Files
multica/packages/views/editor/text-anchor.ts
Naiyuan Qing ea03912baf perf(desktop,issues): single-router tab sessions (MUL-4741 Phase 2) + trace-driven surface mount/render overhaul (MUL-4474/4750 reland) (#5403)
* Reapply "perf(issues): virtualize inbox/list/board/swimlane (MUL-4474, 方案2) (#…" (#5395)

This reverts commit c10bfa8f56.

Co-authored-by: multica-agent <github@multica.ai>

* fix(issues): seed virtualized lists so route-return doesn't flash blank (MUL-4750)

The relanded MUL-4474 virtualization flashed an empty card area (group /
column headers present, rows blank) when returning to /issues or crossing
inbox<->issues. Two stacked blank windows caused it:

1. The scroll element reaches Virtuoso via a callback ref that lands in
   state, so the first render after a remount has customScrollParent === null
   and the code rendered nothing.
2. Even once mounted, Virtuoso renders 0 rows until its post-paint
   ResizeObserver measures the viewport.

Fix both, four surfaces (list / board / swimlane / inbox):

- New shared <VirtuosoSeed> renders a bounded slice of the real rows while the
  scroll parent is still null, reusing each caller's own itemContent /
  computeItemKey so a seeded row is identical to its virtualized counterpart.
- Pass initialItemCount={Math.min(len, SEED)} so the measurement frame keeps
  those rows instead of collapsing to empty.

SEED is capped at 30 and floored by Math.min, so small workspaces
(hasMore=false) and short columns never over-mount — the path that crashed on
real Desktop before. restoreStateFrom (tab-switch Activity restore) is
intentionally out of scope for this round.

Verified: @multica/views tsc --noEmit, eslint, and full vitest (1936 tests)
pass. Real-Desktop route-return / DnD / keyboard / scroll-position regression
pass still owed on device.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>

* perf(issues): defer per-card popup mounting, share one context menu per surface

Tab-switching to a board froze the main thread for seconds: every card
eagerly mounted ~6 popup roots (context menu, pickers, hover cards) plus
per-card query subscriptions, multiplied by seed x columns x remount.

- DeferredPopup: pickers render a pixel-identical static trigger and mount
  the real popover on first pointerenter/keydown (Base UI opens on click,
  so the warm mount always wins the race)
- AssigneePicker/PriorityPicker/DateOnlyPicker defer when uncontrolled;
  AssigneePicker's members/agents/squads/frequency subscriptions now only
  start on interaction
- IssueActionsContextMenu: one controlled ContextMenu per surface anchored
  at the cursor via a virtual anchor; items delegate (issue, position) up.
  Known debt: iOS Safari long-press no longer opens it
- ActorAvatar hover cards warm-mount on pointerenter with a manual
  first-dwell timer matching Base UI's OPEN_DELAY

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(issues): stabilize board column scrollbar across mounts

Column scrollbars redrew visibly on every surface mount (route switches,
first open): the seed frame's scroll height covered only the seeded cards,
then Virtuoso spaced out the full count.

- VirtuosoSeed: optional estimatedItemHeight renders a trailing spacer so
  the seed frame's scroll height already approximates the full list
- Board columns: seed capped at 10 (one column viewport of ~110px cards,
  not the 36px-row-sized generic 30) and the same estimate feeds Virtuoso's
  defaultItemHeight so both phases agree until real measurements land
- Columns at <=30 cards skip virtualization entirely and render plainly
  (same itemContent), making their scroll height browser-measured truth in
  every scenario -- the per-column split Linear ships
  (data-virtual-cluster=false for small clusters)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* perf(editor): reduce issue detail mount cost

Parse long Markdown in smaller chunks without the duplicate initial sync, defer title and empty composer editors until intent, and keep the description editor eager to avoid layout shifts.

* chore(desktop): add navigation boundary lint rule (MUL-4741 Phase 2 prereq)

The tab Coordinator protocol requires that application code never
navigates directly (invariant 1: a Router location change without a
Coordinator token is a protocol error). Enforce it statically:

- renderer app code may not import useNavigate/Navigate from
  react-router-dom nor call router.navigate; src/platform is exempt
- the five known legacy sites (the RFC §8.1 migration checklist,
  cross-validated: the rule fires on exactly those and nothing else)
  carry inline eslint-disable directives tagged MUL-4741 — the Phase 2
  migration removes them one by one, and this rule holding with zero
  disables is the machine check that the migration is complete

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(views): open deferred pickers on click, align triggerRender types

Pointerenter warm-mounting swapped the trigger element mid-gesture: real
browsers re-hit-test so the click lands on the new trigger, but synthetic
pointer sequences (tests, assistive tech) keep dispatching on the detached
node and the first click dies. Upgrade now happens on click/Enter/Space
only — the same timing as Base UI's own trigger — with the in-flight click
stopped so the popup's just-mounted outside-press dismissal doesn't close
it in the same breath.

Also widen triggerRender to ReactElement<Record<string, unknown>> (React
19 defaults ReactElement props to unknown) and mount the
IssueContextMenuProvider in the swimlane test harness like IssueSurface
does in production.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* refactor(desktop): single-router tab sessions with Coordinator protocol (MUL-4741 Phase 2)

Replace the per-tab-router + <Activity> keep-alive model with the approved
single-router session architecture:

- TabSession: tabs are pure serializable state (url, resourceKey, virtual
  history stack, scroll memento). Persist v4 does the one-time legacy
  view-state import from v3; mountGeneration is deliberately unpersisted.
- Coordinator (platform/tab-coordinator.ts) is the only router writer: it
  reconciles THE app router to the active session URL with navigation
  tokens; a location change without a token is a protocol error handled by
  bounded recovery (invariant 1). The router history is never used — every
  reconcile is a replace; back/forward are session-stack operations.
- ActiveTabHost mounts exactly one tab, keyed on tabId:mountGeneration.
  reload() = generation bump + active-scope query invalidation (never
  router.revalidate, never a global cache invalidation). Warm switches
  restore scroll pre-paint; cold restores pre-size containers from the
  memento's saved scrollHeight and settle when data lands.
- resourceKey dedup (pathname only) replaces exact-path dedup: opening
  /slug/issues?filter=b focuses the existing issues tab (RFC §8.2,
  deliberate semantic change).
- All five §8.1 legacy navigation sites migrated (index <Navigate>, error
  page recovery, workspace-layout login bounce, overlay parking, shell
  back/forward); their MUL-4741 ratchet eslint-disables are removed, so the
  navigation boundary rule now holds with zero exemptions outside
  src/platform.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(issues): register board columns and list scroller for scroll mementos

Per-container scroll registration for the MUL-4741 tab session memento:
board columns key by group id (each column's offset restores
independently, the per-column split Linear ships), the list view keys as
"list". Chat and issue-detail already carry the marker.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(tabs): pull-based scroll restoration fed into virtualized lists (MUL-4741)

Rebuild the restore side of the memento protocol on first principles (the
model Linear ships): a saved offset is an INPUT to the mounting view, not a
post-hoc DOM mutation from outside.

- ScrollRestorationProvider (views/platform): views pull their saved offset
  while mounting. Virtualized lists feed it into Virtuoso's initialScrollTop
  so the first render already materializes the rows around it — this
  replaces the pushed spacer+scrollTop hook, whose foreign spacer deadlocked
  against Virtuoso's own height model (restore landed the viewport in
  phantom space, Virtuoso rendered nothing, and the spacer's removal
  condition could never be met → blank issue detail). Plain containers
  assign the offset at ref-attach, pre-paint. Web has no provider and
  behaves as before.
- Memento keys gain a route dimension (`${pathname}::${containerKey}`) and
  capture now also fires before in-tab navigation, so back/forward restores
  each route's own offsets and same-named containers on different routes
  no longer collide.
- commitScrollMemento uses REPLACE-per-route semantics: a container
  scrolled back to 0 clears its stale offset instead of resurrecting the
  old position on the next visit.
- List view gets the same estimate alignment as the board (36px rows into
  the seed spacer and defaultItemHeight), which keeps the shared scroller's
  height truthful from the first frame so the restored offset sticks.

Known gap: chat's bottom-anchored list captures offsets but has no restore
consumer — intentional, it re-anchors to bottom on mount.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* perf(editor): make unlabelled code fences plaintext instead of auto-detected

Follow-up to the issue-detail mount work: lowlight's highlightAuto runs
every registered grammar over the full block for code fences without a
language, which dominated mount cost on code-heavy comments. Extract a
shared syntax-highlight module whose auto fallback deterministically
renders plaintext; explicitly labelled languages highlight as before.
Also ignore .gstack/.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* perf(issues): lazy-mount column-header popups, right-size swimlane seed

Trace analysis of tab/view switching (30s session): a swimlane mount spent
its largest slice on eagerly-mounted header machinery — ~170 tooltip roots
(one per lane x status cell add-button), 13 column dropdown-menus, and up
to 30 fully-materialized lanes from the generic seed count.

- DeferredTooltip (views/common): renders only the trigger until first
  hover, then mounts a controlled Tooltip anchored to the SAME element
  (no trigger swap, so mid-gesture events never land on a detached node);
  ui TooltipContent grows an `anchor` passthrough for it.
- Board/list/swimlane header add-buttons and hide-column dropdowns now
  defer via DeferredTooltip / DeferredPopup (which gains an ariaHasPopup
  option for menu triggers).
- Swimlane lane seed drops 30 -> 6 (a lane row is ~300px+; a viewport fits
  ~3) on both the pre-scroll seed and Virtuoso's initialItemCount.
- openTab gains an `activate` option so "open and focus" paths (pinned-tab
  redirect, explicit open-in-new-tab) are ONE store write instead of
  openTab + setActiveTab back-to-back — one full subscriber pass per user
  action instead of two.
- Dev-only breadcrumb logs IssueSurfaceContent's remount key: the trace
  showed the surface mounting twice inside one task, and the my-issues
  relation toggle is one confirmed key-flip source; the log ties the next
  trace's mounts to exact key transitions.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* perf(issues): stop double full-tree render passes on surface interactions

Trace forensics on view switching showed every interaction paying TWO full
surface render passes (React's own Cascading Update marker sits between
them): entering swimlane flips controller-level loading state (loadProjects
enables the projects query), and any such flag flip re-rendered the entire
unmemoized view tree (~600-1000ms dev per pass).

- Memoize BoardView / ListView / SwimLaneView: controller/data outputs are
  already useMemo/useCallback-stable, so a controller flag flip now
  re-renders the header, not the whole board. The one unstable prop —
  BoardView's inline assigneeGroups.flatMap — moves into a useMemo.
- Selection reset on mount swapped the initial empty Set for a NEW empty
  Set, buying a guaranteed extra full pass per surface mount; functional
  bail keeps the reference when nothing was selected.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* perf(ui): drive sidebar resize by direct DOM writes, drop motion/react

Trace analysis showed sidebar motion.div mounts costing ~1s across a
session. Width previews during drag now write straight to the two layout
shells (CSS disables their transitions while data-sidebar-resizing is set);
React only sees the single committed width on pointer-up, and framer-motion
leaves the sidebar entirely.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(issues): wire swimlane's outer scroller into tab scroll restoration

Review blocker on #5403: board/list/issue-detail register their scroll
containers with the tab session memento protocol, but the swimlane outer
scroller did not — under the single-router architecture an inactive tab
unmounts, so a deep-scrolled swimlane returned at top after a tab switch
or reload.

Same wiring as the other surfaces: data-tab-scroll-root="swimlane" for
capture, useRestoredScrollRef in the scroller's attach callback for the
pre-paint assignment, and the saved offset into the lane Virtuoso's
initialScrollTop. Regression test asserts both the capture marker and the
restored offset.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-15 16:30:51 +08:00

159 lines
6.1 KiB
TypeScript

/**
* Text-anchor caret mapping between the readonly stand-in and the editor.
*
* The lazy-editor swap used to land the caret with raw pixel coordinates
* (`posAtCoords`), which assumes the readonly render and the Tiptap render
* are pixel-identical. They never are on long documents: per-block height
* differences (tables, code headers, mention cards) accumulate top-to-bottom,
* and NodeViews/images keep reflowing after `onReady` — so clicks landed
* lines away from the target. A text anchor sidesteps layout entirely:
* record WHICH character of WHICH top-level block was clicked, then resolve
* that logical position in the ProseMirror document.
*
* Offsets count NON-WHITESPACE characters only. The two renderers disagree
* about whitespace inside container blocks — react-markdown keeps the HTML
* source's inter-element newlines as text nodes ("URL: x\nTitle: y"), while
* ProseMirror's textContent concatenates with no separator ("URL: xTitle: y")
* — so raw character offsets drift one per list item / nested block. The
* non-whitespace character sequence, however, is identical on both sides.
* The whitespace ambiguity this creates at word boundaries ("end of this
* word" vs "start of the next") is resolved by `bias`, captured from what
* the user actually clicked next to.
*
* Known drift: content the two renderers textify differently (a mention's
* display name is text in the readonly render but an atom — zero text — in
* the doc) shifts offsets within that one block; resolution clamps to the
* block end, so worst case the caret lands earlier in the same block, never
* in another block.
*/
import type { Node as PMNode } from "@tiptap/pm/model";
export interface TextAnchor {
/** Index of the top-level block (readonly root child ↔ doc child). */
block: number;
/** Count of non-whitespace characters before the caret within the block. */
offset: number;
/**
* Which neighbor the caret attaches to when whitespace separates the
* offset-th non-whitespace character from the next one: 1 = just before
* the next non-whitespace character (user clicked a line/word start),
* -1 = just after the previous one (user clicked a line/word end).
*/
bias: 1 | -1;
}
const NON_WS = /\S/;
function countNonWs(s: string): number {
let n = 0;
for (let i = 0; i < s.length; i++) if (NON_WS.test(s.charAt(i))) n++;
return n;
}
/**
* Resolve a click point inside the readonly render to a text anchor.
* `root` is the markdown container whose element children correspond 1:1 to
* the document's top-level blocks. Returns null when the point yields no
* caret position (unsupported API, click on empty margin below content) —
* callers fall back to coordinate focus.
*/
export function anchorFromPoint(
x: number,
y: number,
root: HTMLElement,
): TextAnchor | null {
const doc = root.ownerDocument;
let node: Node | null = null;
let nodeOffset = 0;
// Standard API (Chromium/Firefox), then the WebKit legacy one.
if (typeof doc.caretPositionFromPoint === "function") {
const p = doc.caretPositionFromPoint(x, y);
if (p) {
node = p.offsetNode;
nodeOffset = p.offset;
}
} else if (typeof doc.caretRangeFromPoint === "function") {
const r = doc.caretRangeFromPoint(x, y);
if (r) {
node = r.startContainer;
nodeOffset = r.startOffset;
}
}
if (!node || !root.contains(node) || node === root) return null;
// Climb to the top-level block (the direct child of `root`).
let blockEl: Node = node;
while (blockEl.parentNode && blockEl.parentNode !== root) {
blockEl = blockEl.parentNode;
}
if (blockEl.parentNode !== root || blockEl.nodeType !== Node.ELEMENT_NODE) {
return null;
}
const block = Array.prototype.indexOf.call(root.children, blockEl);
if (block < 0) return null;
// Text before the caret within the block. Range.toString() handles both
// text-node carets and element carets (offset = child index), and yields
// a prefix of the block's textContent — which lets us peek at the char
// right after the caret for the bias.
const range = doc.createRange();
try {
range.setStart(blockEl, 0);
range.setEnd(node, nodeOffset);
} catch {
return { block, offset: 0, bias: 1 };
}
const before = range.toString();
const nextChar = (blockEl.textContent ?? "").charAt(before.length);
return {
block,
offset: countNonWs(before),
bias: NON_WS.test(nextChar) ? 1 : -1,
};
}
/**
* Resolve a text anchor to a ProseMirror position. Out-of-range values clamp
* (block → last block, offset → block end) so a stale or drifted anchor still
* lands inside the intended block instead of throwing.
*/
export function posFromAnchor(doc: PMNode, anchor: TextAnchor): number {
if (doc.childCount === 0) return 0;
if (anchor.block >= doc.childCount) return doc.content.size;
const blockIndex = Math.max(0, anchor.block);
let blockStart = 0;
for (let i = 0; i < blockIndex; i++) blockStart += doc.child(i).nodeSize;
const block = doc.child(blockIndex);
const contentStart = blockStart + 1;
const wanted = Math.max(0, anchor.offset);
let seen = 0;
// Caret position right after the `wanted`-th non-whitespace character
// (block content start while none have been passed yet).
let afterPrev = contentStart;
let resolved: number | null = null;
block.descendants((child, pos) => {
if (resolved !== null) return false;
if (!child.isText) return true;
const text = child.text ?? "";
for (let i = 0; i < text.length; i++) {
if (!NON_WS.test(text.charAt(i))) continue;
if (seen === wanted) {
// This is the first non-whitespace character AFTER the caret.
resolved = anchor.bias === 1 ? contentStart + pos + i : afterPrev;
return false;
}
seen++;
afterPrev = contentStart + pos + i + 1;
}
return true;
});
if (resolved !== null) return resolved;
// Caret past the last non-whitespace character: exact end lands after it;
// an overshoot (renderer textified something the doc doesn't, e.g. a
// mention label) clamps to the block's content end.
return seen === wanted ? afterPrev : blockStart + block.nodeSize - 1;
}