Best practices for building 3D map applications with Navara (@navaramap/three). Use whenever writing or reviewing code that uses ThreeView, sources/layers/materials, plugins, mesh/effect/light descriptors, feature evaluation, picking, or geodetic math — in application code, examples, or documentation code snippets.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Best practices for building 3D map applications with Navara (@navaramap/three). Use whenever writing or reviewing code that uses ThreeView, sources/layers/materials, plugins, mesh/effect/light descriptors, feature evaluation, picking, or geodetic math — in application code, examples, or documentation code snippets.
Using Navara (@navaramap/three)
Navara is a 3D globe map engine: reusable GIS logic lives in a Rust/WASM core, and drawing is delegated to a swappable CG-rendering library. @navaramap/three is the Three.js-based binding (currently the only one; more rendering engines are planned — avoid wording that fixes Navara to Three.js in prose). Its public API is ThreeView plus a declarative Source/Layer/Descriptor model.
Packages
Package
What it provides
When you need it
@navaramap/three
ThreeView (default export), Color, geodetic math utils, MeshDesc/EffectDesc/LightDesc base classes, handle types, built-in attribution UI (view.attribution, on by default)
Install every package your code imports from, explicitly. @navaramap/three-default-plugin depends on @navaramap/three-default-descs, but importing descriptor types from the latter with only the plugin installed fails to resolve under pnpm (transitive deps are not importable). Peer deps to install alongside: three, postprocessing.
When DefaultPlugin is used, parameterize the view as new ThreeView<DefaultDescriptions> so descriptor keys are typed.
Start from a recipe when building a scene
For "make it look good" goals, use the proven compositions in references/recipes.md (e.g. the photoreal base scene: DefaultPlugin + photoreal scene + Re:Earth quantized-mesh terrain + EOX satellite imagery + attribution) instead of assembling pieces from scratch.
Four capability tiers — pick the lowest that solves the problem
Declarative API — add sources and layers with plain config objects. Covers basemaps, terrain, vector data, 3D Tiles. → references/declarative-api.md
Plugins — reusable purpose-built features (photoreal scene, first-person walk, DOM overlays, attribution), and how to write your own. → references/plugins.md
Custom descriptors — your own meshes/effects/lights, with access to the render pipeline, depth buffer, and normal/G-buffer (MRT). → references/custom-desc.md
Critical gotchas (apply everywhere)
addPlugin() after init() throws. All plugin init()s run in parallel during view.init().
Updates are partial merges:Layer.update(), Source.update() and mesh/effect/light handle .update() all merge into the current config — omitted materials and omitted fields within a material are preserved (verified: layer.update({ point: { color } }) keeps size/clampToGround). Note: Layer.update()'s JSDoc says "the entire configuration is replaced" — that text is outdated; trust the merge behavior shown in its own @example.
Vector materials consume only their native geometry unless geometryTypes opts in more.point/billboard/text default to ["point"], polyline to ["line"] — polygon data with only a polyline material renders nothing. Opt in with e.g. polyline: { geometryTypes: ["line", "polygon"] } (polygon boundary rings incl. holes render as closed polylines) or point: { geometryTypes: ["point", "line", "polygon"] } (one point per vertex, ring-closing duplicate skipped). The array replaces the default, so include the native category to keep it. Boundary polylines are at ring base height only — extruded-polygon edges remain polygon.outline's job. Two caveats: (1) geometryTypes is construction-time config — layer.update() applies a new value only to tiles loaded afterwards (resident tiles keep their old geometry, and toggling back never removes already-built derivations); set it at layer creation or delete/re-add the layer. (2) On tiled paths derivation walks tile-clipped rings; boundary polylines drop clip-introduced edges automatically (draped and non-draped are both clean), but vector-tile-derived points can still appear at clip-introduced vertices near tile edges.
Sources are reference-counted:source.delete() returns false while any layer still references it. Updating a source resets and reloads every referencing layer.
Inline GeoJSON data does not reload via source.update() (verified 2026-08: readouts driven from the same code ran, but the rendered features never changed; URL-based sources do reload). Re-pointing the existing layer at a fresh source also failed to render. The working pattern for dynamic inline data is rebuild both: . For measurement-style dynamic overlays (a handful of markers + a line), mesh descs are the simpler tool: add/delete markers and a arc via handles — see .
Lighting — when it's needed and which light to add
Lighting only affects surfaces that carry normals. Anything without normals is shaded uniformly (fullbright) and needs no light. But once normals are present, no light means the surface renders black. Add at least one light when you have:
meshes / GLTF models that ship vertex normals
quantized-mesh terrain loaded with requestVertexNormals: true
raster-dem terrain rendered with a hillshade material
the bare globe when the view was constructed with useNormal: true (init-only) — needed for the sun to shade the globe when no terrain or hillshade layer supplies normals
Built-in lights — register via DefaultPlugin, then view.addLight<T>({ ... }):
Descriptor
Key
Use
AmbientLightDesc
ambient
flat fill — raise overall brightness or light everything uniformly; no direction, no shadows
LightProbeDesc
lightProbe
pseudo-IBL from spherical-harmonics coefficients (e.g. a baked night ambient)
SkyLightProbeDesc
skyLightProbe
dynamic sky ambient that follows the sun through the atmosphere system
SunLightDesc
sun
directional sun; its direction is derived from atmosphere.date, and it casts CSM shadows
addDefaultPhotorealScene() already adds a sun + skyLightProbe — start there for realistic scenes and add ambient/lightProbe to taste.
A sun light also works standalone, without addDefaultPhotorealScene(): its direction still follows atmosphere.date, and its color is computed from the atmosphere transmittance texture, which loads on demand (verified 2026-08 shading quantized-mesh terrain with only view.addLight({ sun: {} })). Note view.toneMappingExposure has no visible effect in such a scene — tone mapping is inactive without the photoreal effect stack, so don't reach for exposure to fix brightness there; use the light's intensity.
atmosphere.date is local-time-sensitive. It's a plain JS Date, so new Date("2024-06-21T12:00:00") is read in the device's timezone — the sun lands in a different place per machine. Pin a global instant with an explicit UTC string (new Date("2024-06-21T12:00:00Z")). To hold the same time-of-day or sun elevation while flying the camera to another country, don't recompute by hand — use view.atmosphere.setDateFromCameraAt({ lng }) (keeps local solar time) or setElevationFromCameraAt({ lat, lng }) (keeps sun elevation).
Where to verify — never guess API details
Primary reference: the docs site — https://navara.world/docs/ (Japanese under /ja/). Sections: /three/ (core API: sources, layers, materials, camera, events), /three_default_descs/ (every built-in mesh/effect/light Descriptor and its options), /three_default_plugin/, /three_plugins/.
Do not guess material or config property names — this skill shows patterns, not exhaustive option lists. Verify exact fields against the docs site, or the TypeScript definitions in node_modules/@navaramap/* (.d.ts).
Working inside the Navara repository? The docs source is docs/src/content/docs/ and runnable examples are web/navara_three/example/pages/ — reference paths in this skill starting with example/pages/ refer to that examples directory.
Non-draped vector polylines are lit; draped ones are unlit.clampToGround: false polylines render through a lit shader (material.lights = !isTexturized), so with zero lights they draw black — invisible on a dark basemap and a classic "the line is missing" trap. Draped (clampToGround: true) lines bake into tile textures and ignore lights. Fix: add view.addLight({ ambient: { intensity: 1 } }) (needs DefaultPlugin) to show the line at its plain color.
Polyline width is screen px, but clamped by maxWidth in world meters (default 1000). From a whole-globe camera 1 px ≈ 11 km, so the default clamp makes any line sub-pixel — pass maxWidth (e.g. 100_000) for high-altitude views. Draped polylines additionally vanish entirely at globe-scale zoom (bake resolution) regardless of maxWidth (verified at camera height ≈ 9,000 km; they render fine at ≈ 2,000 km) — use non-draped lines for global arcs.
WASM-backed math throws before view.init().degreeToRadian, EllipsoidGeodesic, geodeticToVector3, … call into navara_wasm_api, which view.init() initializes — a module-top-level call fails with e.g. "Cannot read properties of undefined (reading 'angleToRadian')". Plain lat/lng/height constants are fine at module top level (they're just degree numbers); call the WASM-backed functions on them only after init.
Layer render order = add order (e.g. add terrain before the raster basemap draped on it).
elevationHeatmap: omitting logBoundary breaks software-GL backends.Math.log(mat.logBoundary) (navara_three mesh/tile/index.ts) turns the missing value into a NaN uniform, and the branch-free ternary in elevation_pars_fragment.glsl evaluates the log path even with logarithmic false. ANGLE Metal flushes the NaN (renders fine); SwiftShader — i.e. headless Playwright, the screenshot pipeline, CI — propagates it and every pixel renders one constant color (verified 2026-07 on both backends with identical code). Until the engine guards that computation, pass logBoundary explicitly when headless captures matter. The material itself is unlit and needs no terrain layer or lights; globe.elevationColormap supplies the colors.
Deferred lighting recipe: set view.lit = false so materials output albedo while the lit pipeline keeps writing normals and the shadow buffer, then re-light in a late post pass from getNormalTexture() + getShadowTexture(). A per-mesh/material lit: true opts something out, and the shadow buffer's G channel is the per-pixel mask for it — transparent unlit materials cannot be masked this way (they are already blended into the albedo). Toggling lit recompiles, so it is a config switch, not a per-frame control. Worked effect: example/pages/debug/buffers/run.ts; option semantics: docs three/api/threeview-properties#lit.
aerialPerspective.irradiance: true IS a deferred lighting pass — always pair it with view.lit = false. It sets sunLight/skyLight on the AerialPerspectiveEffect, which re-lights the G-buffer from the precomputed atmosphere using the normal buffer. With view.lit left at its default true the scene is lit twice (forward + atmosphere) and washes out, which is what forces the unnaturally high exposures (10+) seen in older snippets; unlit + irradiance sits around 3. Verified 2026-08: the deferred pass does not consume the shadow G-buffer, so sun.castShadow cascaded shadows vanish under lit = false (cloud shadows survive — the effect samples those from atmosphere.shadow). Close-up scenes that need cast shadows or transparency keep irradiance: false and the default lit.
SnowMeshDesc / RainMeshDescmaxHeight is an opacity fade, not a spawn ceiling. Every frame the material opacity is multiplied by max(1 - cameraHeight / maxHeight, 0), so with the default maxHeight: 3000 a camera at or above 3 km renders zero visible snow while everything still runs — a silent empty screenshot. Particles also spawn in an areaWidth x areaHeight box around the camera and size is world-space (sizeAttenuation), so flakes only read at their nominal size when the camera is inside the weather, not looking at it from above.
sun.shadowFar defaults to 50 km, spreading the 4 cascades so wide that person- or car-scale shadows never resolve. For character/street-level scenes use shadowFar: ~1000 with shadowLambda: 1 (fully logarithmic split) — that is what makes a walking character cast a visible shadow (example/pages/examples/plugin/person-view).
Never reconstruct a G-buffer normal with xy * 2 - 1 — it is octahedral; decode with unpackVec2ToNormal from NORMAL_PACKING_SHADER. The wrong reconstruction looks plausible and costs a long debugging session. Known inconsistency to watch for: instancedSprite.frag.glsl / sdfText.frag.glsl define a LOCAL xy*0.5+0.5 packer that does not match.
Allocating any optional G-buffer silently re-routes meshes into the MRT pass (so their own normal/shadow/mask is written instead of the terrain's showing through). Expect the first addEffect that needs a buffer to recompile and to shift where meshes render. Self-shadow acne (striping) on large mesh faces in the shadow buffer is tamed with sun: { shadowNormalBias: ~3 }.
A custom effect should inherit scene config, not re-declare it — ctx.findLight("sun") / findEffect / findMesh resolve active descs by registered key, so a lighting effect reads the sun's real intensity and view.atmosphere.sunDirection instead of growing its own options.
Toggling terrain on/off: delete/re-add the layer.terrainLayer.update({ terrain: { show: false } }) has no visible effect (verified 2026-07); layer.delete() drops the surface to the flat ellipsoid and draped raster layers re-drape automatically.
Handle events vs desc events: mesh/light/effect handles (BaseHandle) only emit deleted — desc-specific events (load / error / animationReady on GLTF, instanced-GLTF and splat descs) live on the desc, so subscribe via handle.ref.on("load", ...).
Never write to view.camera.raw frustum fields (fov etc.) — the engine overwrites them and Rust-side culling desyncs. Use the view.camera.fov/near/far setters.
Units: every public lat/lng API takes degrees — setCamera/flyTo, geodetic, terrain sampling, geodeticToVector3/vector3ToGeodetic, EllipsoidGeodesic (headings in degrees too). Mesh position is ECEF meters. Three.js-level angles stay radians (Euler rotation fields, rotateAroundAxis/rotateAround — that's what degreeToRadian is still for); batch IDs are 24-bit.
Ground placement needs sampleTerrainMostDetailed: the synchronous sampleTerrainHeight reads only render-resident tiles, so from a distant camera it returns a coarse-LOD height (tens of meters off) — await view.sampleTerrainMostDetailed(source, positions) (source = the registered terrain source's handle or id, always explicit) fetches the source's max-LOD tiles regardless of the camera. See references/imperative-api.md.
Placing a mesh geographically: use geodetic — { lng, lat, height, heading, pitch, roll, scale } in degrees and metres, the same convention as setCamera. It builds a West-Up-North (Y-up, +Z north) frame, which matches glTF's front=+Z / up=+Y convention, so an unmodified glTF asset needs noRx(+90°) up-axis correction. heading is the compass bearing the asset's front faces, and pitch/roll also cover in-frame flips — a Y-down splat capture is upright with pitch: 180 instead of a separate rotation offset. Add heightReference: "terrain" to make height relative to terrain (a live subscription that re-clamps as tiles stream — for a one-shot placement, sampleTerrainMostDetailed is cheaper). position/rotation/scale stay available as offsets inside the frame, and geodetic, matrix and matrixWorld share one placement slot — setting any two of them throws ConflictingTransformError.
Low-level placement: raw position/rotation/scale are earth-centered ECEF metres, so a bare position won't sit upright at a lng/lat. Set matrixWorld to a tangent-frame matrix when you need a frame geodetic doesn't build — eastNorthUpToFixedFrame (ENU, Z-up), northEastDownToFixedFrame (NED), northUpEastToFixedFrame (NUE, Y-up but +X north), northWestUpToFixedFrame (NWU), westUpNorthToFixedFrame (WUN, what geodetic uses). All are exported from @navaramap/three. Note the Z-up frames need an Rx(+90°) correction for glTF assets; WUN does not. See references/imperative-api.md.
Init-only options cannot change after init(): shadow, maxSse, segments, useNormal.
Optional G-buffers: declare, never assume. They exist only while an active effect lists them in static requiredBuffers, so read them through the ViewContext/MRTPassEffectDesc getters (undefined when disabled) and re-fetch every frame — a config change rebuilds the attachments. Never hardcode an attachment index; they are packed and shift. Switching the set reallocates attachments and recompiles shaders, so add an effect once and tune via handle.update() rather than add/remove per frame. Only MRT-pass fragments write the G-buffer at all — a mesh sitting in the opaque/transparent scene contributes nothing to it.
Effect compatibility:hideUnderground: false and logarithmicDepthBuffer break some effect descriptors — test, and prefer defaults.
picking: true (constructor, default on) is required for the featureClick (click or touch tap) and featureHover/featureEnter/featureLeave (pointermove) events and pickable meshes. Hover picking activates only while a hover-family listener is registered.
API stability tiers:ThreeView = Tier 0 (stable). Plugin + ViewContext = Tier 1 (may break between minor versions). Keep app code on Tier 0 where possible.
Runtime assets (atmosphere/cloud/noise/water) bundle automatically — each file is referenced via a static new URL(..., import.meta.url), so Vite/webpack emit them into the app build; no copying or config needed. To self-host instead, pass atmosphereAssetsUrl (Atmosphere) / assetsUrl (Clouds) pointing at a directory that keeps the original filenames (transmittance.exr, local_weather.png, …); both default to undefined = bundled assets. There are no directory URL constants (appending filenames to a directory can't survive bundler hashing) — only per-file constants exist (ATMOSPHERE_TRANSMITTANCE_URL, CLOUD_LOCAL_WEATHER_URL, STBN_URL, WATER_NORMAL_URL, STARS_ASSETS_URL, …).
