Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
2 changes: 2 additions & 0 deletions Docs/DESIGN.md
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,8 @@ folder where a subject needs several. A file here is binding, not background.
| [design/files/modified-and-branch.md](design/files/modified-and-branch.md) | Modified against Branch, what Changes and file tabs count from, the Branch read and its pace, failures, no base branch, staged marks under Branch, merge base reads, scope switches, costs left. |
| [design/files/image-tabs.md](design/files/image-tabs.md) | Which files open as images, Before and After, the 32 MiB cap, decoding, find and keyboard on an image. |
| [design/files/review-comments.md](design/files/review-comments.md) | Comments inline in a file tab, Review and its Send, a file then a mention, the picker's default, a file per worktree, anchoring by quote, Outdated and Lines changed, scope, gaps in the row geometry and the split, the composer's draft, the command line. |
| [design/files/all-changes-tab.md](design/files/all-changes-tab.md) | The All changes tab: a third tab kind, its file list and lazy reads, a list of per-file diff views, the Changes list as its index, review cards in it, Split, its row cap |
| [design/files/all-changes-tab-reviewed.md](design/files/all-changes-tab-reviewed.md) | Review findings on the All changes tab settled without a change, and why. |
| [design/tabs-and-groups/groups-and-dragging.md](design/tabs-and-groups/groups-and-dragging.md) | Pane trees, groups, the active versus shown tabs, dragging a tab to another worktree or into a group, drop bands, springing back, the drag preview, live reorder, the keyboard following a dropped tab, refused drops, `tabs move`. |
| [design/tabs-and-groups/tab-strip.md](design/tabs-and-groups/tab-strip.md) | A strip out of room, scroll arrows, split buttons, sideways wheel scrolling, the New Tab plus menu, renames from the command line, file tabs in the strip, tab titles, the X button, the shown tab's hairlines, the no-tab placeholder. |
| [design/terminals/sessions.md](design/terminals/sessions.md) | A session's state and what clears it, Done and Failed, closing a tab and how a shell ends, notification banners, agents launched through the login shell, the captured login environment, auto-start and warming sessions, terminal editor tabs, each pane's PATH, inherited shell state, the helper hand-off. |
Expand Down
3 changes: 3 additions & 0 deletions Docs/design/agents/command-line.md
Original file line number Diff line number Diff line change
Expand Up @@ -229,6 +229,9 @@ The command line's requests on the reports' socket. Newest at the bottom.
first tab starting out of view as a held-back one does after its hook, and the
keyboard staying put. Added 2026-11-14 so a script can make several at once
without each pulling the view to itself.
- **`tabs open-all-changes` opens or turns to a worktree's All changes tab**,
and `tabs list` reports it as kind `all-changes`; `tabs step` and `tabs find`
reach it as a file tab (files/all-changes-tab.md).

## Reviewed and left as is

Expand Down
87 changes: 87 additions & 0 deletions Docs/design/files/all-changes-tab-reviewed.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,87 @@
# The All changes tab: reviewed and left as is

Findings from the review rounds on [all-changes-tab.md](all-changes-tab.md) that
were settled without a change, each with why. A review reads this first.

- **An opened fold is not moved when lines are added above it** (review round
2): it is kept by the file on disk's line numbers, as the bullet on folds
says, so an edit above shifts which lines show open. Re-anchoring folds as
comments are would cost a quote per fold for a run-only convenience.
- **Find carried twice when a read moves comments** (review round 3, unproven):
setting the comments searches again from the old text, and the read then
carries the result from it once more. A test inserting a line above a repeated
match moved the selection with no comment involved, so this showed nothing of
its own. Round 17 found the cause: the carry went by line number across files.
Round 18 tried a read moving a comment with one row and five above the match,
and the selection held (`aReadMovingCommentsKeepsTheFindSelectionOnItsMatch`).
- **The document's staging marks taken from its first file with rows** (review
round 3, unproven): under Branch a file in an embedded repository counts from
`HEAD` and marks staging while the worktree's own do not. No case has been
shown; each file's own rows carry its marks.
- **A file deleted after the tab read it keeps its last text** (review round 4):
its file tab does the same
(`aFileDeletedUnderAnOpenTabKeepsItsLastTextUnmarked`), and each file reads as
its tab would; read afresh it shows its lines removed.
- **A step's box dropped once the unread file it sat on is read** (review round
5, unproven): the next step counts from the top of the clip, which can step
back into a short file above. Not shown in a test.
- **"Fold" for unchanged lines hidden in a file** (review round 5): the design,
the mockups and `git` itself say fold; the Files panel's "folded" is case
folding, an adjective in another area, and renaming either is churn for no
reader's gain.
- **A file shown whole keeps its gap items at document rows in its slice**
(review round 6): no such file gets a fold or notice, being all one kind of
line, and rebasing by the document's rows would be wrong for a removed file,
whose Split rows do not count as its Unified ones do.
- **Each update regroups the worktree's comments once per built file** (review
round 7): about ten files and a few hundred comments make it a few thousand
cheap steps; caching the groups would need every comment write to keep them
fresh.
- **A scroll to a row can land short of it where files it builds grow by their
cards** (review round 9, unproven): heights measured a turn later move the
files below. Not shown in a test; the next scroll or step corrects it.
- **Spans shifted into document offsets and back when a file is coloured**
(review round 9): it runs once a parse, the parse kept by path; not measured.
- **Every update lays the list out again, estimating each unbuilt file** (review
round 9, not measured): an estimate is a lookup and a sum, and a layout
already runs on every scroll that builds a file.
- **Unproven in review round 11, each tried or traced and not shown**: a bare
caret scrolling a long line back to its left edge (a test kept the sideways
scroll); a file's wheel axis lock met mid-gesture keeping an old axis (rests
on AppKit sending one gesture to two scroll views); a pending caret rect
scrolling the list during a file's layout (none is pending once laid out).
- **Cmd+Up or Cmd+Down in a focused file scrolling the list twice** (review
round 12, unproven): the caret's move and the list's jump to its end were
counted in a test, and the list reported one scroll.
- **Every read lstats every listed file** (review round 13, not measured): the
signature also drops a submodule's folder from the list, and a few thousand
lstats run off the main thread.
- **A path listed twice merged both here and in `FileTree`** (review round 13):
the two differ on purpose, the tree keeping a nested repository's row that
this list leaves out, and each merge is two lines.
- **A row cap of zero or less makes the reader spin** (review round 16): only
tests pass a cap, and each passes a positive one.
- **The canvas's gutter and gap layers framed at the document's height** (review
round 17, unproven, not measured): a scroll redraws only what shows.
- **File-tab names on code the All changes tab shares** (review round 17):
`rerunFileFind(in: FileTab.ID)`, `FileTabReadLog` and the like take its ids
too; renaming every shared call touches dozens of files for no change in
behaviour, and one per-kind seam for them is a larger refactor than the tab.
- **Each read rebuilds the whole document and resolves its git bases again**
(review round 18, not measured): a scroll that brings unread files near reads
them in one batch, which 2 or more files make cheaper than reading alone.
- **Two git diffs a chunk, twice `runsAtOnce` processes** (review round 18):
measured on 6,000 changed files and 12 cores, 8 chunks took 0.307 s and 4 took
0.334 s, so the pair stays.
- **A `git check-ignore` refusal on a path inside a submodule drops the batch**
(review round 19, unproven): a scratch repository in round 16 showed the
reader leaves the submodule's folder out, so its paths never reach it.
- **Two bounded task pools, the reader's and the batched patches'** (review
round 19): the reader's also lands files in order against the row cap, so one
shared pool would need both shapes.
- **Find kept in step by each change to the layout** (review round 19): a rerun
inside `allChangesDocument(of:)` would search from a getter; the four callers
share `relayingOutAllChanges` or pass the old layout themselves.
- **`DiffColumn.hasPendingScrolls` read by a test alone** (review round 20): a
file in the list cannot scroll on its own, so no scroll shows a pending one;
the property is the only way to see `forgetShownFile` clear all three.
Loading
Loading