- Ebike replacement parts: cooling core (correct, anchored at refroidisseur)
+ four distractors anchored at cabledroit/cablegauche/pucehaut/pucebas.
Removes the ad-hoc gant_l/talkie distractors in favor of consistent
case-anchored visuals.
- Pylon replacement parts: cable1 + cable2 (alternative correct, both with
caseLockGroup 'pylon-cable' for upcoming soft-lock) + refroidisseur and
two puce distractors anchored to packderelance.
- Farm replacement parts kept as-is (caseAnchor undefined falls back to
placeholder slot positions for backward compatibility).
- RepairGame threads anchors from RepairCaseModel through RepairMissionCase
to RepairRepairingStep; replacement-part initial position now resolves
to the anchor world position when caseAnchor is set, falling back to the
legacy slot index otherwise.
- RepairMissionConfig.requiredReplacementPartId (string) is replaced by
requiredReplacementPartIds (readonly string[]) so a mission can accept
several alternative correct parts (e.g. pylon will accept either cable).
- RepairMissionPartConfig gains optional caseAnchor (where the standalone
spawns inside packderelance), caseLockGroup (mutually exclusive parts),
and targetNodeName (snap onto a node of the broken model rather than a
placeholder slot in the case).
- RepairScannedBrokenPart gains targetNodeName so scan results can carry
this hint through to the repairing step.
- RepairRepairingStep validation logic (placed/wrong/feedback) now matches
any id in requiredReplacementPartIds. Existing data is migrated mechanically
(single-element arrays); part-level new fields are wired in subsequent
commits.
- Fix REPAIR_CASE_LID_NODE_NAME from "partiesup" to "partsup" (the actual
node name in packderelance.gltf), restoring the lid open/close animation
that was silently no-oping since introduction.
- Add REPAIR_CASE_PART_ANCHOR_NAMES (cabledroit, cablegauche, pucehaut,
pucebas, refroidisseur) and REPAIR_CASE_PART_ANCHOR_FALLBACKS for
case-local positions used when the GLTF lacks a node (refroidisseur).
- RepairCaseModel now resolves these anchor nodes on mount, hides existing
meshes underneath them, and creates lightweight Object3D placeholders for
missing names so the anchoring pipeline is uniform.
- Each frame, anchor world positions are converted to step-local space and
emitted via the new onAnchorsChange callback (debounced via signature like
placeholders). Consumers added in subsequent commits.
The browser MediaPipe model (hand_landmarker.task float16) was failing
to detect hands at 320x240 — MediaPipe ran for 10s straight without
ever returning a hand. Bumping the requested camera resolution to
640x480 (browser only — backend still ships 320x240 JPEGs over the
WebSocket) makes the model reliably pick hands up.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
(1) HandTrackingProvider always renders the same JSX root
(HandTrackingRuntime) so toggling `enabled` no longer remounts the
<Canvas> below — that remount was destroying the WebGL context every
time the player entered an interaction zone.
(2) Add HAND_TRACKING_LINGER_MS (2s) cooldown on `enabled` so brief
walk-throughs of a trigger zone don't tear down MediaPipe before it
has time to initialize the webcam + model + first frame (cold start
~800ms).
Resolves the WebGL context lost + respawn loop and restores visible
hand tracking in the backend runtime. Browser JS runtime detection
quality is a separate follow-up.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
In dev, <StrictMode> intentionally mounts → unmounts → remounts each
effect to surface non-idempotent code. The hand tracking hooks were
calling getUserMedia and creating MediaPipe / WebSocket runtimes on
every mount, which in practice ran the full start/stop/start cycle
inside a few milliseconds and pushed WebGL over its limit on top
of the loaded scene → context lost.
Add HAND_TRACKING_RUNTIME_START_DELAY_MS (80ms) and delay the actual
start() call behind a setTimeout in both useBrowserHandTracking and
useRemoteHandTracking. The cleanup clears the timer, so a fast
mount/unmount never reaches start(). 80ms is invisible to the user
(<5 frames at 60fps) and also absorbs rapid `nearby` toggles at
trigger borders.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Several mitigations against the WebGL context lost that fires when
hand tracking starts on a loaded scene:
- Canvas: fixed DPR [1,1], antialias off, scoped id="game-canvas",
context-lost handler releases MediaPipe and logs GPU memory counters
- optimizeGLTFScene: cap anisotropy at 2 and stop forcing mipmaps /
needsUpdate on every pass — avoids massive texture re-uploads
- MediaPipe: force CPU delegate (HAND_TRACKING_BROWSER_DELEGATE),
cache the landmarker instance, and expose releaseBrowserHandLandmarker
- useBrowserHandTracking / useRemoteHandTracking: idempotent cleanup
guarded by a cleanedUp flag, try/catch around the detect loop, and
release of the landmarker on stop
- World: mount HandTrackingGlove only when the matching hand is
actually present in the snapshot (status connected + hands.length > 0)
- HandTrackingGlove: drop the eager useGLTF.preload that was running
at startup whether or not hand tracking was used
Does not yet absorb the React StrictMode double-mount — that is the
follow-up commit.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Store the lil-gui Source controller so setHandTrackingSource() from
the settings menu can refresh its display, and log the transition
for traceability.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PointerLockControls now targets #game-canvas and respects the
settings menu, and document.exitPointerLock() only runs when a
pointer lock is actually active. The terrain ground snap in
PlayerController is gated on sceneMode === "game" so it doesn't
fight the physics test scene's flat floor.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lower grabbable spawn, expose GPS preview position/rotation as
constants, fix physics spawn to use PLAYER_EYE_HEIGHT, and silence
the console.log noise around waypoint loading in TestMap.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Solid character mesh does not need transparency; BLEND combined with doubleSided
caused depth-sort issues rendering inside-faces over front-faces (visible as
triangular orange artifacts on the face). Default OPAQUE single-sided is correct.
- Talkie does not need a LOD swap; remove the entry from
MAP_LOD_MODEL_PATHS so the same model.glb is used at any distance.
- The ebike repair distractors are meant to be other disassembled bike
parts, not random props. Drop the talkie radio distractor and keep
only the (thematically plausible) insulation glove alongside the
correct cooling core replacement.
The talkie folder now ships a single binary GLB; update the four call
sites (TalkieModel, gallery, two repair mission entries) to load
model.glb. The talkie-LOD path is unchanged.
The exported sapin-LOD and buisson-LOD models do not match their full
detail counterpart's baseline scale, so they appear oversized once the
LOD swap kicks in. Add MAP_LOD_SCALE_MULTIPLIERS keyed by map name
(sapin: 0.5, buisson: 0.8) and apply it on top of the chunk's existing
scaleMultiplier whenever the resolved model path is the LOD variant.
- Bump high to 30m chunk / HD<20m and ultra to HD<30m so HD models
persist further before swapping.
- Render the 5 graphics preset cards on a single row.
- Vegetation LOD selection now uses the distance to the nearest instance
in each chunk instead of the chunk centre, matching MapInstancingSystem
and avoiding premature LOD swaps when the camera enters a chunk.
Register the three new vegetation LOD models in MAP_LOD_MODEL_PATHS and
extend VegetationSystem with per-chunk distance-based LOD selection
(mirroring MapInstancingSystem). Chunk model paths are re-evaluated on
the existing CHUNK_CONFIG.updateInterval cadence and Suspense keys
include the resolved path so a swap unmounts the previous instanced mesh
cleanly.
Switching alphaMode from BLEND to MASK with a 0.5 cutoff lets the lampe
material participate in instanced shadow rendering correctly. BLEND is
known to interact badly with InstancedMesh shadow casting, producing
incorrect or missing shadows on the pylone lampe.
Restore ultra to its original behaviour (50m chunk streaming, HD within
20m, no fog) and introduce a new max preset that disables chunk streaming
entirely (loads all chunks unconditionally) and pushes the HD/LOD swap
distance to 50m. Add chunkStreamingEnabled flag to GraphicsPresetConfig
so the streaming gate honours the preset. The settings card label shows
'All' when streaming is off.
Shadows occasionally failed to render on initial load and the Fabrik
doorway sometimes blocked the player. Both issues are tracked down to
geometry that mounts after Lighting:
- Shadows: GLTFs and the merged static map mount imperatively after
Lighting, so materials get compiled against a renderer state that
pre-dates the final scene and bake a 'no shadow map' permutation,
silently dropping shadows. A WebGL context-restore cycle fixes it,
but is too invasive. New 'useShadowMapWarmup' hook replays it
cheaply: once the scene mesh count has been stable for ~1s, it
disposes the directional shadow map (three.js reallocates it on
the next render) and marks every material 'needsUpdate' so shaders
rebind to the freshly created shadow sampler.
- Doorway: the door slab + its Solidify-modifier frame (children of
the 'Thicken' parent in the LaFabrik GLTF) sat inside the doorway
AABB and prevented the player from walking through. Stripped from
the collision octree alongside the existing 'porte' slab; visual
rendering is unaffected.
Also: extract sun-relative-to-camera placement into a small helper,
remove the temporary diagnostic logs, and document the shadow warmup
in three-debugging.md.
- Lighting: replace single-frame needsUpdate with a 3-rAF warmup that forces
scene.updateMatrixWorld + sun.shadow.needsUpdate + gl.shadowMap.needsUpdate.
This restores the SceneShadowWarmup behaviour (deleted in 777e51e) inline,
so shadows survive Physics Suspense remounts and webglcontextrestored.
- octreeCollisionConfig: remove (comment out) the thin LA_FABRIK interior box
at x=-6.93 that was sealing the doorway despite the mesh hole; fabrik mesh
octree already provides surrounding wall collision.
- DebugOctreeVisualization: add Fabrik-only filter to inspect interior
collisions/non-collisions in isolation.
The previous talkie-overlay refactor lost the rest/active behaviour and
left the radio-shake CSS animation running constantly. Restore the
intended polish:
- Position lerp between TALKIE_REST_Y (idle) and TALKIE_ACTIVE_Y
(raised when narrator is speaking) with a subtle floating bob.
- Subtle z-axis float at idle, faster shake when active.
- Gate the CSS radio-shake on the new --active modifier so the talkie
is calm when no narrator dialogue is playing.
- Keep the face-camera rotation [0.18, PI, -0.08] from the original
overlay version.
Visibility still kicks in at the reveal step (no regression on the
recent fix).
Shadows still go missing intermittently despite the per-frame
needsUpdate fix. Add temporary console logs to narrow down the cause:
- [shadow:mount]: one-shot snapshot of renderer shadow flags and a
scene.traverse count of meshes with castShadow / receiveShadow at
Lighting mount time.
- [shadow:tick]: every 2s during useFrame, log shadow map enabled flag,
autoUpdate, sun.castShadow, sun intensity, shadow map texture
presence, sun and target world positions, and renderer draw calls.
To be removed once the root cause is identified.
Default visualization was unreadable because every node from depth 0 to
maxDepth was rendered with rainbow-coloured edges. Add three filters
exposed in the Debug folder:
- Octree Leaves Only (default true): skip internal nodes
- Octree Min Depth (default 4): hide the largest enclosing boxes
- Octree Opacity (default 0.35): tone down line density
Also skip nodes without triangles, drop the per-depth HSL palette in
favour of a uniform cyan, and bump default Octree Max Depth to 8.
Regression introduced by the narrator-video revert (1ad0c4d): the talkie
overlay was hidden whenever no narrator subtitle was active. Restore the
prior behaviour where the talkie stays visible from the reveal step
onward and only the --raised modifier and signal lines depend on the
active narrator dialogue.
Update three-debugging.md to reflect that the shadow intermittence
is resolved by explicit sun.shadow.needsUpdate = true at mount and
in useFrame after updateMatrixWorld.
Restore sun.shadow.needsUpdate = true at mount and in useFrame
after updateMatrixWorld. Lost during SHADOW_CONFIG centralization.
Matches develop's belt-and-suspenders pattern; autoUpdate alone
is insufficient because the sun follows the camera (matrix dirty
every frame) and three.js can skip shadow map re-render.
- Drop SceneShadowWarmup section, document centralized shadow config.
- Document the localized Suspense boundaries in World.tsx.
- Document the new player model and octree debug visualizations.
- Open note about intermittent first-load shadow rendering.
Late asset loads inside GameStageContent (e.g. EbikeSpeedometer's
useTexture) and the spawn-player block were bubbling Suspense up to the
root boundary in pages/page.tsx, which unmounted World mid-load and
triggered a redundant octree rebuild + shadow re-config. Localize the
suspension by wrapping each block in its own Suspense fallback.
Also mount DebugOctreeVisualization conditionally on the new debug
toggle.
- Show Player Model: render the main character GLTF in camera-local
space so it stays visible at any pitch.
- Show Octree: overlay the collision octree as colored line segments,
one wireframe per spatial cell, colored by depth.
- Octree Max Depth: cap recursion to keep the scene readable.
- Extract SHADOW_CONFIG into lightingConfig.ts (bias=0, normalBias=0,
cameraSize=95) matching the historically working values from develop.
- Drop SceneShadowWarmup; rely on sun.shadow.autoUpdate=true for
steady-state refresh.
- Enable cloud castShadow and traverse Ebike meshes for cast/receive.
Tom Boullay
math-pixel