Easter Eggs Consolidation & Overhaul - #184
Merged
Merged
Conversation
Why this change
===============
The easter-egg games were reachable only via Ctrl+Shift+1/2/3 keyboard
chords, which we are retiring: hidden chords are undiscoverable by design,
and the product direction is that eggs should be found by users who pay
attention to details (clicking real UI elements, trying phrases in chat).
This commit is the structural groundwork for that direction; the actual
new discovery triggers land on top of it.
What was wrong structurally
===========================
- Three byte-identical chord hooks (useDinoShortcut, useGame2048Shortcut,
useSpaceInvadersShortcut) differing only in which digit they matched.
- Two near-identical modal wrappers (DinoGameModal, SpaceInvadersModal)
plus a third variant for the 2048 iframe.
- The dino "waiting game" logic (unlock flag + Space trigger + auto-close)
was copy-pasted across ChatPage and OnBoardingPage (~40 lines each),
including duplicated localStorage/dinoUnlockChanged/storage sync code.
- The chat phrase easter eggs ("do a barrel roll", matrix variants) were
matched with inline string comparisons inside ChatPage.
- All game code was statically imported through DashboardPage, so ~1,600
lines of canvas/iframe game shipped in the main bundle even for users
who never open an egg.
What replaces it
================
features/easter-eggs/ now owns the shared machinery:
- registry.ts: every modal egg as { id, label, kind, component }, with all
components lazy-loaded so each game becomes its own chunk fetched on
first open. Adding a future egg = one registry entry + one component.
- components/EggModalShell.tsx: one Framer Motion modal for every egg.
Canvas games render bare (they own their chrome and Escape handling via
onExit); iframe games get a header bar + close button + window-level
Escape (the iframe swallows keys, so the shell must listen). One
Suspense wraps both branches because an unsuspended lazy child would
tear down the whole tree.
- components/Game2048Frame.tsx: thin iframe wrapper so the self-contained
vanilla-JS 2048 page fits the shared { onExit } game shape.
- hooks/useDinoWaitingGame.ts: useDinoUnlocked() (live localStorage flag
synced via dinoUnlockChanged + storage events) and useSpaceOpensDino()
(Space opens while armed+unlocked, typing-guarded, closes when the flag
flips off — via React's adjust-state-during-render pattern, not an
effect, matching ChatPage's existing style and the lint rule).
- hooks/useRepeatClicks.ts: generalized N-consecutive-clicks detector;
first consumer is the dashboard header icon (triple-click = dino), the
same gesture language the Settings cogwheel already uses.
- lib/eggPhrases.ts: phrase matcher extracted verbatim from ChatPage so
behavior is unchanged there and other surfaces (buddy chat later) can
reuse it.
Consumer changes
================
- DashboardPage: all three chords removed; triple-clicking the page-header
icon now opens the dino runner (ungated, deliberately - finding it IS
the unlock). 2048/invaders lose their only triggers for now; new ones
follow in upcoming commits.
- NotFoundPage: no longer force-opens Space Invaders on load; uses the
shared shell instead of its own modal instance. A visible teaser lands
separately.
- ChatPage / OnBoardingPage: waiting-game logic replaced by the shared
hooks; OnBoardingPage additionally closes the game when a regeneration
starts (same as before) and keeps its DinoGame mount.
- useDinoEasterEgg (Settings cogwheel): rewritten onto useRepeatClicks +
shared unlock helpers. Its dinoUnlockChanged event is now dispatched on
a microtask because the old synchronous dispatch ran during a React
render phase (setState-in-render hazard flagged by lint). Consumers
listen for the event asynchronously anyway.
Verification
============
npm run build: PASS (pre-existing chunk-size warning only)
npx eslint src tests: PASS (0 problems)
npm run unit: 205 files / 1814 tests PASS
npm run a11y: 52 files / 60 tests PASS
format:check: my files clean; 5 unrelated files were already drifted on
feature/minor-upgrades (verified against committed blobs), untouched here.
Why this change
===============
Space Invaders used to force itself open the moment the 404 page loaded,
covering the actual error content before the user could read it. That is
the opposite of what an easter egg should be: an egg is a reward for
paying attention, not something thrown at you. It also contradicted its
own copy - the page invited the user to "save the galaxy" but never let
them choose to.
What changed
============
- The modal now starts closed. Arriving on 404 shows exactly the error
content: headline, copy, Return to Dashboard.
- Below the copy sits one small teaser row, deliberately styled to blend
in (subtle text color, small size): "While you wait for your manager's
approval... [rocket emoji]". Clicking anywhere in the row opens Space
Invaders via the shared EggModalShell.
- The whole row is the button so the hit target stays generous; the
emoji is aria-hidden and the button carries an explicit accessible
name ("Open Space Invaders") instead of reading out the ellipsis.
- Hover/focus states follow the app's standard surface-hover + focus ring
pattern.
Tests
=====
- NotFoundPage.test.tsx: asserts the game is NOT mounted on arrival,
that the teaser reads correctly, and that clicking it mounts the shell
with eggId "space-invaders" (mocked shell keeps the test hermetic -
no lazy game chunks load).
- NotFoundPage.a11y.test.tsx: mock updated from the deleted
SpaceInvadersModal to EggModalShell for the same reason.
Why this change =============== The easter-egg overhaul moved discovery from keyboard chords to real UI details. But a trigger nobody notices is just a dead button, so the icons that hide a game need the smallest possible tell - enough for an attentive user to think "huh?" and click, not enough to stop being an easter egg. The Settings cogwheel (dino) is the first consumer; the dashboard header icon gets the same treatment with its own game in the next commit. What changed ============ PageHeader gains an `eggHint` prop (default false, purely opt-in). When set on an icon that already has `onIconClick`: - The icon grows ~15% on hover (the button already had transition-transform + active:scale-95, so this rides the existing motion system rather than adding a new one). - On mount it plays exactly ONE 0.5s brand-colored glow pulse via a drop-shadow keyframe. One-shot by construction (`1 both`, never re-triggered), so it reads as a flicker of "this thing is special", not as a notification or badge. Implementation notes ==================== The effects are two small CSS rules keyed off classes + data-[egg-hint="true"] in styles/index.css, next to the barrel-roll keyframes - not Tailwind arbitrary variants - because both the hover descendant selector and the prefers-reduced-motion override read better as plain CSS and the file is already the home for one-off app-wide animations. Accessibility: the hint is decorative only. It is deliberately NOT reflected in aria labels or roles - screen readers get the same plain "Settings icon" button as before (asserted by a dedicated test). Users with prefers-reduced-motion get neither the hover scale nor the pulse; both overrides sit inside one media query next to the base rules. subtitle became optional in PageHeaderProps (was required): ChatPage already renders PageHeader without one and passed runtime undefined; typing it optional makes the contract honest. No behavior change - empty subtitle renders nothing, exactly as before. Tests ===== New tests/unit/components/layout/PageHeader.test.tsx covers: plain icon stays non-clickable, onIconClick turns it into a named button, the data-egg-hint attribute appears exactly when eggHint is set, and the accessible name is unchanged by the hint. Existing PageHeader a11y scan, SettingsPage tests and the full unit/a11y suites pass unchanged.
Why this change
===============
When the chords died, the dashboard's triple-click-the-header-icon
trigger was parked on the dino runner as a placeholder. The plan assigns
2048 to that spot: the dino already has two discovery homes (the Settings
cogwheel unlock plus Space during AI thinking / onboarding generation),
while 2048 had none left after losing Ctrl+Shift+2. The chart-column icon
over a grid of widgets also rhymes nicely with a game about merging tiles
in a grid.
What changed
============
Same gesture, same target grammar ("header icons hide games"), different
game: three quick clicks on the dashboard page-header icon now open 2048
in the shared EggModalShell instead of the dino modal. Still ungated -
finding it is the fun, no localStorage flag involved. The empty-board /
edit-mode concern does not apply: the header icon exists on the dashboard
in every state, unlike the edit-mode-only Add-widget button considered
earlier.
Dino therefore has no chord-free dashboard entry anymore - intentional;
its homes are the cogwheel and the waiting screens.
Verification
============
npm run build PASS, npx eslint src/tests PASS (0 problems),
npm run unit 206 files / 1819 tests PASS, npm run a11y 52 files / 60
tests PASS. Dashboard/PmDashboard/PageHeader suites re-run individually.
The AI chat already offers the dino runner (press Space while the AI
thinks, when unlocked) and two phrase easter eggs ("do a barrel roll",
"the matrix"). They stayed stuck on the AI chat: typing a phrase to the
*buddy* did nothing, and while the buddy churned there was never another
game to click — so the two chats drifted apart, the exact inconsistency
the BuddyProvider exists to prevent (dock vs. /buddy page sharing one
conversation).
What changed
------------
1. App-level effect layer. The barrel roll and Matrix rain used to be
local useState inside ChatPage. They now play through a shared
`eggEffectBus` (playEggEffect / useActiveEggEffect / clearEggEffect)
and render from one EggEffectsLayer mounted once in App.tsx — so any
surface, not just the AI chat, can fire them. ChatPage was migrated
onto the bus (~35 lines of local state/effect deleted; phrase matching
routes through matchEggPhrase + playEggEffect instead of calling
setIsBarrelRolling/setIsMatrixActive). ChatPage keeps its MatrixRain
import deleted; the layer owns it now.
2. Buddy waiting-game. useBuddyConversation now calls useSpaceOpensDino
with (isThinking || isStreaming, dinoUnlocked) — Space opens the
runner while the buddy works, and it auto-closes when the answer
arrives (mirrors ChatPage's render-adjust close). The props flow up
through its return value and down through BuddyThread → Buddydock/
BuddyConversation → BuddyWidget → BuddyDock, so both the dock and
the /buddy page show the running game inline under the typing
indicator.
3. Buddy phrase interception. useBuddyConversation.handleSubmit checks
matchEggPhrase(draft) BEFORE sending: a match clears the composer,
fires playEggEffect, and swallows the message — no request, no reply.
Same silent contract as the AI chat, so "do a barrel roll" / "the
matrix" behave identically in both chats.
4. Tests. New eggEffectBus.test.tsx (4 cases: idle→barrel, barrel
timeout+cleanup, matrix mount/unmount, clear on unmount) and
buddyEasterEggs.test.tsx (phrase swallowed silently vs. normal send
reaching the wire, dino Space trigger while thinking). useDinoUnlock
coverage moved into the bus test's storage-event case.
Design notes
------------
- EggEffectsLayer is mounted outside `showSidebar` (i.e. always, even on
/login) because a fired effect must always have its renderer; the
barrel-roll body class is harmless on login, and matrix rain on a
login screen is the kind of secret that rewards attention.
- The waiting-game uses React's documented "adjust state when a value
changes" pattern (guarded setState during render) rather than
set-state-in-effect, matching ChatPage's existing approach — keeps the
close-on-answer synchronous with the render that knows the turn ended.
- dinoGameActive/closeDinoGame are optional on BuddyDock (defaults false /
no-op) so the existing dock test mocks, which don't pass them, keep
compiling and green.
Tests: unit 208/1825, a11y 52/60, eslint 0 problems, tsc clean.
Why this change =============== The chord retirement left two kinds of dead weight behind. The moments feature still carried its own five development-only chords (Ctrl+Shift+4..8) — a tuning aid, marked "remove before merging" in its own header, whose reason to exist (replaying progress-gated moments) has nothing to do with shipping. And the docs/comments still described a world that no longer exists. What is removed =============== - `useMomentDevShortcuts.ts` (whole file) and its call in `MomentsProvider`. The hook was DEV-only (`import.meta.env.DEV` compiled it out of production), so this changes nothing for users; it just deletes the tuning scaffolding before the branch lands. - The `EggEffects = Record<EggEffectId, ReactNode>` type from eggPhrases.ts — a leftover from the pre-bus design where each page rendered its own effects. The bus replaced it; nothing imported it. eggPhrases.ts now re-exports `EggEffectId` from the bus instead of declaring a second copy of the union. - A stale comment block in SpaceInvaders.tsx pointing new readers at the deleted SpaceInvadersModal and the retired Ctrl+Shift+3 chord. Verification ============ npm run build (tsc -b + vite), eslint src tests, prettier --check on touched dirs, unit suites for easter-eggs/buddy/moments/pages (373 tests) all pass.
Why this change
===============
The phrase easter eggs had two entries (barrel roll, matrix) and the
bus had exactly one renderer per id — adding a third effect is the
proof that the EggEffectsLayer design scales without new plumbing.
"Party" was picked over more game-adjacent ideas because it needs no
new surface, no persistence, and no trigger UI: someone types "party"
into the AI or buddy composer and the app celebrates. It is also the
first alias set that includes an emoji ("🎉"), which is deliberately
the most guessable trigger in the whole system — no language barrier,
and typing one character into a chat box is a very low-risk experiment
for whoever suspects there is something hidden.
Trigger
-------
matchEggPhrase gains ["party", "party time", "let's party", "🎉"] →
"party". Matching stays exact after trim/lowercase, so "partying" and
"a party for two" remain normal messages. The union type on
EggEffectId makes the compiler force the new branch into
EggEffectsLayer — no silent fall-through possible.
ConfettiBurst
-------------
Same architecture contract as MatrixRain: all mutable particle state
lives in local variables owned by one effect, a single delta-time-scaled
rAF loop renders it, DPR-aware canvas, cleanup cancels everything.
- Two cannons (bottom-left/bottom-right) fire up-and-inward, 75
particles each, staggered across the first 150ms so the burst reads
as two pops rather than one wall.
- Particles are rects and circles with per-particle spin, gravity
(900 px/s²), air drag, and a sinusoidal horizontal flutter whose
phase advances faster while falling — that flutter is what makes it
read as paper instead of ballistics.
- Colors come from the computed CSS custom properties of the theme
(brand plus the extended accent hues), read once at spawn: confetti
automatically matches light/dark mode and any future palette change.
A hardcoded five-color fallback covers SSR/test environments where
getComputedStyle resolves nothing.
- When every particle has died, the canvas fades out (~250ms) and the
effect clears itself off the bus; the layer keys ConfettiBurst by
seq, so re-triggering mid-burst replays cleanly instead of being
swallowed by equal-state bailing.
Reduced motion
--------------
Particle animation IS the effect, so there is no honest way to keep
it: under prefers-reduced-motion the component renders a static
"🎉 Party!" status chip that fades via opacity only and clears the
bus after 1.5s. Nobody who types "party" ends up staring at nothing.
Tests
-----
- eggPhrases.test.ts: all four aliases match, plus trim/case
invariance; "partying" and substring cases stay null.
- ConfettiBurst.test.tsx (via EggEffectsLayer): canvas mounts
pointer-events-none + aria-hidden; the full lifecycle
(spawn → simulate → fade → bus clear) completes under fake timers
using a stubbed 2D context and faked rAF/performance (jsdom has no
canvas, so without the stub the loop would exit before doing
anything); the reduced-motion chip renders instead of the canvas
and also clears its bus effect when its timer runs out.
Verification
------------
build ✅ · eslint src tests ✅ 0 problems · prettier (touched files)
✅ · unit 209 files / 1836 tests ✅ · a11y 52 files / 60 tests ✅
Why this change =============== The easter-egg overhaul replaced keyboard chords with real UI details a user can stumble on. The strongest remaining surface for that grammar is the app logo itself: it already reacts to hover (badge lift, rocket nudge, flames light up), so it reads as interactive — but clicking it did nothing. That gap between "looks alive" and "is clickable" is exactly where an egg belongs, and repeated clicks on a decorative element is the same discovery gesture as the cogwheel and dashboard-icon triggers. Research note: the logo only *looks* clickable — it is decorative markup at all three usage sites (desktop sidebar header, mobile top bar, Keycloak login template). Per the agreed approach it stays decorative: plain onClick plus cursor-pointer, no button, no tab stop. Rationale: the egg is pure whimsy with zero function, and adding a focusable element to the login page's keyboard path (of all places) would make navigation worse for every screen-keyboard user to serve a secret. The animation ------------- A four-phase gravity sequence on the outer badge div, driven by framer- motion's onAnimationComplete rather than hand-tuned delays, so each beat starts exactly when the previous one lands: 1. falling (~0.45s, easeIn): y +140px with a slight forward tumble — gravity accelerates, it does not slide. 2. impact (~90ms): sideways squash (scaleX 1.14 / scaleY 0.8), the classic cartoon beat. 3. bouncing (~0.62s): two diminishing bounces via keyframed arrays (34% then 16% rebound heights) with re-squashes that shrink per landing. 4. hopping (spring, stiffness 260 / damping 11): back to the shelf with a small overshoot while rotation settles to 0. The fixed 140px translateY keeps the effect layout-independent — parents are overflow-visible, so the badge visually escapes its header row without portals or measured rects. Deliberately NOT routed through the egg effect bus: barrel roll and matrix are whole-window effects owned by EggEffectsLayer; this one is scoped to a single element that three different layouts place differently. Element-scoped animation, element-owned state. Interplay details ----------------- - The rocket's permanent idle float lives on the inner <svg>, the drop on the outer div — separate DOM nodes, transforms compose. - whileHover="lift" is suppressed while any phase is playing so hover never fights the drop for the same transform; it re-arms at idle. - Clicks during a run are absorbed by the phase gate (only idle arms a new drop); useRepeatClicks resets after firing either way. - prefers-reduced-motion skips the sequence entirely — pure motion has no honest static equivalent — though the click counter still consumes. - One targeted behavior kept from review: DROP_PHASES is passed as the animate target object directly (not by variant name), because the phases are not part of the badge's rest/lift variants family. Tests ----- SidebarLogo.test.tsx asserts the trigger contract under a motion-stripping proxy: the logo remains decorative markup (no button role) with cursor-pointer; four clicks leave it untransformed; five rapid clicks do not throw even under reduced motion. Pixel choreography is not asserted (jsdom has no animation loop — same boundary as the game tests). Verification ------------ build ✅ · eslint src tests ✅ 0 problems · prettier (touched files) ✅ · unit 210 files / 1839 tests ✅ · a11y 52 files / 60 tests ✅
…ose for 2048 - Enable eggHint on DashboardPage PageHeader so the ChartColumn icon triggers the opening glow pulse and scales on hover - Enable Escape key close handling for 2048 game modal via postMessage from 2048.html and keydown listeners in Game2048Frame - Add unit test coverage in DashboardPage.test.tsx and Game2048Frame.test.tsx - Apply Prettier formatting across codebase
…ture/easter-eggs-overhaul
This was referenced Aug 28, 2026
Merged
…ture/easter-eggs-overhaul
daniilperkin
marked this pull request as draft
August 29, 2026 12:44
Collaborator
Author
|
should be merged after confluence, so undraft after it also |
Collaborator
|
Draft, so no approval — but I went through the branch's own commits ( The consolidation is genuinely good: one The diff against
|
daniilperkin
changed the base branch from
dev
to
feature/269-confluence-connector
August 31, 2026 13:37
Collaborator
Author
|
Claude is right, i will point this PR into feature/269-confluence-connector |
Brings 141 commits of dev — industry detection, starter-work, the TanStack
Query migration, smooth page navigation, the chat conversation rail — together
with the easter-egg overhaul (one shared egg module behind a bus, the waiting
games, the 404 rocket teaser).
Eight files conflicted. Resolved as follows.
App.tsx — kept both sides. dev renamed the shell's auth gate from `showSidebar`
to `signedIn`, so the rocket pet reads that now. `EggEffectsLayer` deliberately
sits outside that gate: a fired effect must always have its renderer, whoever
is looking. The comment above the pet points at `AppearanceSection`, since the
branch's `MomentsSection` no longer exists.
DashboardPage.tsx — took the egg shell (`EggModalShell` + `useRepeatClicks`)
over dev's three per-game modals and their Ctrl+Shift chords: this branch
deleted `src/features/{game2048,dino,space-invaders}` and moved the games into
`features/easter-eggs`, so six imports had nowhere to point. dev's
`WidgetPickerModal` replaces `AddWidgetModal` (the JSX already said so).
ChatPage.tsx — kept dev's conversation rail, `surfaceFromPathname` and
`useNewConversationShortcut`; dropped dev's `MatrixRain` import, because matrix
rain is now fired through the egg bus and drawn once, app-wide, by
`EggEffectsLayer`. dev's inline barrel-roll / matrix / dino wiring in this file
is superseded by `playEggEffect` plus `useSpaceOpensDino` / `useDinoUnlocked`.
OnBoardingPage.tsx — dev wrapped the generator in `useCallback` and
invalidated the onboarding status on both the success and the recovery path;
this branch had replaced the inline dino state with `useSpaceOpensDino`. Both
kept: `closeGame()` closes the waiting-game, `invalidateMyOnboardingStatus()`
still runs in both places (that was the stale-status bug dev fixed), and
`closeGame` joins the dependency array — it is a stable `useCallback`, and
leaving it out left an `exhaustive-deps` warning in a repo that runs at zero
warnings. The branch's `eslint-disable` for that array is gone with it.
Buddy chain (BuddyThread / BuddyConversation / BuddyDock / BuddyWidget) — both
features thread props through the same four components, so each now forwards
both concerns: the dino waiting-game pair and the visit divider's
clear-previous action. The dock forwards the clear action without naming a
keyboard chord, and that is deliberate: over a page like /chat that chord
starts a chat, not a visit.
Verified before committing: `tsc -b` silent, `eslint .` 0 errors / 0 warnings,
2328 unit tests, 63 a11y tests, production build green. The five files still
failing `prettier --check` are dev-side drift this branch never touched and are
left alone on purpose.
daniilperkin
marked this pull request as ready for review
September 15, 2026 19:52
…ne game at a time Review findings on the consolidation branch, fixed in one commit. Each was reproduced against the branch head (46ff21a) before being changed, and the reasoning is repeated in the code comments where a future reader will look. Buddy: Space can open the waiting-game at last - The chat blurs its composer on submit precisely so the Space trigger can fire (see ChatPage's focus effect), but the buddy composer never gave the caret up: after Enter the textarea still held focus, and the trigger's "don't eat the space bar while somebody is typing" guard refused every press. The waiting-game this branch brought to the buddy was therefore only reachable after clicking away from the box. - BuddyComposer now blurs on submit and takes the caret back when the turn ends — the same two-way dance the chat does — and only when *this* box gave the caret up in the first place. The dock auto-focuses its composer and is mounted on every page, so an unconditional refocus would steal the caret out of whatever the hire was actually typing in. - "Busy" is isThinking || isStreaming, matching the chat's notion of a live turn. BuddyConversation takes a new isStreaming prop (the page and the dock already had the flag) instead of folding streaming into isThinking, which would change what the thread draws. Barrel roll stops spinning for prefers-reduced-motion - .barrel-roll-active had no reduced-motion guard, so a full-screen 360° rotation of the whole app played for users who explicitly asked for less motion — while everything else this branch touched respects it (the confetti falls back to a static chip, the logo's drop is skipped outright, the header hint's glow is disabled). - The layer skips the class and clears the effect immediately: the spin has no honest still frame, the same call the logo's drop makes ("the counter still consumes, nothing plays"). Typing the phrase still works; there is simply nothing to see. - MatrixRain is deliberately left alone: it is a dismissible overlay the player asked for, it is pre-existing behaviour and this branch did not change it. Left as a residual rather than silently rewritten. One waiting-game at a time, app-wide - Three surfaces arm the trigger against their own turn and all listen on the window, so a single Space press with two of them busy opened two dino games at once (chat streaming with the dock open mid-answer). The first host to claim the shared slot plays; the rest stay shut. - The slot is released on close, on the unlock toggle and on unmount, so a page that goes away mid-game cannot leave every other surface unarmed. - A module-level symbol rather than context: the hosts are a page, a floating dock and a wizard step, with no component above all three to own it. 404 teaser: the accessible name now contains the visible words - aria-label="Open Space Invaders" replaced the visible sentence, so the button's name was not in its text (WCAG 2.5.3). The hint moved into an sr-only suffix inside the name instead, and the test asserts the contract rather than the mismatch. 2048 frame: one mechanism per Escape, same-origin only - The frame attached the same Escape handler to both the frame's window and its document, so one press called onExit twice — and the frame's own page already reports Escape through EGG_EXIT, which made the injected listener redundant as well as doubled. What is left is the focus on load, so arrow keys work without a click first, plus that message path. - The message listener now insists on event.origin === window.location.origin. The game page is ours, so a message from anywhere else can only be some other frame closing a modal that is not its business. Hygiene and docs (no behaviour change) - registry: iframeSrc was dead config (the wrapper hardcodes the one path it has) and the `dino` entry had no call site left — the waiting-game is played inline in chat, onboarding and buddy, never in the shell. EggId is now the two eggs that actually open as modals. - EggModalShell: dropped the unused id on the iframe header (the dialog is named by aria-label) and made the close button's testid egg-relative instead of naming 2048 inside a shared shell. - PageHeader's icon button is type="button"; it defaulted to submit. - ConfettiBurst states why three components now share its name. - Stale references are gone from SpaceInvaders (the Ctrl+Shift+3 chord and the deleted SpaceInvadersModal), MatrixRain (ChatPage's old inline arrow) and the README, whose feature list and structure tree still named dino/, game2048/ and space-invaders/ after this branch deleted them. - The Keycloak login copy of the logo now says what it is: the same mark, deliberately without the gravity egg, still synced by hand. Tests: Game2048Frame covers both message directions (our own origin closes, a foreign one is ignored); EggModalShell covers the two eggs that exist.
npm run format:check fails on five files that predate this branch: ProjectIndustryPanel, useStarterWorkReview, projectService, ProjectIndustryWidget.test and starterWorkService.test. Formatting only — npx prettier --write, nothing else — and deliberately kept out of the easter-egg review commit so the behaviour change and the churn do not hide each other in blame. CI never ran format:check (ci.yaml runs lint, build, unit, a11y and the docker build), which is how the drift shipped in the first place. These five were the whole of it on this branch, so the check is green from here.
The shared slot behind the Space-to-play dino game was released on host unmount and when the unlock flag flipped off, but never when the wait itself ended. The game's DOM belongs to the host's waiting state, so it disappears on its own the moment the answer arrives — leaving `gameActive` true kept the slot claimed by a game nobody could see, and the next surface to arm found the trigger dead. The onboarding generation step is where that bites hardest: it arms on `loadingState === "generating"`, has no close of its own, and stays mounted after the path arrives, so one finished generation ate Space for the rest of the visit — chat, buddy and onboarding alike. Chat and buddy had each grown a copy-pasted "close when busy flips" block to work around it; both are gone now that the hook owns the close, which is what the buddy hook's own comment already claimed it did. The close is a render-phase adjustment next to the existing unlock flip, so there is no extra effect, no cascading render, and no host that has to watch `armed` itself. A second armed host can open its game once the first one's wait has ended — a test that fails without the release.
The bus has carried a `seq` since it was written, and ConfettiBurst was
keyed by it — but neither of the other two ids ever saw it. `MatrixRain`
was rendered without a key, and the barrel-roll effect listed only
`effect?.id` in its dependencies, so firing the same phrase again while
its effect was running handed the second trigger the first one's
deadline: the rain ended on the first instance's 6s timer, a second
barrel roll was cut short by the first 2s timer. The layer's own
documentation promises the opposite ("re-triggering replays the effect"),
and the two re-fireable ids were precisely the ones that did not.
MatrixRain is keyed by `seq` now and the barrel-roll effect depends on
it, so a re-fire restarts the clock. The body class itself stays
idempotent: cleanup drops it, the effect puts it straight back.
ConfettiBurst's fade timer is ref-held and cleared on unmount, for the
same class of reason. It fires into the bus rather than into the
component, so the cleanup that looked complete — rAF, both spawn timers,
the resize listener — still let a burst that was remounted (or replaced
by another effect) clear the *new* effect a few hundred ms in.
Three tests, one per re-fire path, each fake-timer driven and each
failing without its fix.
Reviewed the 16 owner-authored commits since the last push (buddy team mode, the GitHub repository facet, the dino waiting game evolution and the closed-task card) with three parallel read-only reviews, verified every finding against the sources, and fixed what was real. BabuPlk's three commits (arriving via the #248 re-merge) were excluded as other people's work. Buddy team mode (major + minors): - useBuddyConversation: a refused hire action's "Try again" button was dead. The card renders it only for a resolved refusal (!action.ok && "action" in action), but confirmAction's guard let only idle/error through - every click hit the guard and returned. The component-level test stubbed onConfirm, so it passed while the hook refused the retry. Now a resolved refusal that still carries its action passes the guard; a permanent refusal (no action payload) and everything confirming/confirmed/declined stays spent. The per-card flight lock already cleared in finally, so a retry runs. - useBuddyConversation: teamTargetRef survived a signed-in-user change. The userId effect set team mode directly, bypassing setTeamMode - the only place that cleared the target - so the next user's adopt effect compared their selection against the previous user's binding and overwrote the restored preference with a bogus exit. The ref now lives above the effect and is cleared on every subject change. - useBuddyConversation: the dock's "Clear the earlier conversation" control is handed to BuddyThread unconditionally, and startFreshVisit's guard checked greeting/decisions but not a live reply - clicking it mid-stream silently dropped the tokens, citations and proposed actions still arriving. The guard now also refuses while isThinking/isStreaming (one-line fix at the hook; the durable stream-abort redesign stays a separate piece of work). - BuddyWidget: stale comment claimed the mode switcher still carries the restore audit; it moved into the session - the comment now points at the session. Dino waiting game (major + minors): - DinoGame: the window keydown/keyup handlers had no typing-target guard, and the game now outlives its turn - exactly when the composer regains focus. Space/w/s/arrows were prevented and eaten mid-sentence, and a stray Escape closed the game while typing. Both handlers now bail when the event lands in a text field, reusing the same isTypingTarget guard the trigger applies (now exported, widened to EventTarget for event.target). - The completion badges claimed success from "not busy anymore": - ChatPage promised "Reply ready" while tokens were still streaming (isThinking flips false at the first token) and on a failed or stopped turn; now it waits for the stream to actually end. A provider-level clean-end flag would fix the error/stop labels too - left as a follow-up, failures surface as toasts so there is no state to read from the page today. - OnBoardingPage promised "Path ready" for a failed generation (status error also leaves running); now only status done claims it. - SourceDetailsPanel promised "Sync complete" when a sync failed into attention or was disabled; now only state connected claims it. - ThinkingIndicator: the suppression comment claimed the ReasoningPanel displays the thinking state - it does not render thinkingState, so a post-reasoning tool label has no surface. The comment now documents the accepted gap instead of a fiction. Knowledge base (minors): - githubMetadata: getArtifactRepository re-parsed and re-warned the raw metadata JSON on every card render, drawer render and facet memo evaluation - the WeakMap cache only covered matchesRepository. The parse now runs once per artifact object for every reader (compute core + cache wrapper), which also caches the org-login lookup. - orgMetadata: parseOrgMetadata accepted a blank-string login (unlike the repo parser's blank rejection) and returned it untrimmed, silently missing the case-folded owner-half match against a trimmed repositoryFullName. It now rejects blank logins and trims like the repo parser does. - useKnowledgeBase: the facet count allocated a Set inside the per-candidate filter - O(repos x candidates) throwaway allocations per evaluation; now one per offered repository. - githubMetadata.test: the matchesRepository tests leaked real console.warns for the malformed-metadata case, masking unexpected warnings; silenced like the parse tests, asserting the warn fired. Verified: npm run build, npm run lint, npm run format:check green; unit suite 2911/2911 across 353 files. Not pushed, per instruction.
…PATH_STEP board, buddy quotes, AI skill suggestions)
roll, dialog focus contract, and a truthful confetti palette Fact-checked pass over the six strict-review findings, keeping the ones that survived verification and refuting the rest: - ConfettiBurst: the fifth palette candidate read a variable that does not exist (--color-app-yellow-text); the real token is --color-app-highlight-yellow, so the burst now paints all five app hues instead of silently running on four. The claim that the whole reader was dead is refuted: the --color-app-* tokens are defined via @theme inline in styles/index.css, so getComputedStyle resolves them. The inert animate-fade-in / shadow-card classes (defined nowhere) are dropped. - Buddy waiting game: the one surface where the game died at the first token. The hook now arms useSpaceOpensDino with keepActiveUntilExit like the chat and the drawer; the thread keeps the typing bubble mounted while the game is open even after thinking ends; isStreaming flows from both callers (dock and full page) into the thread, and replyReady is only true once thinking AND streaming are done — so the dots row stops lying once the reply has arrived and the game's completion badge takes over. - EggEffectsLayer: re-triggering the barrel roll mid-run removed and re-added the class inside one commit, which the engine answers with a shrug — the animation kept its already-finished state. A forced style recalc between cleanup and re-add makes the second roll actually roll from the top. - EggModalShell: focus contract without fighting the games — focus lands on the dialog on open (tabIndex={-1}, so it joins no Tab order) and returns to the opener on close; Tab wraps inside the dialog only. Tab is deliberately the only trapped key: the canvas games own arrows/w/s/space, and keydowns inside the 2048 iframe never reach this window, so neither surface can collide with the trap. - OnBoardingPage: the queueMicrotask around the last-generation snapshot stays — fact-checking showed it looks redundant but exists to satisfy react-hooks/set-state-in-effect, which errors on the direct setState the reviewer suggested. A comment now records that so the next reader does not "simplify" it away. Refuted by verification, left unchanged: the useRepeatClicks window finding — the inter-click-gap semantics are the documented design ("the guard only has to rule out accumulation across a session"), and the proposed total-window reading would tighten a deliberate gesture. Verified: npm run build, npm run lint, npm run format:check green; unit suite 3017/3017 across 360 files.
daniilperkin
added a commit
that referenced
this pull request
Sep 24, 2026
…ogins Three small hardening changes to the GitHub metadata readers, plus a format-check exclude for local worktrees. They were first written on feature/easter-eggs-overhaul (#184) against the KB code already on dev, where they had no business living: two of the touched files overlap this branch and made #184 and #264 conflict whichever landed second. They now live here, next to the code they harden, and #184 drops them. - githubMetadata: getArtifactRepository now reads through the same per-artifact WeakMap cache as the org-login lookup. The card list, the drawer and the facet pass all call it on every render, and each call re-ran JSON.parse (and re-logged the warning for a malformed payload) for metadata that never changes for a given artifact object. The uncached core is one private computeRepositoryInfo both public readers share, so the two can no longer disagree. - orgMetadata: a blank login is rejected and a padded one is trimmed. The owner-half repository match compares against a trimmed `owner/repo` string, so " org " silently never matched its own repositories. This is the same normalisation the repo parser already applies to repositoryFullName. A new test pins both cases. - githubMetadata test: the malformed-blob case now silences and asserts the expected parser warning, so real warnings stay visible in the test output. - .prettierignore: skip .worktrees. eslint and vitest on this branch already exclude it, but format:check still walked every per-task worktree copy and flagged files formatted against other branches. Not ported: #184's facet-count Set hoisting in useKnowledgeBase. This branch moved facet counting server-side, so that loop no longer exists. The vite/eslint worktree excludes are already here in this branch's own form. Gates: build, lint (0 warnings), format:check clean; unit 362 files / 3131 tests.
This branch carried seven files that have nothing to do with easter eggs: hardening in the knowledge-base metadata readers, a Set hoist in useKnowledgeBase, and .worktrees excludes for vite, eslint and prettier. They were review fixes to code already on dev, parked here while this branch was the only open one. Now #264 is open and rewrites the same KB code: merge-tree showed useKnowledgeBase.ts (about 240 lines) and vite.config.ts conflicting whichever PR landed second. All seven now match origin/dev again: - githubMetadata/orgMetadata hardening + its test: moved to #264 (9cd5081) with a new trim test the original lacked. - useKnowledgeBase Set hoist: dropped. #264 moved facet counting server-side, so the loop it optimised no longer exists. - vite/eslint .worktrees excludes: #264 carries its own versions. This also drops the duplicate `vitest/config` import the vite change left. - .prettierignore: moved to #264 with the KB commit. No worktree nests inside this branch's own checkout, so its gates are unaffected.
…ailed chunks The modal shell promised a focus contract it could not keep, and the lazy game chunks could take the whole app down. - Focus trap: keydowns inside the 2048 iframe never reach the parent, so Tab from its last link walked out of the aria-modal dialog onto the page behind it (WCAG 2.4.3). A document focusin guard now pulls focus back to the dialog's first or last control, depending on the direction it left. - Focus restore: on a cached reopen the iframe focused itself (a fresh about:blank already reads readyState "complete") before the shell's effect ran, so the shell recorded the iframe as the opener and restored focus to a removed node. The shell now captures the opener in useLayoutEffect, drops the guard before restoring, and restores only an opener that is still connected. Game2048Frame focuses on the iframe load event only. - EggErrorBoundary: a game chunk that fails (stale tab after a deploy, offline) used to unmount the root, since the app has no boundary. The shell now shows an EmptyState "Couldn't load <game>" with a working close. The 2048 header renders outside Suspense so it can always be closed. - Reduced motion: MatrixRain ignored the preference the effects layer documents. It now shows a static chip like confetti, via the new shared ReducedMotionEffectChip. - Space trigger: isInteractiveTarget exempts buttons, selects, links and ARIA widgets, so Space presses the focused control instead of opening the dino. Modifier keys, auto-repeat and already-handled presses are also ignored. isTypingTarget keeps its text-field-only meaning for the game's own key filter. Unlock reads are guarded against a throwing localStorage. - useRepeatClicks: the window now spans the whole gesture, not each gap, which is what the docs said (3 clicks in 3 s, not 3 s per gap). - Confetti: its status now mounts empty and fills a tick later so screen readers announce it. The pale highlight yellow is dropped; the hex fallback is kept and documented as jsdom-only. - Hardening: ui/Button close control. 2048 messages must come from our iframe; the page posts to location.origin, not "*". Phrases accept curly apostrophes (iOS "let’s party"). Tests: focus contract (open, restore, removed opener, Tab wrap, focus pulled back), failing chunks, two armed hosts on one Space, focused controls keep Space, load-only frame focus, foreign messages, matrix reduced motion, curly apostrophes, gesture window.
The unlock popover's live region was mounted together with its text
inside AnimatePresence, so screen readers often skipped the "unlocked"
announcement entirely. The emoji's aria-label also repeated the
"shh…" copy, so it was read twice.
- The status region is now always mounted; only the bubble animates.
The emoji is aria-hidden.
- useDinoEasterEgg returns an explicit kind ("unlocked" | "locked").
The popover used to pick lock vs unlock by matching substrings of
the toast text, so any wording change would flip it silently.
- The popover uses the floating shadow-lg and app tokens instead of
shadow-black/10 with dark: overrides.
…evived icon - SidebarNavIcons: StarterWorkIcon came back in merge fix-up 44f3cc5 after dev deleted it (d922f8e, Starter Work became Hire Setup). Nothing imports it. Removing it and its Target import leaves the file byte-identical to dev, so this PR no longer touches it. - SidebarLogo springs: the hover springs use hoverSpringToken instead of inline 400/16 and 420/14. The five-click drop really needs a bouncier landing, so it gets a named logoHopSpringToken in tokens.ts instead of magic numbers. The comment claiming the egg lives on the login page too now says two in-app surfaces; the login copy deliberately has none. - SidebarLogo tests: the old motion mock turned every tag into a div and dropped `animate`, so "idle below five clicks" could never fail. The logo root now exposes data-drop-phase. The tests check idle after four clicks, falling after five, each stage in turn, no restart mid-drop, and staying idle under reduced motion (a stubbed useReducedMotion, since framer caches matchMedia). - NotFoundPage: the Space Invaders teaser said "While you wait for your manager's approval…", which makes no sense on a 404. It now reads "While you're lost in space…" with a screen-reader-only "play Space Invaders" suffix, so the accessible name still starts with the visible text (WCAG 2.5.3). It is a ui/Button ghost instead of a hand-styled button.
The game now stays open past the end of the wait, which exposed bugs in every surface that hosts it. Each one is fixed and pinned by a test. - Chat and buddy refocus: when the reply landed, the composer took focus back mid-run (during play focus sits on body). DinoGame then ignored every key aimed at the text field: the dino died, Esc was dead, and Space typed into the box. The refocus now waits until the game closes, then returns focus if nothing else holds it. - Buddy visibility: Space was armed on isThinking alone, but the session lives app-wide and the dock mounts only while open. So a minimised dock plus Space opened an invisible game that held the one-game slot for good. Surfaces now register themselves (the open dock, BuddyPage); Space arms only while one is visible, and the game closes when none is. The hire/team conversation switch also closes it. - Outcome labels: "Reply ready" used to show after Stop and after failed streams too. The chat records each turn's outcome (finished / stopped / failed, exposed per chat by useChat.turnOutcome); buddy derives it from the last message's error. The badge now reads "Stopped" or "Reply failed", in words not just colour. The shared label lives in dinoOutcome.ts. - Onboarding: a failed generation fell back to the last snapshot, whose working phases kept isGenerating true. Phases froze, the timer kept running, and retry never appeared. Failure now wins and closes the game; on success the snapshot shows every phase completed. - Data sources: "Sync complete" ignored an update still in flight; replyReady now also needs !isSyncing. - Drawer Escape: SidePanel's document listener closed the drawer before DinoGame's window listener saw the key. DinoGame now takes Escape in the capture phase, and SidePanel skips a key already handled, so the first Esc closes the game and the second the drawer. - Screen readers: the whole game sat in role=status, so the score was announced about 12 times a second. The game is out of the live region now; one status line gives the result and final score, the running score is aria-hidden, and the typing bubbles keep a short "Thinking…". - Consistency: the game's raw emerald/white colours are now app tokens, its buttons ui/Button, the buddy bubble uses centralSpringToken, and dino-game* test ids are added. BuddyDock's prop docs are back on their own props. ChatPage closes the game on a chat switch from an effect instead of during render.
The drawer test for "an Escape already handled inside is ignored" was appended to SidePanel.test.tsx. #264 also adds imports and tests at the end of that file, so whichever PR merged second hit a conflict on its import lines and on its final block. The test moves unchanged into SidePanel.escape.test.tsx, and SidePanel.test.tsx matches dev again. merge-tree against feature/KB-updates-final is now conflict-free.
The trigger's "this control owns Space" guard also covered disabled controls. Space on a disabled button does nothing, and a busy ui/Button (`loading`) is disabled while it can still hold focus. So clicking "Update repo" in the source drawer and then pressing Space never opened the game, during the one wait that is exactly what the game is for. isInteractiveTarget now skips `:disabled` and `[aria-disabled="true"]` controls. Text fields stay protected even when disabled: the typing guard is unchanged. This was also why the SourceDetailsPanel "update in flight" test was flaky. In jsdom, blur() does nothing on a disabled element, so focus stayed on the busy button. The test passed only when the drawer's initial-focus frame happened to land after the click. The test now waits for that frame, clicks, asserts focus is still on the busy button, and presses Space once, without the retry loop. Passed 8/8 runs alone and 3/3 with its folder.
daniilperkin
marked this pull request as ready for review
September 24, 2026 12:23
daniilperkin
requested review from
BabuPlk,
DavidLeuter and
kiranfin
and removed request for
DavidLeuter
September 24, 2026 12:23
This was referenced Sep 24, 2026
… out of Space Invaders Addresses the verified findings from the PR #184 review. 1. useSpaceOpensDino released the shared game slot during render. `releaseGameSlot(host)` mutated the module-level `gameHost` inside the render-phase "adjust state when a prop changes" blocks (unlock flipped off, wait ended). Render must stay pure: a render React throws away (concurrent retry, StrictMode double render) would free a slot the committed game still holds, and a second armed surface could then claim it and open a second game - exactly what the slot exists to prevent. The render-phase blocks now only call `setGameActive(false)`; a new effect releases the slot whenever `gameActive` is false after commit. `close()` still releases eagerly because it runs in an event handler, where mutating module state is fine. The suggested alternative (move both the release and `setGameActive(false)` into one effect) was not used: a setState inside an effect trips `react-hooks/set-state-in-effect` and costs an extra render with the stale game still on screen. 2. Space Invaders handled keys with no modifier or typing guard. Ctrl/Cmd+A (select all) and Alt+D matched `KeyA`/`KeyD`, got `preventDefault` and moved the ship. The keydown handler now ignores Ctrl/Meta/Alt presses and keys typed into a text field, the same guards DinoGame already has. keyup stays unguarded on purpose so a held key is always released. Escape moved to a capture-phase window listener with `preventDefault` + `stopPropagation`, again mirroring DinoGame, so one press closes only the game and never a surrounding surface too. New tests/unit/features/easter-eggs/SpaceInvaders.test.tsx covers movement keys taken, modified presses and text-field typing left alone, and Escape claimed before document listeners. Findings checked and deliberately not changed: - Raw <button>s in SpaceInvaders: game surfaces are a listed exception in FRONTEND_CODING_STANDARDS.md section 4. - `text-white` on `bg-app-brand`: the same pairing `ui/Button` primary uses; `text-app-text-inverse` would turn dark in dark mode. - EggLoadingFallback `w-[680px]`: the dialog is `max-w-full overflow-hidden`, so it clips instead of widening the viewport; `w-full` inside that shrink-to-fit dialog would collapse the reserved box to zero width. Gates: prettier format:check clean, npm run build OK, npm run lint 0 problems, npm run test 368 files / 3082 tests passed.
DavidLeuter
reviewed
Sep 25, 2026
DavidLeuter
left a comment
DavidLeuter
left a comment
Collaborator
There was a problem hiding this comment.
Went through the full source diff against dev (87 files) — really solid work. The registry + single shell, the effect bus with one renderer, the shared game slot and the capture-phase Escape are all clean, and the test coverage for the tricky bits (focus contract, slot release, honest badges, drawer Escape) is impressive. CI is green on 061276ed.
One behavioural finding (inline on GenerationScreen.tsx) and a few non-blocking nits:
Nits (non-blocking)
isTypingTarget/isInteractiveTargetlive inhooks/useDinoWaitingGame.ts, butSpaceInvadersandDinoGameimport them from there too. They're generic key guards —easter-eggs/lib/would be a more natural home than a dino hook.- Import specifiers are touched in both directions:
ThinkingIndicator/BuddyMessageadd.tsx/.tsextensions, whileNotFoundPageremoves them. Pure churn — maybe pick one and leave untouched imports alone. registry.ts: Space Invaders'lazy()is hoisted into a const while 2048's is inline — small inconsistency.- The Verification section still says head
17cdf9b6; the head is now061276ed(CI is green there, so just the description).
Happy to approve once the GenerationScreen point is addressed or you explain why it can't happen with the current stage stream.
…es the game, key guards to lib Responds to review 5317855360 on PR #184. Every point was checked against the code. 1. GenerationScreen derived "still generating" from the phases (real). `isGenerating` was `!isCompleted && (no phases || some phase waiting/working)`. OnboardingJourneyProvider builds `phases` incrementally from `onStage` events (unknown names are pushed) and a phase's state comes from its detail line, so every phase can read done while `generation.status` is still "running": after the last phase reports "Completed..." but before `onDone` (path still persisting), and in any gap before the next phase's first stage event. In that window an open game showed "Path ready", the trigger disarmed, the Space hint vanished and the clock stopped - although the run could still fail. GenerationScreen now takes a required `isRunning` prop, OnBoardingPage passes `generation.status === "running"` (which it already computed), and `isGenerating = isRunning && !isCompleted`. The phases are purely presentational. Made the prop required rather than optional so no caller can silently fall back to the old phase-based guess. New test: all phases done but still running -> hint shown, game opens, no "Path ready"; existing tests now pass `isRunning` explicitly. 2. isTypingTarget / isInteractiveTarget moved to easter-eggs/lib/keyTargets.ts. They are generic key guards used by DinoGame, SpaceInvaders and the Space-to-play hook, not dino-hook internals. No re-export from the old module: all three importers (plus the hook test) now import from lib. 3. Import-specifier churn reverted. ThinkingIndicator's untouched imports are back to dev's extensionless form and NotFoundPage's are back to dev's `.tsx` form; the one import each file genuinely adds follows that file's style. BuddyMessage was not churn - its extension-bearing lines are all new imports - so it is left alone. 4. registry.ts: Space Invaders' lazy() is inline like 2048's instead of a hoisted const. Not changed here: the PR description's Verification head SHA is updated on the PR itself after this push. Gates: format:check clean, npm run build OK, npm run lint 0 problems, npm run test 368 files / 3083 tests passed.
Collaborator
Author
|
Thanks @DavidLeuter. Addressed in
|
DavidLeuter
approved these changes
Sep 26, 2026
DavidLeuter
left a comment
DavidLeuter
left a comment
Collaborator
There was a problem hiding this comment.
Thanks @daniilperkin — checked b3dd474e against my review, all points are resolved:
- GenerationScreen:
isGenerating = isRunning && !isCompletednow comes from the real run status, and the phases are purely presentational. I also walked the status transitions onOnBoardingPage:errorunmounts the screen, andidleis only ever reached viaclearGeneration()afterdone/error(it's a no-op while running), wherelastGeneration.completedkeeps "Path ready" honest. So there's no path left where an open game claims completion early. Making the prop required is a nice touch, and the new "all phases done but still running" test covers exactly the case I raised. - Key guards in
easter-eggs/lib/keyTargets.tswith no re-export, import churn reverted to each file's dev style, registrylazy()inline, description updated. Fair point onBuddyMessage.
CI is green on the head. Approving 🚀
Reviewed with Claude Code.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
📌 What this PR is
Two things in one branch, by design:
src/features/easter-eggs/): a central registry for modal games, one accessible modal shell, a single app-wide effects layer driven by a bus, the dino waiting-game shared by chat / onboarding / buddy, discoverable triggers,prefers-reduced-motionhonoured everywhere, and the legacy per-feature implementations deleted.devbackmerges. The branch's diff againstdevis now 86 files, +5346 / −1075. It contains nothing that lives on another branch; the knowledge-base and tooling tweaks it used to carry moved to feat(knowledge-base): server-side pagination, URL-state filters, sort, date/language facets, bulk delete and AI status #264 (see Upgrade round), and the branch merges into feat(knowledge-base): server-side pagination, URL-state filters, sort, date/language facets, bulk delete and AI status #264 without conflicts.🔀 What was absorbed along the way
dev(5 merge rounds, last =d531e21e)Conflict policy throughout: incoming (dev/PR) changes win, except where the branch itself carries newer consolidated work (the egg architecture, the sidebar's Blueprints entry beside dev's Hire Setup, the reviewed buddy-session fixes). After every round the tree is swept for the known trap — dev-side branches re-importing the pre-overhaul dino state in
ChatPage— and the sweep is at zero dead markers on this head.🥚 The easter-egg overhaul (core changes)
Architecture
registry.ts): modal eggs (game-2048,space-invaders) addressed by id everywhere; each game component isReact.lazy-loaded, so a game's chunk is fetched only on first open. The dino waiting-game is deliberately not a registry entry — it plays inline in the chat thread, the onboarding step and the buddy, never in the shell. Verified: it stays its own lazy chunk.EggModalShell.tsx): one accessible shell (role="dialog",aria-modal, scroll lock) replaces three hand-rolled modal implementations.eggEffectBus.ts,EggEffectsLayer.tsx,ConfettiBurst.tsx): phrase detection (eggPhrases.ts) fires app-wide effects drawn once, app-wide — barrel roll, matrix rain, confetti.EggEffectsLayersits outside the auth gate deliberately: a fired effect always has its renderer.dino/,game2048/,space-invaders/and the stale global shortcut hooks are deleted; README updated.Behaviour
useDinoWaitingGame.ts):Spaceopens it while an AI turn runs, shared by chat, onboarding generation and buddy; a shared app-wide slot guarantees one game at a time; it can outlive its turn (keepActiveUntilExit), which is exactly why both the trigger and the game's own key handlers ignore keystrokes landing in text fields — the composer regains focus the moment a turn ends, and Space/w/s/arrows/Escape keep working mid-sentence.status === "done", ingestion only when the source really reachedconnected— not merely "busy flipped off" (a failed sync/turn no longer promises a reply, path or sync that does not exist).eggHintglow pulse on header icons, triple-click 2048 on the dashboard, cogwheel triple-click dino unlock in Settings, sidebar-logo physics roll, 404 rocket teaser, egg phrases in chat and buddy.Escapemechanism per press — the iframe captures it and posts a same-originEGG_EXITmessage; the frame also takes focus on load so arrows work without a click first.prefers-reduced-motion.🧾 Review responses this branch has been through
d1a7ffc9,8a0335e0)Spacetrigger never fired (composer focus parity with the chat, added), and the barrel roll ignoredprefers-reduced-motion. Plus the app-wide one-game-at-a-time slot, a WCAG 2.5.3 accessible-name fix on the 404 teaser, the 2048 double-Escape, and a batch of docs/hygiene items3f701f5e)ba9f7d0d…17cdf9b6)9cd5081e)061276ed)b3dd474e)isRunning) instead of guessing from phases, so "Path ready" can no longer show while the run is still persisting; key guards moved toeaster-eggs/lib/keyTargets.ts; import-specifier churn reverted; registry lazy() inline for both games🧪 Verification — head
b3dd474enpm run build(tsc -b && vite build)npm run lintnpm run format:checknpm run testnpm run a11y(vitest-axe)npm run testrun abovegit merge-treeagainstdevand against #264 (feature/KB-updates-final): no conflictsManual verification (carried from the original pass; surfaces unchanged since)
Escapecloses it from both the iframe and the shell.party/party time/let's partyin Chat or Buddy fires confetti;do a barrel rollandmatrixfire their effects via the app-level layer.Spacewhile the AI thinks, on chat, onboarding and buddy.prefers-reduced-motion: reduce— glow pulse, hover expansion, logo drop and barrel roll included.📋 Definition of Done
npm run lint,npm run format:check,npm run buildgreenprefers-reduced-motioncompliantdevcontains only this branch's own work — the knowledge-base/tooling tweaks moved to feat(knowledge-base): server-side pagination, URL-state filters, sort, date/language facets, bulk delete and AI status #264, which this branch merges into cleanly