Delta and difftastic both rendered the preview normally with the trees a
line ending apart, so only the warning was ever visible. The setting
behind it was core.autocrlf in the system config Git for Windows
installs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The lint failure at PR 6's tip came out of the 2026-09-19 replay, and the
fixups for it sit on two PRs. Write down why the wrapper leaves PR 6 and
comes back in PR 7, and that the PR 7 commit between the two fixups is
meant to stay unbuildable until they are folded in.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The click names its own line, so it needs none of the gate the edit
keybinding needs, and for a modify/delete conflict the file it opens is
in the working tree.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The branch's opening commit belongs in PR 5, and moving it there proved
much cheaper than the note in the plan had assumed. Four more findings
needed code, two of them in gocui.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A fixup on "Show git's own diff when the renderer's can't be acted on"
would have folded two decisions into one commit: which diff to render,
and whether the content can be pointed at.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two problems from testing PR 9: wrapLinesInDiffView governed everything
the two main panes showed, and the hint for a conflict that has to be
resolved by picking a side had a selection over it. Both are answered by
having a render say whether it holds the panel's diff.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two bugs reported from using the stack and a third found while reproducing the
first (PR 2 round 1, PR 5 round 5, PR 8 round 1), then two more rounds as each
fix in turn was tested and found short: the pane answered "is there anything to
select here?" one screenful too early (round 6), and then took the answer over
from the commit before while it still couldn't tell (round 7).
Round 6 carries the measurements from the reported commit, which say what each
fix is worth. §8 gains the diffstat row the user has since decided to keep, and
loses the row round 7 closes.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
PR 10 landed independently of the explorer removal and gained visible feedback for non-suspending editors. Capture its actual dependency, six-commit shape, pending terminal matrix, and the separately deferred click-selection behavior.
Co-authored-by: GitHub Copilot <noreply@github.com>
Four problems, the one root cause behind three of them, the prep refactor the
focus rule needed, and the two gaps left open. Also corrects a claim made
while the round was being scoped, that a diff renderer is re-laid-out
whenever the view width changes.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two defects and their fixes, the testing setup that reproduces them, and
the rebase traps the round turned up.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Record the four review comments the interactive pass came with, what was
decided for each, and where their fixups landed — three of them belong to
PRs 5 and 6 rather than to PR 7.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The staged-side branch grew out of it, PR 7 was rebuilt on top, and the
selection-visibility rule became a question about content rather than about
which kind of task rendered it.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
It sits below PR 7, which is rebased onto it and simpler for it: which side a
pane shows is now a property of the pane, and the focus-follow rule is
symmetric.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Seventeen deviations, of which the ones that bind later work are the split
seam (PR 8's panels satisfy the extension interface), the stale timing fact
about what a refresh has done by the time it returns, and the restore's new
Done hook.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two things bind later work: a restore keeps the task reading to the end of
its input, because a rendering that has to be parsed as a diff can only be
parsed whole, and identities match through a key in which every kind of
content line collapses together - which is what lets a change that becomes
a context line still count as the same place, and a hunk header not.
The interactive pass on PR 5 found nothing to fix; refining the behaviour
of a few special cases can wait until the remaining PRs are written.
Toggling "ignore whitespace" re-renders the diff just like a context-size
change or a renderer switch does, so it belongs in PR 6 as a third
consumer. It is the first one whose anchor can vanish along with its hunk
or its whole file, which invalidates the candidate walk's reason for
stopping at the first change line; the walk goes unbounded instead, so the
restore lands on the nearest surviving line wherever in the diff that is.
And once staging happens in the main view, master's refusal to ignore
whitespace while staging no longer catches it - parked as an open question
for PR 7.
The selection commit is split in four, comments no longer point at the
staging view, and the selection commands are described and enabled only
where they apply.
