The legacy bar rendered MAX_DISPLAY_COUNT FriendBarItemViews and padded
empty slots with null, so an empty online-friends list produced three
identical 'Trova Amici' buttons. The current bar already renders a single
explicit search chip, but harden it: filter null/undefined out of the
online-friends array before slicing/mapping so the search chip is the only
possible source of that affordance — exactly one, always.
The messenger rendered the participant figure straight from the frozen
thread participant, so offline friends (whose look used to be empty)
showed the anonymous/standard avatar. Read the live look from the friend
list via getFriend() - the same source the friends list renders - with
resolveAvatarFigure() as the final fallback, so the real avatar shows
even when offline (pairs with the server fix that now sends offline
looks). Applied to both the avatar-bar tab and the in-thread avatars.
Also fix the avatar-tab head framing: it positioned the head-only image
with full-body geometry (90x130, top:-31px), clipping the head. Render
the head at native size (background-size:auto, no scaling -> not grainy)
and centre it in the 36x36 tab.
The online-friends bar is portaled into the right toolbar nav, which sits
inside `tb-nav-clip` (fixed, `max-w-[calc(50vw-242px)]`, `overflow-x: clip`).
Each online friend adds a fixed `w-[132px]` chip, so the bar grew with every
friend up to MAX_DISPLAY_COUNT (3). Once it exceeded the clip width the right
edge was silently cut off - the scroll arrow and part of the search button
disappeared. The portal slot is `shrink-0`, so the chips never compressed;
they just overflowed and got clipped. Net effect: "more friends online =
broken bar".
Measure the room actually available between the bar's left edge and the
viewport's right edge (re-measured on resize / ResizeObserver) and derive an
effective visible count clamped 1..3, always reserving space for both arrows
and the search chip so nothing clips at any width or friend count. The bar's
left edge is stable (it follows fixed-width toolbar icons), so changing the
chip count never moves it - no measurement feedback loop.
Scroll offset now derives a clamped safeOffset used by every read, so a
stale indexOffset after the list shrinks / the fit grows renders correctly
and self-corrects on the next arrow click (no write-back effect).
React 19 dropped the no-arg useRef overload — the type-only useRef<T>()
form (no initial value) is gone, every call must pass an initial value.
The codebase had 15 occurrences of useRef<HTMLDivElement>() (DOM ref
pattern) all flagged by tsgo as 'Expected 1 arguments, but got 0'.
Mechanical sweep to useRef<HTMLDivElement>(null) — no behavior change,
React still hands out a ref object with .current set to null at mount.
Net tsgo error count: 57 -> 42.
ToolbarView and FriendsBarView declared their motion variant objects
without a type annotation, so tsgo widened transition.type to 'string'
where framer-motion's Variants narrows it to a literal union (spring /
tween / inertia / etc). Every <motion.div variants={...} /> site flagged
the mismatch.
Annotating the constants as Variants makes the literal inference work
('spring' stays 'spring'); also drops the redundant 'as const' on
staggerDirection now that the parent type pins it.
Net tsgo error count: 133 -> 100.
Run eslint --fix across src/ to clear ~1900 mechanical lint errors
surfaced by the @typescript-eslint v8 + react-hooks v7 + react-compiler
upgrade in the React 19 modernization PR.
Issues fixed automatically:
- brace-style (Allman): try/catch one-liners reformatted to multi-line
- indent: tab-vs-space and depth corrections
- semi: missing trailing semicolons
- no-trailing-spaces
No semantic changes. Remaining 701 errors are real-code issues
(set-state-in-effect, rules-of-hooks, no-unsafe-* type checks) that
need manual per-file review.
https://claude.ai/code/session_01GrR87LAqnAEyKG2ZbmQt5Q
simoleo89