The replica framed to a font-derived nominal width, so faces in wide
lanes' interior masonry columns and narrow lanes dragged a preview of
the wrong size and the cursor could sit over empty space beside it.
CardFaceView's existing onGeometryChange now reports size, not height
alone - height still feeds LaneDropRegistry, width feeds the replica -
and BoardMetrics.cardReplicaWidth(measured:bodyPointSize:) keeps the
old figure as the documented not-yet-laid-out fallback. Same row, same
paddings, same lineLimit at the same width means the same height, so
the preview is now an exact overlay of the face it left.
Manual verification pending alongside the lane grab-point card: 3x vs
1x lanes, interior masonry columns, window re-divide, trash rows,
large text sizes, drag-at-creation fallback.
3 tests. 1661 green on both schemes.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
SwiftUI's .onDrag(_:preview:) exposes no anchor or grab-point API and
lays the preview centered over the grabbed view, so a full-height lane
replica grabbed by its ~1.5-line title bar hung half a lane above the
pointer. DragPreviewAnchor is the pure fix: transparent one-sided
padding computed so the padded image's center IS the replica's title
bar - the system's centering then lands the bar under the cursor and
the body over the lane it was lifted from. LaneView measures the real
bar height in the geometry observer it already runs, accounting for the
accent band when the lane's color resolves.
Pixel-exact grab preservation would need re-homing the whole gesture
stack onto an AppKit beginDraggingSession path; declined for a polish
card - this gets the cursor onto the grabbed bar under either of
SwiftUI's possible placement rules. Manual verification pending (no
display here): grab specific bar pixels, colored + uncolored lanes,
large text sizes, multi-lane fan, cross-board drop geometry.
9 tests. 1661 green on both schemes with the paired card-width fix.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Closes the gap found at m10: enterUnwritableLock existed with zero call
sites. WritabilityProbe classifies the cause volume-first - a board on a
read-only DMG is also permission-denied by access(2), and "you don't
have permission" would send the user to a Get Info panel that cannot
help - with a pure classify(volumeIsReadOnly:isWritable:) truth table
and a two-syscall probe that rebuilds its URL to defeat NSURL resource
caching. ReadOnlyLockReason.unwritableLocation now carries the cause;
BannerCenter phrases the two ("this board's volume is read-only" vs
"you don't have permission to change this folder").
The probe wires once in BoardStoreRegistry.acquire, immediately after
the store loads - every open path funnels through it, and running
before the loose-file relocation and agent-guide hooks makes the
skipped-with-log guide write true by construction (its isWritableFile
pre-check demotes to second line of defense). The board still opens:
lock, not refusal.
The reconciling re-probe is now symmetric per 02's settled text - a
volume gone read-only mid-session raises the lock at the next probe
(sibling locks settle first, so a root returning read-only lands the
honest lock); the stale "deliberately one-way" comment and its pinning
test are gone. Save as Template's carve-out predicate extracted to a
testable allowsSave (behavior unchanged); Duplicate stays disabled.
11 tests added. 1649 green on both schemes.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
User-ruled 2026-07-29: the ledger builds now in base, pre-release
(DESIGN/02 - Components - EchoLedger; DESIGN/10 - Live board
announcements). Receipts drop inside BoardWriter's four disk primitives
(atomic replace, folder move, removal, attachment copy) into a @TaskLocal
ledger that BoardStore.performWrite binds for the bracket's duration -
no call-site bookkeeping, and performWholesale deliberately binds
nothing per 02's bracket exemption. Classification is a pure function of
two snapshots: an item whose folder, index.md bytes, or attachment
listing differs is an observed change; disk matching the receipt is
app-mediated (receipt consumed), no receipt or mismatch is foreign.
Byte-identical foreign overwrites classify app-mediated (unobservable,
accepted); a foreign edit over a fresh app write classifies foreign.
The announcer now consumes per-file facts on every reload origin - the
WatchOrigin gate is gone (ReloadFacts.origin removed outright; nothing
read it after the gate fell). Reconciling sweeps announce their
receipt-less findings as foreign, closing both interim holes
(debounce-window absorption, reconcile silence). The vanishing-focus
rung gates on the ledger too: "deleted externally" would be a lie about
an app-mediated delete, and the subject's own verdict decides.
Divergence flagged: attachment imports hash the landed file right after
FileManager.copyItem rather than during the copy (the bytes do not
stream through the app); an unreadable read-back records nothing, the
direction that biases toward foreign.
30 ledger tests added, announcer suite reworked to the ruling. 1638
green on both schemes.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Three ruled behavior changes (DESIGN/10, resolution session 2026-07-29):
- The board-change digest covers the trash while View > Show Trash is on:
BoardDiff.between gains includingTrash, keying its card index by
ItemPath so foreign purges, restores, and Empty Trash join the digest;
crossings of the trash boundary still read deleted/restored, never
moved, on both sides of the toggle. BoardStore.land passes the store's
own isTrashVisible - no new injection seam.
- A vanished head with surviving co-selection is still named: naming and
recovery are independent axes, so BoardAnnouncer's vanished-focus rung
fires on all branches while the survivors-veto now gates only the
recovery half (recovery implies vanished, no longer both-or-neither).
- Banner-row buttons are literal FKA Tab stops: BannerRow.controls is
the row's testable button inventory, BannerRowView renders from it
with .focusable() on each button; the combined VoiceOver element stays
unconditional - custom actions and Tab stops are independent surfaces.
19 tests added, 2 expectations updated to the rulings. 1607 green on
both schemes.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The automated half: AccessibilityAuditTests runs performAccessibilityAudit
over all eight surfaces DESIGN/10 names — board with trash hidden and
shown, card window in Preview/Edit/raw source, welcome, template chooser,
board popover. One audit per test, .all audit types, no issue handler —
nothing waived; a future false-positive excusal must match one element on
one surface with its reason beside it. Navigation is menu-bar titles and
the arrow grammar; no accessibility identifiers added to production code.
The suite launches with --ui-test-fixture-board: the sandbox forbids
handing the app a temp-folder path (no bookmark behind it), so the flag
carries no payload and the app builds a known board inside its own
container through the ordinary BoardWriter door — three lanes, six cards,
a rich Markdown body with attachment, one card already in .trash/ — with
the registry redirected to the same scratch directory so audit runs never
pollute real recents. LaunchPlan replaces the restore Bool (welcome /
restoreBoards / uiTestFixture, fixture wins outright), decided once in
KanbanApp.init and dispatched by RestoreBootstrapView; pure and pinned by
UITestLaunchTests, and the fixture itself is materialized and read back
through BoardLoader in units — the only proof available headlessly.
The manual half: KanbanUITests/AccessibilityVerification.md is the one
document — the audit suite at the top (it needs a real display and
Accessibility permission), the per-release VoiceOver smoke script with
expected utterances quoted from AccessibilityPhrases, and the
consolidated m11 checklist from all four implementation cards.
1588 unit tests green, UI target compiles, both schemes build. The audit
run and smoke script await a real display — the manual pass is the
user's.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
DESIGN/10's ≥ 4.5:1 rule gets its owner. ContrastMath is the pure seam:
WCAG relative luminance (piecewise sRGB linearization), symmetric
contrast ratio, source-over compositing (an #RRGGBBAA board colour
resolves over the window background of the active appearance), and
inkChoice — native label if it clears AA, else the other appearance's,
else the higher ratio with meetsAA false (a mid-grey hex can max out
below 4.5 against both 85%-alpha labels; the app paints the best
available rather than overriding the user's colour). BoardTextInk is
the board's application: the decision is a ColorScheme, not a Color —
the text on the board fill is a hierarchy (.primary/.secondary/
.quaternary), and overriding the subtree's scheme moves the whole
vocabulary coherently. Recomputed on appearance change by construction
(read in body); label/backdrop colours resolve inside the asked-for
appearance, Increase Contrast variants included.
Two render sites — the only board text that sits on the user's colour:
the lane header (lanes draw no plate; title, icon, badge, rename field
and the + button land directly on the board fill) and the trash header
(its wash is ~5% effective alpha). Menus, popovers, and drag replicas
deliberately stay native; card faces carry their own opaque plates.
The card's premise fell during implementation: 03's "palette pairs
AA-verified at design time, pinned by a computed-contrast unit test"
was false — no such test existed, and the m4 path drew the native label,
failing AA in one appearance for all twelve wells (obsidian in Light
Mode: 1.0:1). Palette names now route through the same ink selection
(paintedColor delegates to Palette.nsColor — one predicate with
BoardView's paint decision), and PaletteContrastTests pins that the
chosen ink clears AA for every well in both appearances — plus
theNativeLabelIsNeverEnough, which would have failed on the m4 code.
Filed on the Redesign board for ratification. 1579 unit tests green,
both schemes build.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Full relative text scaling per DESIGN/10: BoardMetrics is the board
strip's geometry as a pure function of the body point size
(CardWindowMetrics' twin) — lane plate/header/band, card
corner/stripe/padding, masonry spacing, the drop model's nominal card
height, resize-handle geometry, trash hatch pitch, and both window
floors all derive from an em; CardFaceMetrics folded in. The two fixed
font sizes (welcome brand/glyph) went relative; the toolbar search
field is 17 ems like the transient bar's. The no-horizontal-scroll
invariant is pinned by test at six text sizes by twelve lane counts.
Accommodations is Motion's sibling for the visual settings: Increase
Contrast adds a flat point to strokes (monotone, hierarchy-preserving),
gives borderless card/lane plates a resting separator hairline, and
takes faded accents to full alpha; Reduce Transparency turns the
transient search bar's glass solid and does the same for the alpha
washes that composite over a user-chosen board background (trash plate,
hatched header, drag shadow). Reduce Motion audited — every animated
surface already routes through Motion with a reduced variant; no gaps.
Full Keyboard Access: the template chooser's tiles were pointer-only —
now focusable, arrow-navigable (clamped, StyleWellGrid's rule), Space
picks, Return stays the sheet's default action, focus names the
selection one-way. The board's single tab stop shows its focus ring
under FKA (focusEffectDisabled inverts). Style editor verified already
conformant. Edge accents verified text-free; trash hatch pitch now
font-derived so it still reads as hatching at large text.
1549 unit tests green, both schemes build.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The board speaks when files change under the user, per DESIGN/10 § Live
board announcements. BoardDiff is the pure snapshot summarizer (identity
sets for cards/lanes added/edited/moved/deleted — ids, not tallies, so
pro-m1's semantic commit engine can build on it; edited = rendered
content only, moved beats edited, implied events don't steal the
subject). BoardAnnouncer is the decision seam: focusOutcome computes the
vanishing-focus sentence and the walk-up-then-sideways recovery (next
lane by order, else previous, board container only when none remain,
never the trash); speech(for:) is the one-sentence precedence ladder —
raised condition > bracket completion > cleared condition > vanished
focus > digest — foreign-only for the last two rungs, so app-mediated
echoes stay silent.
BoardStore.land assembles ReloadFacts and posts exactly one sentence per
reload through the injectable announce outlet (AccessibilityAnnouncer,
medium priority, never interrupting). Selection recovery layers on top
of ItemReferenceSet re-resolution — survivors veto, the emptied
selection lands on the vanished item's lane and re-arms ⌘N's active-lane
memory. performWholesale(announcing:) arms a completion phrase consumed
by the closing reload — nil on every base bracket today; pro-m1 fills
git phrasings. Locks raised outside the reload path (vanished root,
unwritable location) announce through the same ladder, and the banner
strip is a labeled "Board status" container whose row labels are the
announced sentences (AccessibilityPhrases.bannerLabel — one string for
eye and ear).
Announcements classify at reload granularity (WatchOrigin) as a
deliberate interim: DESIGN/02's EchoLedger (per-file classification, the
announcer's specified input, git-free) was scheduled with the
auto-committer that the edition split moved to pro-m1 — filed on the
Redesign board for a ruling. 1533 unit tests green, both schemes build.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The board window's accessibility tree per DESIGN/10: lanes are containers
labeled "<title>, lane, N cards" (filter-aware count = renderedCards, the
badge's own collection); cards are one flattened element each — label =
title or the untitled placeholder, value = attachment count + "cut,
pending paste", selection via trait; face icon, stripe, and paperclip are
decorative and hidden. Masonry never leaks into traversal: slots carry
order-keyed accessibilitySortPriority, so a wide lane reads by card order,
not column-major. Lane titles carry the heading trait for the rotor.
VO-Space is the ⌘-click analogue routed through the existing
BoardStore.click funnel (SelectionGrammar stays the single answer for
toggle and container-boundary rules) — cards and lane headers both.
Context-menu rows double as custom accessibility actions, each calling
the same private method as its menu row so the surfaces cannot drift;
trash cards expose Delete and Reveal in Finder and never Open. The trash
column is pinned last via sort priority 0, its label/value re-routed
through the new AccessibilityPhrases seam; toggling trash visibility
posts a one-line announcement from the store seam (both command faces).
The invisible lane-resize drag strip leaves the tree — the stepper and
menu items are the accessible width path.
AccessibilityPhrases is the pure vocabulary seam (labels, values, plural
folding shared with TrashModel.phrase), pinned by its own test suite.
Both schemes build; 1466 unit tests green.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The v5 guide prose, verified two ways. A scripted walkthrough gave a
fresh agent nothing but the guide and a demo board: it created a card
(fresh lowercase UUID, correct bottom rank), moved one to a lane top
with modified and modified-by re-stamped, deleted one into .trash/,
attached a file into attachments/, picked `fern` off the palette table,
and quoted a colon title — and the resulting board loads through
BoardLoader with zero warnings. The walkthrough's one finding is fixed:
the trash-arrival rule now reads formulaically ("smallest order minus
1024") instead of the spatially ambiguous "below the smallest order".
Content drift-guards join the suite: every palette name the app resolves
must appear in the guide (a Palette rename now fails a test instead of
teaching agents dead colors), the rewrite's conventions are present by
name (.trash/, attachments/, modified-by, CLAUDE.user.md, the
stage-only-your-own-paths rule), and the pathfinder's retired vocabulary
(media/, tombstones) cannot resurface — the only deleted: mention is the
warning never to write it.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The app-owned agent guide at every board root (DESIGN/08 ▸ The agent
guide): version-gated by a first-line marker (v5, superseding the
pathfinder's v4 guides on real boards), rewritten when missing or older,
byte-for-byte untouched when current or newer. A markerless CLAUDE.md is
displaced to CLAUDE.user.md when that name is free — never clobbered —
and the guide write is skipped with a log when it isn't. Symlinks,
folders, and read-only volumes are skipped in silence; the write rides
performWrite's bracket as an app-mediated Writer operation (new
WriteOperation.agentGuide), so the echo lands appMediated and the
Pro-era committer can attribute it honestly later.
Hooked at store acquire (beside the loose-file relocation, after the
watcher exists) and on every successful reload — the guide self-heals
from foreign deletion or rollback, pre-wiring 06's acknowledged undo
bounce. The refresh memo arms before each attempt and clears on a
successful write, so a failing write can't hot-loop and a foreign
deletion stays healable. First-line-only marker parsing (no Regex);
guide content is one swappable literal, finalized under the next card.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The chooser completes its three tiers: bundled by template order, then
keyed user templates, then keyless boards by display name — and a
malformed user template still lists, by folder name with the loader's
own sentence on the row, never failing its neighbours. The store is
re-scanned on every presentation and on app activation, the Reveal
round trip made honest without watching a folder 09 deliberately
leaves unwatched; Reveal lives in the chooser's header and mints the
store on first press. Save as Template repeats Duplicate's sequence —
progress row with Cancel, flush, detached cancellable copy — through
the engine: mint the store, read the next user order before the copy
can count itself, Finder-ladder the name, copy excluding .git and
.trash/, then stamp the whole template: mapping on the landed copy
through updateIndex, with no bracket because the copy lives outside
every watched board. Folder attributes deliberately don't carry — the
one lock the command stays live under is the read-only-DMG one, and
carrying its mode bits would mint a read-only template in the user's
own store; the command gates instead on the real hazard, unsaved card
content. A signpost names the template only when the ladder renamed
it. One name ladder now serves Duplicate and the store.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The pathfinder's catalog transcribed verbatim as real board folders —
lane titles, icons, palette colors, board icons, and blurbs read
straight from its BoardTemplate.all, every color already in the
palette, no invented fields, no starter cards it never had. Basic had
drifted from the source in the engine card's version: it carried
icon/iconColor keys the pathfinder's key-less Basic never wrote
(visually identical via the fallback, byte-level wrong) and a
rewritten blurb — both restored. Orders run 100 to 1000 in inventory
order, Basic first; fresh lane UUIDs, gapped ranks, one consistent
timestamp; the discovery test now pins all ten slugs, lane counts,
Basic's primacy, and distinct ascending orders through the real
loader.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
A template is a board folder the ordinary loader reads — no second
schema, no Swift catalog. BoardTemplate became exactly that: a loaded
BoardModel with chooser-facing derivations, the lane-title stub gone.
TemplateEngine instantiates by the copy-remint-restamp walk: .git and
.trash excluded at top level only — both names mean something at a
board root and nowhere else, and .gitignore must survive — every
materialized folder reminted, created/modified stamped fresh (born
today, not forked), modified-by cleared, the template: key carried
inert, the blurb and style inherited, and loose card files normalized
at this import boundary per the paste precedent so a new board never
opens with a notice about a mess its own birth made. Legacy deleted:
keys copy through verbatim to the one migrator — stripping would
resurrect, skipping would destroy. Atomicity is construct-then-clean:
a sibling temp can be sandbox-refused and a cross-volume rename is
just a second copy, so the call removes what it created on every
non-board exit and never touches an occupied destination. The
cancellable per-item walk extracted into BoardTreeCopy serves
Duplicate and instantiation with two parameters — top-level exclusions
and folder-attribute carriage, the only axes they differ on.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Phase 3 finishes the pivot at the surface. One card face serves two
containers: CardFaceView extracted with a role — board or trash — so
stripe, tint, chip, selection stroke, cut dim, marquee registration,
and drag are shared by construction, the trash side differing only in
its absences: no Open, no rename, no Style, no file-hover highlight,
and a Delete that goes through the confirmation host. The column
rewrote around the lanes' own single-column masonry so drag reflow
reads as positional slides; chrome stays the hatched header, symbol,
and count — 11 gives Empty Trash to the File menu alone. Two real
grammar bugs die here: plain Backspace on a trash selection purged
without the confirmation the menu raises, and the context menu's
Delete resolved against the standing selection, so right-clicking a
trash card under a board selection silently did nothing — it now
stages the clicked set explicitly. Open, Rename, Style, and Empty
Trash validation became testable store seams; the column is one named
accessibility container of ordinary card elements. The tombstone era
is swept: deleteItem, restoreItem, stripTombstonedChildren — dead
since lane copies stopped nesting trash — the restore verb, the
unreachable put-back banner row, and every quasi-lane doc comment.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Phase 2 swaps every consumer: Liveness and its ancestor walk are gone,
replaced by ItemContainer — a UUID set plus the container side it
lives on, presence the whole test, one selection boundary instead of
the old liveness law. Deletion stages by place: board cards move to
the trash at a store-minted head rank, trash-side delete is permanent
behind its confirmation, Delete Immediately skips the trash from
anywhere, lane delete captures the subtree and removes the folder.
Restore has no method at all — moveCards resolves members in either
container, so drag-out and cut-paste are the ordinary moves 13 calls
them, registering ordinary Move steps. The delete inverse moves the
card back to its captured lane and rank; redo replays the captured
trash rank, a value the gesture actually wrote; lane undo recreates
the subtree byte-faithfully in session. Purges register nothing —
where 13's trash section contradicts its own Rules on that, Rules
wins, filed for ruling. Staleness collapsed to present-or-absent: a
container is a path, so a foreign restore fails the delete step's
expectation structurally. Legacy tombstones migrate on the loose-file
tail hook, cards oldest-first so minting above top reproduces the
retired newest-first column, lanes returning live, one folded loss
row naming both directions. Put Back, restoreByDrag,
receiveRestoredCards, TrashEntry, and the kind machinery are deleted;
the trash column renders the container correctly with its full face
rework left to phase 3.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Phase 1 of the trash pivot: the file format learns .trash/. The loader
parses the reserved root container — cards only, one shared parseCard
for both containers so fail-fast, attachments, and verbatim documents
are literally the same code; absent means empty; symlinks and
lane-shaped nestings fall out as strays by construction. BoardModel
grows snapshot.trash as a plain rank-ordered card list — the container
has no identity to carry. Legacy deleted: keys keep flowing through
the retiring flag path so every tombstone consumer stays green, and
are additionally reported through LoadResult.legacyTombstones in the
loose-file idiom for phase 2's migration scheduling — nothing vanishes
from view before its folder has actually moved, which is also 01's
lock-deferral posture. Writer primitives land value-passing: move to
trash with caller-minted rank and the deliberate modified stamp,
tombstone migrations that surgically remove the key, physical lane
removal, per-card and whole-container purge that leaves strays
verbatim, and byte-faithful whole-subtree capture/recreate for lane
undo. Board-wide identity now spans the trash, so an import colliding
with a trashed UUID remints instead of colliding. The watcher already
delivered .trash events — isGitInternal tests a component, not a dot —
now stated and pinned rather than relied on.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The command surface was already almost entirely platform machinery —
this card proves it and pins it. Headless probes established that
NSWindow.validateMenuItem answers enablement AND rewrites the row title
from the delegate-supplied manager, so 'Undo Move 3 Cards' flows step
phrase to Edit menu with no code of ours; under the lock the rows dim
and keep their names, the correct reading of the-stack-survives. The
toolbar twins validate through validateUserInterfaceItem, which never
touches labels — 03's static-label exception proven rather than
asserted — and their specs' enablement abstention is pinned so nobody
later adds a second, disagreeing answer. The one link a headless run
cannot close is the nil-target key-window resolution itself: standard
responder-chain behavior with none of our code in it, left as the
manual check. Base-edition 'disabled without undo' scaffolding is
reworded away — every base board has undo now. New suites cover the
trash's two doors (delete-then-undo byte-identical to Put Back's
effect), position-preserving restore of a middle card, and the
capstone: five gestures forward, five presses back to the origin
board, five forward again, the menu phrase asserted after every press.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Every crossing validates its expectations before writing: each step
carries per-item HistoryExpectations — folder, effective ancestor-walked
liveness, and exactly the fields the gesture set — and a mismatch pops
the step, posts the signpost ('Undo skipped — Fix login changed outside
Lanework'), and falls through to the next. Validation reads disk, not
the in-memory snapshot: the snapshot is by construction one reload
behind every app write, so a rapid second undo would false-skip against
the pre-state — disk is what current can honestly mean at press time.
Stale and failed part ways: a stale step is one the board moved past,
so dropping it loses nothing; a failed one is refused by a usually
momentary condition, so it stays put and the crossing stops with only
performWrite's own error row — which forced the provider off
NSUndoManager onto two plain arrays, since a popped group cannot be put
back. The read-only lock disables Undo/Redo through the adapter while
the stack survives to resume on clear. Delete and restore validate
presence alone — a machine timestamp is not a decision — and a
malformed field matches nothing, since it is a shape the app never
writes.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The store is the Writer boundary, so it computes and registers
inverses: a weak history sink bound at session composition, one
HistoryStep per gesture at exactly the brackets that were already one
performWrite each — multi-card moves, style batches, width pairs, and
multi-row restores each undo as one plurally-titled step, and the Edit
session registers once at the flip from the bytes disk held before its
first landed write, debounce ticks registering nothing. Crossings run
through performWrite, so an undo brackets the watcher, echoes through
the reload, and reaches every window; every closure captures values,
never snapshots. The inventory follows 13 exactly: moves return to
origin lane and order, renames restore or remove the title key,
restyles and resizes restore field values or absence, tombstones and
restores swap with captured timestamps, and an undone create is a real
removal — no trace — with redo re-materializing the same UUID from
bytes captured at gesture time. Purge, attachments, repair,
bookkeeping, checkbox flips, raw Apply, and the whole arrival family
register nothing, each exclusion documented where it lives. Step names
speak 06's verb vocabulary through the new HistoryPhrase.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The provider seam 12 promised: HistoryProviding speaks 13's vocabulary
— register a HistoryStep (bare 06 phrase plus undo/redo closures
returning applied or skipped), canUndo/canRedo, action names, clear —
and no UndoManager type appears anywhere in it, proven by a fake that
satisfies the seam with counters. The base provider wraps a private
UndoManager with groupsByEvent off so coalescing stays the Writer call
site's decision; undo re-registers the reversed step from inside the
undo, which makes a stale-skipped step vanish for free and the
crossing loop fall through to the next. BoardUndoManager adapts the
protocol to the responder chain — a stackless UndoManager subclass
answering from the provider — so Pro's git provider inherits menu
enablement, dynamic titles, and the nil-target toolbar pair by binding
the protocol. One stack per board session, born in beginSession,
cleared in the close flush; every window over the board answers it
through windowWillReturnUndoManager. Headless probes shaped the
routing: a real NSTextView's own manager wins natively, but a field
editor's does not — BoardUndoRouting answers the per-window text
manager while any NSText is first responder, so a search-field typo
never crosses a board step.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
12's settled ruling supersedes the m4 placeholder: on an ordinary board
the git section is simply absent — the popover is complete in itself —
and on a board carrying an inert .git it shows only the calm one-line
note, 'This board has a git history. Lanework Pro works with it.' The
detection is a pure one-line seam checking the board root at popover
open, deliberately non-live: .git is filtered from the watch by design,
so there is no reload to hang a live fact off, and a quiet signpost
self-corrects on next open.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Two app targets from one source tree — no build flags, no #if in
shared code: an edition difference is a file one target compiles and
the other does not. Base keeps everything it had (dev.rzen.indie.Kanban,
minimal entitlements, AppIcon); KanbanPro compiles the same sources
plus the reserved KanbanPro/ root (Git/, Remote/, Auth/ land with
pro-m1 — libgit2 deliberately not added yet), adds network-client and
its keychain group, and hand-writes its Info.plist with the UTI block
verbatim — base exports the type, Pro imports it, one format either
app opens. The unit-test sources compile twice, once per host, with
Pro's module aliased so 56 test files keep @testable import Kanban
unchanged; scheme Kanban stays the muscle-memory command and
LaneworkPro joins it. InertGitTests pins the base posture with bytes
and mtimes — a full editing session over boards carrying realistic
.git trees at root and nested in a card leaves all twelve entries
untouched, and moves and copies carry them verbatim.
scripts/verify-editions.sh proves the rest: 26 checks over signatures,
symbols, entitlements, identity, and the shared UTI, discounting
Xcode's test-host exceptions by name rather than silently.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
NSToolbar through the existing HostedWindowController rather than
SwiftUI's toolbar — for reasons that are contract, not taste: 03's
transient-search clause is a decision over the toolbar's current
contents, which NSToolbar publishes and SwiftUI's API cannot answer;
Undo/Redo are the system's nil-target responder-chain actions so the
toolbar items validate exactly as the menu rows do (disabled on base
boards, alive in m8 unchanged); and the search item hosts the real
NSSearchField with explicit first-responder control. Customization is
all system furniture — Customize sheet, drag rearrange, display-mode
popup, overflow, autosaved per window kind. Board default: the search
field alone, trailing; catalog adds New Card, New Lane, Undo, Redo,
Show Trash, every action extracted from its menu command so no second
predicate exists. Card default: the Edit Body / Raw Source toggles and
Add Attachment, mirroring their commands' own predicates live via
observation tracking. Removing the search item keeps the promise —
⌘F surfaces the same field as a transient strip under the title bar,
persisting until the query clears, and an overflowed item that cannot
take the keyboard falls through to the strip too.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The sidebar completes: the shared style editor gains a second anchor —
StyleEditorLayout carries the geometry (the popover keeps its settled
268/14/7/8 untouched as the default; the sidebar packs columns to its
width with no inner scroller) while every well, the batch display, the
arrow grammar, and the one applyStyle bracket stay the shared
component's. The card anchor is fixed, not tracking: the target is
this card, and the fate walk retires the window when the card goes.
Details renders every unknown frontmatter key read-only in file order —
Card.document already carried them — showing the author's own bytes
where the raw span is a value and the engine's rendering for block
scalars and empties; reserved enhanced-schema keys are ordinary
unknowns, and no keys means no section. Actions: Delete rides the same
tombstone bytes as Backspace and drop-on-trash through a one-line
seam, says nothing about selection, and lets the fate walk dismiss;
Reveal in Finder resolves through the attachment scope so the two
paths cannot disagree. History reserves its m7 slot without drawing a
header no base board can honor.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The card's complete file inventory: compact QuickLook-thumbnail rows
over Card.attachments — no reference tracking, subfolders tolerated
and unsurfaced — with a quiet header add affordance and the drop hint
empty state. The whole window is the file-drop surface, Edit mode
included (the editor's drag types were already filtered; now tested),
sharing the board's folder-refusal semantics literally: FinderDrop
moved verbatim into its own file so both windows run the same
partition and loss row. Dragged text still lands at the caret and is
inert elsewhere — the window delegate accepts file payloads only.
Rows open on double-click or Return, drag out their file URL, and
Remove is a bracketed write through FileManager.trashItem — the system
Trash, never a hard delete, returning the in-Trash URL so the promise
is testable; the attachment listing is the guard, so traversal and
subfolder names refuse in one line. Keyboard-native per 05: the
section is one Tab stop, arrows walk rows by name, Space toggles the
shared QuickLook panel, Backspace removes. File > Add Attachment
(shift-cmd-A) comes alive through the same import path.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The escape hatch: View > Raw Source (opt-cmd-E) unmounts the whole
content area for the literal on-disk index.md in a plain monospaced
editor with Cancel/Apply. Raw source is window-level state, not a third
body mode — entry rides setMode(.preview), which flushes the Edit
session by construction, then reads the file fresh; exit reveals
Preview, and an empty body after Apply does not reopen Edit (openIfNeeded
already ran). Apply validates the proposed bytes through the loader's
own card checks — parseDocument's strict UTF-8/BOM rejection, schema,
order — deliberately skipping the uneditable-shape refusal, since a
flow-mapping card is exactly what the hatch repairs; invalid bytes
alert in place with the loader's own error and no bracket opens. The
write is byte-for-byte with no modified stamp and no modified-by clear,
per 01's explicit carve-out — the verbatim contract outranks stamping —
and identical bytes write nothing. Escape cancels, cmd-Return applies,
toggle-off applies too, and cmd-E disables while raw is active via a
testable predicate. Tombstoned targets refuse as vanished: a foreign
delete is never reverted by a stale buffer.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The editing surface: the same hosted TextKit-1 text view gains an
editable branch with a per-keystroke line-scanner highlighter — chosen
over a parser re-parse because a mid-typing buffer is usually invalid
Markdown and 05 wants the delimiters themselves dimmed; apply only sets
attributes, so presentation-never-transforms is structural. Saves ride
a ~700ms injectable debounce through BoardWriter.writeBody —
toggleTaskMarker's idiom widened to the body span, frontmatter bytes
untouched, refusing to write when disk already holds that body, which
enforces all three gates (untouched, reverted, echo) at the layer that
owns the bytes with one isDirty predicate above it. Mode grammar lands
whole: ⌘E toggles with a checkmark, Return in Preview enters, Escape
returns, and every flip flushes first; window close flushes through
the existing retry/save-copy/discard modal, and the dismissal flush
deliberately reaches a tombstoned card. Dirty-buffer-wins: disk always
follows the snapshot, the buffer only when clean, both surfaces render
the buffer. Undo is the editor's own session-scoped NSUndoManager;
endEditSession names the pro-m1 one-commit-per-session boundary.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The card body's resting state: swift-markdown (pinned 0.8.0, smart
typography off — Preview renders the bytes on disk) parsed into a pure
BodyMarkup model with UTF-8 source offsets, rendered on one hosted
TextKit 1 NSTextView — chosen because find-in-text is NSTextFinder,
checkbox clicks reuse AppKit character hit-testing, links are .link
attributes, and NSTextTable's automatic layout is exactly the
columns-sized-to-contents rule. The GFM subset renders per 05; HTML
stays verbatim code-styled text; relative images resolve against the
card folder while remote URLs are never fetched, drawing a quiet chip
instead. Task checkboxes are live: a click flips exactly one byte
through a fresh-read, refuse-uneditable, stamp, atomic-replace write —
the app's only offset-addressed write, so a moved target refuses as
staleTarget and what the user saw decides the direction, netting one
toggle on a double-click. Empty bodies open in Edit per CardBodyMode's
opening rule, applied once; the Edit surface itself stays an honest
read-only stub until its card. FindCommand prefers the card body's
find over board search when a card window is focused.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The settle rule 3f4125e introduced — at release the shadow swaps for the
dropped card(s) drawn in place immediately — is withdrawn on review. The
board returns to the presentation the committed-overlay hold has always
had: the shadows keep standing at their landing slot and the originals
keep their lift until the echo reload lands, whatever the operation was,
and the real faces arrive with the snapshot that carries the write.
Gone with it: the DropLanding/DroppedItem seam and the three landing
accessors, the hold's landing/operation payload and the removesOriginals
and keepsIdentity halves it decided, the pickup-captured titles a
cross-board arrival's face needed, LaneSlot.dropped with DroppedFace and
DroppedCardFace, the strip's dropped-lane run, and the trash column's
settled tombstone rows — the same presentation at the delete gesture's
end, which goes for the same reason while the drop-to-delete write and
its shadow row stay exactly as they were.
Kept: the hold itself, which predates the commit, along with the 1500 ms
failed-write timeout, its injectable duration and its extracted expire,
and the guard that keeps a settled release past retargeting. The tests
pinning those stay; the ones asserting what the landing slot drew are
gone, and the suite now pins the hold's own claim — the arrangement does
not move at the release.
The pause between the release and the echo is therefore an open design
question again, filed on the Redesign board.
Claude-Session: https://claude.ai/code/session_01CqjXB7ASoWtbyoGod68k97
The m4 scene plumbing was already honest — one WindowGroup value per
CardWindowRef enforces one-window-per-card, and CardWindowFate's
ancestor walk answered dismissal — so this card fills the window: a
two-column shell whose body column takes all resize flex and whose
sidebar width derives once from font metrics (26 characters of average
body advance plus em gutters), the five 05-ordered section headers as
placeholders, and the card body as selectable plain text until Preview
mode lands. The fate walk now returns a CardPlacement (card + lane), so
one pass answers both liveness and the live board › lane subtitle; a
board rename lands for free through displayName. Card windows remember
their frames per card in the board record (case-folded id keys,
unchanged-writes-nothing), restoring instead of cascading; only
unremembered cards take the last-used size and cascade. Store
acquisition stays gated on liveStore — a card window never opens a
board — and the close-flush hook stands with nothing to flush until the
Edit-session card.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
01's Lanework-owns-the-board carve-out: a regular file beside a card's
index.md belongs in attachments/, and the app moves it there. The
loader detects read-only — a new LoadResult.looseCardFiles channel,
separate from the stray-tolerance warnings because it says the opposite
thing — skipping directories, symlinks, hidden entries, and the
reserved names compared case-insensitively (on APFS, Index.md IS the
index). The relocation rides one performWrite bracket at the tail of
every successful reload, which makes lock deferral free: the reload
that lifts a read-only lock is the reload that relocates. A
lane/card/filename memo keeps a failing relocation from hot-looping —
one one-shot, then silence until disk changes. The notice rides the
loss-row class, phrasing folded by BannerCenter (one file, one card's
files, a multi-card sweep), naming original filenames per the
importAttachment rule. Paste normalizes at the import boundary: staged
snapshots' loose files land in the pasted card's attachments silently,
every arrival path declaring its side via an explicit
normalizingLooseFiles parameter — drag paths decline and fall back to
the destination's own carve-out. checkIsCardFolder closes the hole
where a lane's notes.txt would have been relocated: card depth is
exact, UUID under UUID.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Creation now clears the search by mechanism, not gesture: one seam
(noteUserCreation) states 04's rule once, called from the placeholder
funnel, paste — cards and lanes, after the staleness guard so a stale
paste clears nothing — and Finder file-drop creation; the attach path
deliberately doesn't clear, and cross-board arrivals and New Lane stay
outside the seam (a transfer isn't creation; a lane can't be born
invisible). An open inline rename now survives the filter hiding its
card: the model already kept the editor, but the field renders in the
card's slot, so renderedCards keeps the renaming card's slot exactly as
long as the editor is open — the query stands throughout, and commit or
Escape lets the predicate apply in the same pass. Verified conformant
and newly pinned: query-emptied lanes keep their slot with a 0 badge,
pasteboard staleness (takeover before paste and mid-staging both no-op),
out-transition reachability and the strip pre-divide hold by
construction with comments citing their rulings.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
04's settled clauses were mostly shipped already — the create landing
resolved through DropSlotMath.cardSlot with one nominal shadow per
importable file — but a release on the lane header fell through to the
card zones, which clamp inward, so a scrolled lane could propose behind
the header stripe. FileDropZones now folds header, attach hit-test, and
card-slot resolution into one pure seam asked in that order, the header
answering topmost per the ruling; lane headers register their frames
for it. FinderDrop.shadowCount names the floor-at-one rule. New tests
pin the header boundary, a differential against cardSlot's own zones
(same zones, not similar), and a store-level differential proving a
file landing takes the very ranks a card move there takes.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
04's ruling makes the drag the pointer's delete gesture: the shown
trash column accepts live same-board card drags, the shadow pinned
topmost — honest, since the trash sorts by deleted newest-first — and
release tombstones through the same write path as Backspace, extracted
so the two gestures cannot drift. DropTarget grew a container case for
the quasi-lane (it has no lane id by construction); lane drags,
cross-board arrivals, option-copies (re-checked at release, the one
input that can flip without a callback), trashed-side payloads, hidden
trash, and the read-only lock all refuse — and a refusal falls through
to the strip retarget, never cancelling the drag. The settle draws the
tombstoned rows in the trash under the cards' own GUIDs, so the echo is
an invisible content swap and nothing winks out for a round trip.
Selection needs no surgery: the reload's resolve rule ejects tombstoned
members as the vanish it is, pinned by a test contrasting both gestures.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
03's resettlement reverses the pathfinder carry-over: the
selection-keyed dual presentation proved undesirable, so a card has one
presentation — selection changes styling, never geometry, and the
masonry never reflows on click. Deleted the carousel view (page dots,
glass underlay, scroll-tick monitor), the QuickLook thumbnail cache
(sole consumer), the pure paging/suppression rules, and the
sole-selected animation key — Motion now keys transactions on the
search query and the drop proposal only. The attachment chip stays as
the face's whole attachment story; viewing media is the card window's
job. No carousel state had leaked beyond the view layer.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
03's sharpened settle rule: rendering the arrangement means rendering
the card — at release the shadow swaps for the dropped card(s) drawn in
place immediately, the appear never waiting for the echo reload. The
committed hold now carries the landing (ids, payload titles, operation)
and surfaces read one DropLanding seam: within-board moves draw the
real faces at their proposed slots under the arriving card's own key,
so the echo is an invisible content swap; cross-board card arrivals
draw payload-titled faces keyed positionally, so the echo reads as an
ordinary arrival. Cross-board lane arrivals deliberately keep their
shadow until the echo — a lane's face is a whole column with no honest
payload equivalent. The 1500 ms failed-write timeout is now seamed
(injectable duration, extracted expire) and pinned by tests.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
04's settled ruling: the attachment model is flat top-level files, so a
drag containing only folders never engages — no highlight, no proposal,
the standard incompatible-payload cursor — and a mixed drag proposes for
its files only, importing them at the drop while a loss row names the
skipped folders. Hover reads the providers' registered types (anything
conforming to public.directory refuses, packages included); commit
re-partitions authoritatively from the filesystem, so a synthetic
payload that hides its type still can't land a folder. The create path
now only ever fires with at least one importable file — the
mint-fail-remove dance is gone from the folder case and stays reserved
for genuine mid-batch failures.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
02 rules the trashed side has exactly one definition: the set with
trash rows — a card carrying its own deleted: under a tombstoned lane
is in neither universe, so an anchor or selection can never survive on
an item that renders nowhere. The membership rule now lives once, in
Liveness.walk, and ItemReferenceSet.idUniverse, TrashModel.entries,
paths, and emptyTrashTargets all derive from it — the old
lane-OR-card logic that admitted subsumed cards to the trashed side is
gone, and the two universes deliberately no longer partition the
board. Put Back, Delete Immediately, and Empty Trash outcomes are
unchanged: a tombstoned lane still moves and purges whole, its nested
tombstones with it.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The copy now walks the source tree item by item, checking cancellation
between items, and the in-progress banner row carries its promised
Cancel — a cancelled duplicate removes the partial sibling and never
happened (DESIGN/03 > File menu). A sandbox permission refusal of the
silent Finder-style sibling falls back to an NSSavePanel pre-filled with
the parent folder and the copy name — the panel's grant is the sandbox's
own answer; cancelling the panel cancels quietly, and non-permission
failures keep the ordinary one-shot banner. Refusal classification is
deliberately narrow (NSFileWriteNoPermissionError itself, no underlying-
chain walk) so an unreadable source never masquerades as a destination
refusal. Directories are created writable first with mode and timestamps
restored after the subtree lands, so a read-only source folder can't
strand its own copy. BoardDuplicatorTests grows from 9 to 18 tests.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
02 ratified a warning-tone class for non-failure losses — content that
didn't arrive though nothing failed: a degraded paste, folders skipped
from a Finder drop, their future kin. Loss rows take the one-shot's
dismissable-untimed lifecycle (a loss the user didn't notice is the
harm) and rank below the true failures, above commit and attachment
notices. BannerCenter grows LossBanner, postLoss, and the
skipped-folders phrasing; the degraded-paste notice moves off its
signpost onto the new class, ending its too-quiet ranking.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The committed placeholder's lane slot now takes the arriving card's
identity — .awaitingArrival keys as card:<id>, so when the echo reload
lands the ForEach sees one persisting element whose content swaps from
stub to card face instead of a removal and an insertion with two
scale+fade transitions (DESIGN/02 > overlays: the handoff must read as
one arrival). The awaiting face renders the card's exact chrome from
shared CardFaceMetrics, so any residual branch crossfade is between
pixel-identical renderings; the editing phase keeps its constant key so
typing identity holds. Slot position math is untouched.
CreateHandoffIdentityTests pins the key derivation.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The 2026-07-28 realign card asks the four caret chords (Move Left/Right,
lane-width pair) to yield to any focused text control. The seam already
conforms — caretChordsYield covers the board popover, the search field's
first-responder tracking, and card windows publish no scope — but had
zero regression coverage. CaretChordTests pins the rule: scopeless
windows never yield, either surface suffices, and inline editors remain
acceptsBoardMutations' concern, not this seam's.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The card face's in-place expansion (03-board-ui.md § Card face): a card
with attachments, as the sole selection, grows a paged media band below
its title; every other state stays compact behind the quiet paperclip.
- CardCarousel owns the pure rules: the sole-selection predicate decides
by identity (a sole-selected lane matches no face, no snapshot walk),
and expansion is suppressed outside the animation key while a rubber
band is active — a band names a set in progress, so carousels neither
flicker nor animate under it.
- QuickLook thumbnails generate off-main into a per-window cache keyed to
survive reselection, with the Finder-icon fallback while loading and
for non-previewable types; pages ride the platform paging behavior,
dots (glass underlay, solid under Reduce Transparency) click to page,
and a local wheel monitor turns a discrete tick into one clamped page —
precise trackpad pans fall through untouched.
- The expansion animates under Motion's new carouselExpansion transaction
keyed narrowly on the sole-selected card; Reduce Motion goes instant.
Drop-slot math and the marquee read the expanded height for free — both
re-register on every size change.
928 unit tests (24 new). m5-interactions complete.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Nexus parity, audited row by row (11-command-nexus.md § Menu commands) —
every already-built item's title, chord, and placement matched exactly;
this pass fills what remained:
- The future-window rows, present with stable titles and validation-driven
disablement until their milestones fill the actions: Save as Template
(m9), Add Attachment… ⇧⌘A, Find Next/Previous ⌘G/⇧⌘G, the View-menu
card triplet Edit Body ⌘E / Raw Source ⌥⌘E / History (m6), Board ▸
Pull/Push (git milestones) — one shared disabled-row shape in
FutureCommands.swift so later milestones only flip validation.
- No Print story in v1: the print group is removed.
- Help carries the Nexus's one remap-teaching line — Customize Keyboard
Shortcuts…, opening System Settings' Keyboard ▸ Shortcuts extension
directly (the modern extension URL, verified to launch the appex).
- The launch-restore decision now runs through a pure, tested
AppModel.shouldRestoreAtLaunch gate; the Settings pane's caption rides a
proper Form section footer.
904 unit tests.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The board's live title+body filter per 04-interactions.md § Search:
- SearchFilter — a pure value folding the query once (case- and
diacritic-insensitive substring, locale-stable); title OR body matches,
attachment filenames never searched; only the literal empty string is
inactive.
- One universe: the filter threads through SelectionGrammar's order lists
as a defaulted parameter, so ranges, Select All, arrow navigation, the
marquee, drop zones, count badges, and the shown trash all read the same
filtered set by construction; lanes are deliberately never filtered out
(an emptied lane keeps its slot with a 0 badge). Hidden cards leave the
selection through the existing constrain primitive, run on every query
change and as the last line of the reload resolve; the delete successor
is filtered so ⌫ never selects a hidden neighbour.
- The field: an NSSearchField-backed toolbar item (the toolbar's sole
default item); Edit ▸ Find ⌘F focuses it through a focused-value
presentation; stock field-editor dispatch — Return swallowed, Tab is the
keep-filter path to the board, board commands stay enabled except the
caret-chord pair, now one shared caretChordsYield expression.
- Escape is staged: clear the non-empty query (focus stays), hand an empty
field back to the board, clear an active search from board focus —
before Escape's clear-selection meaning.
- Creating a card clears the search (the placeholder funnel); a rename
deliberately gets no carve-out; filter reflow rides the content spring
keyed narrowly on the query.
903 unit tests (24 new).
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
⌘X/⌘C/⌘V for cards and lanes per 04-interactions.md § Clipboard:
- ClipboardStore stages full folder snapshots eagerly at the gesture into
Application Support (at most the current copy; sweep at launch and on
each copy purges what the pasteboard no longer references; a copy made
before quitting pastes whole after restart) and writes the pasteboard a
JSON manifest — every entry embedding its index.md, lane entries their
cards' too — plus plain-text titles.
- Cut is Finder-style deferred: items dim in place off pendingCut, void on
pasteboard takeover (changeCount, no timers), source-board close, or
per-item external tombstoning; the first armed paste moves the surviving
originals whole (tombstoned interior cards land in the destination's
trash), a second paste materializes copies from staging.
- Paste anchors by the shared flatten-order rule (NewCardTarget's anchor,
extracted); a tombstoned selection never anchors; lane paste reaches the
right end and stays enabled on a zero-lane board; paste into the source
board is the within-board lane duplicate; copies keep created, take
fresh GUIDs, and strip tombstoned cards; trash-sourced copies strip
deleted: at materialization; ⌘X is disabled on the trash side.
- A degraded paste is loud, never silent: staging gone → the embedded
index.md fallback lands content-intact, attachments absent, and a
BannerCenter-phrased row names what was lost.
- The standard Edit items validate through conditionally-attached
onCommand handlers, so AppKit's enablement mirrors the availability
predicates; text fields keep their own clipboard while focused.
879 unit tests (68 new).
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
Files from Finder land on the board per 04-interactions.md § Drag & drop:
- Dropped on a card, they copy into its attachments/ (any type, multi-file),
the face highlighting while hovered; the hovered card is resolved by
hit-testing the same analytic masonry frames the card zones are built
from, so attach-beats-create adds no drop region and cannot drift from
the dispatch.
- Dropped on lane empty space, one card per file — filename minus extension
as the title (a blank stem omits the key), fresh GUID, rank at the drop
position through the ordinary insertion machinery, the file attached —
all in one bracket; a failed import removes the just-minted card, so
creating-then-abandoning never leaves an empty card behind.
- Every board drop surface now declares .fileURL beside the two board
types (the single-target-dispatch rule); tombstoned surfaces are inert;
file sessions ride a distinct session mode with their own watchdog and
no hysteresis, leaving the board-drag machinery untouched.
- Also: four empty fixture directories pinned with .keep files so git
preserves them, and a test-only visibility fix in DragSessionTests.
811 unit tests (12 new).
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
The second half: system drag sessions over phase 1's model, per
DRAG-REORDER.md and 04-interactions.md § Drag & drop.
- Card faces, lane headers, and trash rows drag as NSItemProvider sessions
(two exported UTTypes, JSON payload in flatten order, plain-text titles as
the secondary representation) — replacing m4's custom lane-reorder gesture
and trash drag-out wholesale; the app-wide DragSession carries the members,
the frozen dragged sizes, the live proposal, and the effective operation.
- Three drop delegates (lane masonry, strip, window fallback), each accepting
both types and routing internally per the single-target-dispatch rule; the
cursor is the physical mouse converted to strip space; proposals come from
DropSlotMath with hysteresis threaded through, and the lane-strip proposal
clamps in front of the shown trash.
- Locality picks the default — move within a board, copy across, the badge
tracking live; ⌥ forces copy (ignored on within-board lane drags), ⌘
forces move; trash rows restore within their board (positional), copy out
across boards by default, ⌘ forcing the true restore-move.
- N contiguous shadows with reflow keyed on the proposal; the
committed-overlay hold renders the dropped arrangement until the reload
echo lands (1.5 s dissolution deadline for refused writes); the
re-grounding trio: geometry re-derives per render, proposals re-validate
by liveness at release, an emptied drag cancels itself.
- Edge autoscroll (ticking driver over DragAutoScrollMath, re-targeting per
step), the mouse-up-gated late-event cleanup, and the polling watchdog —
the pathfinder's lifecycle traps, ported.
- Store: moveLanes and multi-card restoreByDrag join the one-bracket drop
commits.
784 unit tests.
Claude-Session: https://claude.ai/code/session_01SR4XGjmBE16ZUYWpfFHXwY
rzen