphabetical", respectively, before). This is the reason we need
to touch so many integration tests. For some of them I decided to adapt the test
assertions to the changed sort order; for others, I added a SetupConfig step to
set the order back to "recency" so that I don't have to change what the test
does (e.g. how many SelectNextItem() calls are needed to get to a certain
branch).
When toggling the value in the UI we simply overwrite the value in UserConfig;
this would be bad if there was ever a chance that we want to write the user
config back to disk, but it is very unlikely that we can do that, because
currently we have no way to tell which parts of the config come from the global
config file and which ones come from a repo-local one.
